diff --git a/.gitignore b/.gitignore
index 4456512..8dbcd91 100644
--- a/.gitignore
+++ b/.gitignore
@@ -1,4 +1,4 @@
-node_modules/
+node_modules
 .lloom/
 .DS_Store
 .env
diff --git a/README.md b/README.md
index b61fd4c..08484a1 100644
--- a/README.md
+++ b/README.md
@@ -3,7 +3,7 @@
 [![CI](https://github.com/enntity/lloom/actions/workflows/ci.yml/badge.svg)](https://github.com/enntity/lloom/actions/workflows/ci.yml)
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE)
 
-LLooM is a local-first LLM gateway for people who run serious open models on their own hardware. It treats NVIDIA systems—including DGX Spark / GB10—and Apple Silicon Macs as first-class platforms. LLooM sits in front of vLLM, SGLang, MLX, MTPLX, llama.cpp, Ollama, image generators, and other local runtimes, then exposes stable OpenAI-compatible and Anthropic-compatible APIs to agent tools.
+LLooM installs, manages, and serves AI models on your hardware. It treats NVIDIA systems—including DGX Spark / GB10—and Apple Silicon Macs as first-class platforms. LLooM sits in front of vLLM, SGLang, MLX, MTPLX, llama.cpp, Ollama, image generators, and other local runtimes, then exposes stable OpenAI-compatible and Anthropic-compatible APIs to agent tools.
 
 The goal is simple: install one bridge, let it inspect the machine, choose the best agentic model recipe from the LLooM community library, install the backend needed for that recipe, download and configure the model, keep it warm, and point Codex, Claude Code, OMP, OpenCode, Hermes, Zero, or any OpenAI-compatible client at one base URL.
 
@@ -15,11 +15,11 @@ The planned public community host is `https://lloom.enntity.com`; source checkou
 
 ## First-Class Platforms
 
-| NVIDIA / DGX Spark                                                                     | Apple Silicon                                                                     |
-| -------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------- |
-| CUDA, Blackwell, DGX Spark / GB10, and Linux NVIDIA hosts                              | M-series Macs with unified memory                                                 |
-| vLLM and SGLang are the primary high-throughput backends                               | MLX, MTPLX, OptiQ, and llama.cpp are the primary native backends                  |
-| Managed Docker runtimes, GPU-memory admission, warm/on-demand lanes, and Spark recipes | Native processes, unified-memory-aware recipes, model-root reuse, and Mac recipes |
+| NVIDIA / DGX Spark                                                                      | Apple Silicon                                                                     |
+| --------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------- |
+| CUDA, Blackwell, DGX Spark / GB10, and Linux NVIDIA hosts                               | M-series Macs with unified memory                                                 |
+| vLLM and SGLang are the primary high-throughput backends                                | MLX, MTPLX, OptiQ, and llama.cpp are the primary native backends                  |
+| Managed Docker runtimes, GPU-memory admission, warm/on-demand lanes, and Spark recipes  | Native processes, unified-memory-aware recipes, model-root reuse, and Mac recipes |
 | See [`docs/dgx-spark.md`](docs/dgx-spark.md) and [`docs/clusters.md`](docs/clusters.md) | See the bundled `apple-silicon-*` recipes                                         |
 
 Both platforms get the same gateway APIs, runtime policy, per-connection telemetry, live dashboard, client integrations, external-provider passthrough, and community recipe/benchmark workflow. Independent LLooM gateways can also form a heterogeneous lab behind one central endpoint; node profiles and optional GPU telemetry degrade cleanly across CUDA, Metal, ROCm, and CPU-only hosts.
@@ -34,7 +34,6 @@ cd lloom
 npm ci
 npm link
 lloom
-lloom up --go
 ```
 
 `npm link` installs the same `lloom` and `lloom-host` commands from your checkout. After the first npm release, `npm install -g lloom` will be the supported package install path.
@@ -48,7 +47,9 @@ curl -sS http://127.0.0.1:8100/v1/models
 
 Dashboard: [http://127.0.0.1:8100/](http://127.0.0.1:8100/)
 
-That is the 1.0 path. A bare `lloom` is a dry run first: it inspects the machine, asks the LLooM community host for the best known recipe pack and backend catalog, shows what will be installed, and refuses writes until you rerun it with `--go`. `up` is the named alias for the same first-run flow. `--go` applies the plan, confirms noninteractive writes, and starts the selected keep-warm runtime after setup. Use `--offline` when you want to ignore the host and select from only the local recipe library.
+On a new installation in an interactive terminal, `lloom` opens a local browser setup. It detects your hardware, asks what you want to use AI for, and recommends a compatible vendor recipe from the bundled library. Review the plan and choose **Set up my AI** to install it. The terminal stays open during setup. Chat setup checks a real response through the gateway before showing **Ready**; media installations show when output still needs verification.
+
+Use `lloom up --browser` to request browser setup explicitly, or `lloom ui` to open an installed gateway. The CLI remains available: `lloom --no-browser` previews the community-based plan; `lloom up --go` installs, integrates, and starts it. Scripted, JSON, offline, and explicit recipe commands keep their CLI behavior. See [the browser experience](docs/browser-experience.md) for the flow and current boundaries.
 
 The default gateway endpoint is `127.0.0.1:8100`; managed backend runtimes default to `8201-8299`. This source checkout defaults community lookup to the local development host at `127.0.0.1:8110`, starts it automatically if it is not already running, serves signed seed host data from `community/`, and requires signed recipe packs by default. Local imports still land in `recipes/` and `benchmarks/community/`. A production package should point at the signed public LLooM host. Most users should not need to care.
 
@@ -60,12 +61,12 @@ Use this path when validating the repository before a package release:
 
 ```bash
 npm install -g .
-lloom up
+lloom up --no-browser
 lloom up --go
 lloom doctor --no-runtimes
 ```
 
-`lloom up` should show the detected machine profile, the trusted community recommendation, the selected recipe, benchmark evidence, and the exact apply command.
+`lloom up --no-browser` should show the detected machine profile, the trusted community recommendation, the selected recipe, benchmark evidence, and the exact apply command.
 
 - On NVIDIA Linux, LLooM detects CUDA devices, compute capability, Blackwell, and DGX Spark / GB10 markers. Spark recipes use vLLM or SGLang, managed Docker containers, and explicit GPU-memory/runtime policy. The checked-in Spark deployment demonstrates a warm primary chat model, warm embedding model, and an on-demand alternate chat lane.
 - On a 96 GB Apple Silicon machine, the bundled development host should recommend `apple-silicon-qwen36-35b-a3b-mtplx` and select `Youssofal/Qwen3.6-35B-A3B-MTPLX-Optimized-Speed-FP16`. Lower-memory Macs should fall back to the 27B MTPLX recipe.
@@ -99,14 +100,16 @@ Then open OMP normally. The generated OMP config points at `http://127.0.0.1:810
 - Backend recipes for vLLM, SGLang, MTPLX, MLX LM, llama.cpp, Ollama, OptiQ, and stable-diffusion.cpp, with dedicated DGX Spark / GB10 and Apple Silicon recipes.
 - Community recipe packs and hardware-matched benchmark evidence so machines can select the best known model/backend recipe automatically instead of blindly chasing global tok/s.
 - Generated client profiles for OMP, OpenCode, Codex-compatible, Claude-compatible, Hermes, Zero, and any OpenAI-compatible client.
-- A small dashboard at `/` for local status and guarded setup actions, with a live topology that can switch between the default columnar racks and an action view whose camera and cards follow live models.
+- A browser dashboard with Live, Models, Machines, Clients, and Settings. Inspect real topology and memory blocks, preview a model’s expected footprint by pointing or focusing, and use models directly through chat or connected apps. LLooM prepares models automatically; optional readiness and memory controls live under Options & details. The action camera follows serving models while preserving manual zoom.
 
 ## Daily Commands
 
 Primary ladder (see `lloom help`; full catalog under `lloom help advanced`):
 
 ```bash
-lloom                         # preview plan
+lloom                         # browser setup on first interactive run
+lloom ui                      # open the installed gateway
+lloom --no-browser            # preview the CLI plan
 lloom up --go                 # install + integrate + start
 lloom down                    # stop the gateway and all managed model backends
 lloom doctor --no-runtimes
@@ -123,7 +126,7 @@ lloom add-model 'openai:http://127.0.0.1:8000/v1#my-model' --default --apply --y
 lloom serve --config ~/.lloom/config.json
 ```
 
-Bare `lloom`, `up`, and `onboard` all route to the same first-run flow. By default, the community request asks for the best known `agentic-coding` recipe with `tools`, `reasoning`, and `long-context`; use repeated `--workload`, `--capability`, or `--tag` flags to target a different kind of local model. `doctor` is the readiness view for humans and automation. `integrate` repairs or writes client configs from the registry. `add-model` imports an ad hoc Hugging Face, local, or Ollama model outside the community recipe library.
+The CLI `onboard` flow remains available independently of browser setup. By default, the community request asks for the best known `agentic-coding` recipe with `tools`, `reasoning`, and `long-context`; use repeated `--workload`, `--capability`, or `--tag` flags to target a different kind of local model. `doctor` is the readiness view for humans and automation. `integrate` repairs or writes client configs from the registry. `add-model` imports an ad hoc Hugging Face, local, or Ollama model outside the community recipe library.
 
 After `~/.lloom/config.json` exists, operational commands such as `doctor`, `models`, `serve`, `integrate`, `add-model`, and runtime controls automatically read that installed config when `--config` is not supplied. Before that file exists, those commands return a `not-installed` report with the exact `lloom up` command to run, instead of silently operating on bundled model defaults. Read-only planning commands such as `lloom`, `lloom up`, `onboard`, and `integrations` can still preview from the packaged gateway shell plus community data. Use `--config` whenever you want to inspect or operate a different config file.
 
diff --git a/SECURITY.md b/SECURITY.md
index 92dbe80..e35b57c 100644
--- a/SECURITY.md
+++ b/SECURITY.md
@@ -24,8 +24,11 @@ The maintainers will acknowledge reports as soon as practical, validate the issu
 - A signature proves which key signed a recipe pack. Trusting keys served by the same community host is equivalent to trusting that host and its TLS connection; use explicit local trusted keys for stronger publisher pinning.
 - Development keys and the checked-in public seed key are not production trust roots.
 - Model weights and external runtimes have their own licenses and security posture. LLooM does not make untrusted model code safe.
+- Browser management requests must come from the gateway's own origin. First-run setup additionally requires a local session token and applies only reviewed plans. Keep the setup terminal and tokens private. See [the browser security boundary](docs/browser-experience.md#local-security).
 - The public community MVP is read-only by design. Proposals arrive through reviewed pull requests; anonymous recipe and benchmark uploads are disabled in production.
 - Remote community feeds must use HTTPS and a locally pinned public signing key. Do not treat a signing key downloaded from the same remote host as an independent trust root.
 - The production community deployment is isolated from inference, databases, and Docker control-plane access. See [`deploy/community/README.md`](deploy/community/README.md).
 
 See [docs/architecture.md](docs/architecture.md) for route authorization and network-binding defaults.
+
+Remote admin writes require `security.allowRemoteAdmin=true` and a key in `security.adminApiKeys`. Inference credentials alone do not grant remote process or installation control.
diff --git a/bin/lloom.mjs b/bin/lloom.mjs
index 40566c2..556a565 100755
--- a/bin/lloom.mjs
+++ b/bin/lloom.mjs
@@ -3,7 +3,8 @@ import { spawn } from 'node:child_process';
 import { closeSync, existsSync, mkdirSync, openSync, readFileSync, unlinkSync, writeFileSync } from 'node:fs';
 import path from 'node:path';
 import { fileURLToPath } from 'node:url';
-import { Agent as UndiciAgent } from 'undici';
+// Dispatcher and fetch must come from the same undici copy (see server.mjs).
+import { Agent as UndiciAgent, fetch as undiciFetch } from 'undici';
 import {
   backendIds,
   defaultBackendVariables,
@@ -36,14 +37,18 @@ import {
   selectedRecipeIdFromCommunityPlan
 } from '../src/community-client.mjs';
 import { loadConfig } from '../src/config.mjs';
+import { mutateConfigSource } from '../src/config-mutation.mjs';
 import { runtimeControlTimeoutMs } from '../src/control-timeout.mjs';
 import { createDoctorReport } from '../src/doctor.mjs';
+import { createBrowserSetup } from '../src/browser-setup.mjs';
+import { shouldOpenSetup, openLocalBrowser } from '../src/first-run.mjs';
 import {
   ClusterCoordinator,
   currentNodeId,
   detectNvidiaSyncCluster,
   federatedNodeConfigFromSnapshot,
-  nvidiaSyncClusterConfig,
+  mergeNvidiaSyncClusterDiscovery,
+  nvidiaSyncDiscoverySummary,
   validateClusterConfig
 } from '../src/cluster.mjs';
 import { applyInit, defaultUserConfigPath } from '../src/init.mjs';
@@ -78,6 +83,7 @@ const __filename = fileURLToPath(import.meta.url);
 /** Command tiers used for help + installed-config policy. Aliases resolve before dispatch. */
 const COMMAND_REGISTRY = [
   { name: 'up', aliases: [], tier: 'primary', needsInstalledConfig: false },
+  { name: 'ui', aliases: [], tier: 'primary', needsInstalledConfig: true },
   { name: 'onboard', aliases: [], tier: 'advanced', needsInstalledConfig: false },
   { name: 'down', aliases: [], tier: 'primary', needsInstalledConfig: true },
   { name: 'doctor', aliases: [], tier: 'primary', needsInstalledConfig: true },
@@ -86,6 +92,7 @@ const COMMAND_REGISTRY = [
   { name: 'suspend', aliases: [], tier: 'primary', needsInstalledConfig: true },
   { name: 'resume', aliases: [], tier: 'primary', needsInstalledConfig: true },
   { name: 'route', aliases: ['routing'], tier: 'primary', needsInstalledConfig: true },
+  { name: 'fleet', aliases: ['profiles'], tier: 'primary', needsInstalledConfig: true },
   { name: 'integrate', aliases: [], tier: 'primary', needsInstalledConfig: true },
   { name: 'integrations', aliases: [], tier: 'primary', needsInstalledConfig: true },
   { name: 'add-model', aliases: ['model-add'], tier: 'primary', needsInstalledConfig: true },
@@ -147,8 +154,11 @@ function usage() {
   return `LLooM — local-first LLM gateway
 
 Primary commands:
-  lloom / lloom up                 Start installed LLooM; preview setup on first run
+  lloom / lloom up                 Start installed LLooM; guided setup on first run
   lloom up --go                    First run: install, integrate, and start the model
+  lloom up --browser               Open guided local setup on first run
+  lloom up --no-browser            Keep setup in the terminal
+  lloom ui                         Open the installed gateway dashboard
   lloom down                       Stop the gateway and all managed model backends
   lloom doctor                     Readiness report (blockers, warnings, next actions)
   lloom serve                      Run the gateway (reads ~/.lloom/config.json)
@@ -301,6 +311,7 @@ const INSTALLED_CONFIG_COMMANDS = new Set([
   'keep-warm',
   'model-add',
   'models',
+  'fleet',
   'remove-model',
   'route',
   'routing',
@@ -331,6 +342,7 @@ const OPERATIONAL_CONFIG_COMMANDS = new Set([
   'keep-warm',
   'model-add',
   'models',
+  'fleet',
   'remove-model',
   'route',
   'routing',
@@ -1072,7 +1084,7 @@ async function gatewayRequest(config, pathname, { method = 'GET', body, timeoutM
     bodyTimeout: timeoutMs
   });
   try {
-    const response = await fetch(`${gatewayUrlFor(config)}${pathname}`, {
+    const response = await undiciFetch(`${gatewayUrlFor(config)}${pathname}`, {
       method,
       headers: {
         ...gatewayAdminHeaders(config),
@@ -1118,15 +1130,22 @@ function gatewayProcessPaths(configPath) {
   };
 }
 
-async function startGatewayBackground(configPath) {
+async function startGatewayBackground(configPath, { requireOwned = false } = {}) {
   const gatewayConfig = await loadConfig(configPath);
   const url = gatewayUrlFor(gatewayConfig);
   const alreadyHealthy = await gatewayHealth(gatewayConfig);
   const paths = gatewayProcessPaths(configPath);
   if (alreadyHealthy?.ok) {
+    const ownedPid =
+      requireOwned && existsSync(paths.pidPath) ? Number(readFileSync(paths.pidPath, 'utf8').trim()) : null;
+    if (requireOwned && (!ownedPid || alreadyHealthy.pid !== ownedPid))
+      throw new Error(
+        'The gateway port is already used by another server. Choose another port or stop that server before retrying setup.'
+      );
     return {
       status: 'already-running',
       url,
+      pid: alreadyHealthy.pid,
       health: alreadyHealthy,
       ...paths
     };
@@ -1158,6 +1177,8 @@ async function startGatewayBackground(configPath) {
       ...paths
     };
   }
+  if (requireOwned && health.pid !== child.pid)
+    throw new Error('Another server answered on the gateway port. The installed gateway has not been verified.');
   return {
     status: 'started',
     url,
@@ -1448,6 +1469,24 @@ async function main() {
     },
     up: async (context) => {
       const { args, configPath } = context;
+      if (
+        shouldOpenSetup(args, {
+          installed: Boolean(configPath),
+          interactive: Boolean(process.stdin.isTTY && process.stdout.isTTY)
+        }) &&
+        (!wantsOnboardingPlan(args) || hasFlag(args, '--browser'))
+      ) {
+        const options = await firstRunCliOptions([...args, '--offline']);
+        const setup = createBrowserSetup(context.config, options, {
+          gatewayStarter: (configPath) => startGatewayBackground(configPath, { requireOwned: true })
+        });
+        await setup.listen();
+        const url = setup.bootstrapUrl();
+        console.log('LLooM setup: ' + url + '\nKeep this terminal open while setup runs. Press Ctrl+C to stop.');
+        const opened = await openLocalBrowser(url);
+        if (!opened) console.log('Open the local setup URL above in your browser.');
+        return;
+      }
       if (!configPath || wantsOnboardingPlan(args)) return handlers.onboard(context);
       const gateway = await startGatewayBackground(configPath);
       const report = {
@@ -1473,6 +1512,16 @@ async function main() {
       }
       if (!report.ok) process.exitCode = 1;
     },
+    ui: async ({ args, config }) => {
+      const url = gatewayUrlFor(config);
+      if (wantsJson(args)) {
+        console.log(JSON.stringify({ url }));
+        return;
+      }
+      console.log('LLooM: ' + url);
+      if (!hasFlag(args, '--no-browser') && !(await openLocalBrowser(url)))
+        console.log('Open the gateway URL above in your browser.');
+    },
     onboard: async ({ args, config, command: _command }) => {
       const go = wantsGo(args);
       const apply = hasFlag(args, '--apply') || go;
@@ -2351,6 +2400,59 @@ async function main() {
       if (!result) throw new Error(`route switch failed through ${gatewayUrlFor(config)}`);
       console.log(JSON.stringify({ ...result, applied: true }, null, 2));
     },
+    fleet: async ({ args, config }) => {
+      const action = positional(args)[1] ?? 'list';
+      const name = positional(args)[2];
+      const apply = hasFlag(args, '--apply');
+      const yes = hasFlag(args, '--yes');
+      if (!['list', 'show', 'use', 'save'].includes(action)) {
+        throw new Error(`Unknown fleet action ${action}; use list, show, use, or save.`);
+      }
+      if (action === 'list') {
+        const result = await gatewayRequest(config, '/gateway/fleet/profiles', { timeoutMs: 10000 });
+        if (!result) throw new Error(`LLooM gateway at ${gatewayUrlFor(config)} is not reachable`);
+        console.log(JSON.stringify(result, null, 2));
+        return;
+      }
+      if (!name) throw new Error(`Missing profile name for fleet ${action}.`);
+      if (action === 'show') {
+        const result = await gatewayRequest(config, `/gateway/fleet/profiles/${encodeURIComponent(name)}`, {
+          timeoutMs: 10000
+        });
+        if (!result) throw new Error(`LLooM gateway at ${gatewayUrlFor(config)} is not reachable`);
+        console.log(JSON.stringify(result, null, 2));
+        return;
+      }
+      const isApply = action === 'use';
+      const plan = isApply
+        ? { action: 'use', profile: name, applied: false, next: `lloom fleet use ${name} --apply --yes` }
+        : {
+            action: 'save',
+            profile: name,
+            description: argValue(args, '--description') ?? '',
+            overwrite: hasFlag(args, '--overwrite'),
+            applied: false,
+            next: `lloom fleet save ${name}${hasFlag(args, '--overwrite') ? ' --overwrite' : ''} --apply --yes`
+          };
+      if (!apply) {
+        console.log(JSON.stringify(plan, null, 2));
+        return;
+      }
+      if (!yes) throw new Error(`Refusing to ${action} a fleet profile without --yes after reviewing the plan`);
+      const result = await gatewayRequest(
+        config,
+        `/gateway/fleet/profiles/${encodeURIComponent(name)}${isApply ? '?apply=1' : ''}`,
+        {
+          method: 'POST',
+          body: isApply
+            ? { yes: true }
+            : { yes: true, description: plan.description, overwrite: plan.overwrite },
+          timeoutMs: 60000
+        }
+      );
+      if (!result) throw new Error(`fleet profile ${action} failed through ${gatewayUrlFor(config)}`);
+      console.log(JSON.stringify({ ...result, applied: true }, null, 2));
+    },
     runtimes: async ({ args, config, command: _command }) => {
       const runtimeId = positional(args)[1] ?? 'all';
       const manager = runtimeManagerForCli(config);
@@ -2377,18 +2479,50 @@ async function main() {
     cluster: async ({ args, config, command: _command }) => {
       const action = positional(args)[1] ?? 'status';
       if (action === 'discover') {
-        const discovery = await detectNvidiaSyncCluster();
+        const explicitLocalId =
+          process.env.LLOOM_NODE_ID ??
+          config.cluster?.nodeId ??
+          (Object.keys(config.cluster?.nodes ?? {}).length === 1 ? Object.keys(config.cluster.nodes)[0] : undefined);
+        if (process.env.LLOOM_NODE_ID && config.cluster?.nodeId && process.env.LLOOM_NODE_ID !== config.cluster.nodeId)
+          throw new Error('LLOOM_NODE_ID conflicts with configured cluster.nodeId');
+        const discovery = await detectNvidiaSyncCluster({ localNodeId: explicitLocalId });
         if (!discovery) throw new Error('No NVIDIA Sync cluster was detected from this node');
-        const cluster = nvidiaSyncClusterConfig(discovery, {
+        const options = {
+          discovery,
+          localNodeId: discovery.nodeId,
           id: argValue(args, '--id'),
-          apiKeyEnv: argValue(args, '--api-key-env') ?? 'LLOOM_CLUSTER_KEY'
-        });
+          apiKeyEnv: argValue(args, '--api-key-env')
+        };
+        // Keep authoring forms such as environment references out of expansion.
+        const source = JSON.parse(readFileSync(config.sourcePath, 'utf8'));
+        const plan = (raw) => mergeNvidiaSyncClusterDiscovery({ ...options, cluster: raw.cluster });
+        let cluster = plan(source);
+        const validationErrors = validateClusterConfig({ ...config, cluster });
+        let changed = false;
         if (hasFlag(args, '--apply')) {
-          const source = JSON.parse(readFileSync(config.sourcePath, 'utf8'));
-          source.cluster = cluster;
-          writeFileSync(config.sourcePath, `${JSON.stringify(source, null, 2)}\n`, { mode: 0o600 });
+          if (validationErrors.length) throw new Error(`Invalid discovered cluster: ${validationErrors.join('; ')}`);
+          ({ changed } = await mutateConfigSource(config, (latest) => {
+            // Replan from the current raw source inside the serialized mutation.
+            cluster = plan(latest);
+            latest.cluster = cluster;
+          }));
         }
-        console.log(JSON.stringify({ ok: true, applied: hasFlag(args, '--apply'), discovery, cluster }, null, 2));
+        console.log(
+          JSON.stringify(
+            {
+              ok: validationErrors.length === 0,
+              applied: hasFlag(args, '--apply'),
+              changed,
+              discovery,
+              summary: nvidiaSyncDiscoverySummary(discovery, { currentConfig: config }),
+              validationErrors,
+              cluster,
+              next: 'Discovery does not change model placement or gateway listeners. Verify endpoints before serving.'
+            },
+            null,
+            2
+          )
+        );
         return;
       }
       if (action === 'add-node') {
@@ -2714,7 +2848,7 @@ async function main() {
   handlers['pack-export'] = handlers['recipe-export'];
   handlers['recipe-pack'] = handlers['recipe-import'];
   handlers['pack-submit'] = handlers['recipe-submit'];
-  handlers['model-add'] = handlers['add-model'];
+  handlers['profiles'] = handlers['fleet'];
   handlers['routing'] = handlers['route'];
   handlers['runtime-status'] = handlers['runtimes'];
   handlers['cluster-status'] = handlers['cluster'];
diff --git a/config/default.json b/config/default.json
index e84e8de..feccb4d 100644
--- a/config/default.json
+++ b/config/default.json
@@ -31,7 +31,12 @@
     "enabled": true,
     "autoEvict": false,
     "reserveMemoryGb": 12,
-    "protectActiveRequests": true
+    "protectActiveRequests": true,
+    "memorySafety": {
+      "mode": "enforce",
+      "maxMemoryUtilization": 0.9,
+      "pollIntervalMs": 250
+    }
   },
   "community": {
     "hostUrl": "http://127.0.0.1:8110",
diff --git a/docs/architecture.md b/docs/architecture.md
index 52bfe22..08c5ac2 100644
--- a/docs/architecture.md
+++ b/docs/architecture.md
@@ -48,6 +48,31 @@ tool call has no reasoning to replay; a new user turn permits thinking again.
 This opt-in compatibility setting preserves tool constraints without inventing
 reasoning history. Ordinary calls outside those cases keep their thinking settings.
 
+### OpenRouter provider restriction
+
+Set `openrouterProvider` on a dedicated OpenRouter backend to restrict its
+upstream providers across Chat Completions, Responses, and Anthropic Messages,
+including streaming requests:
+
+```json
+{
+  "type": "openai",
+  "baseUrl": "https://openrouter.ai/api/v1",
+  "apiKeyEnv": "OPENROUTER_API_KEY",
+  "openrouterProvider": { "only": ["z-ai"], "allow_fallbacks": false }
+}
+```
+
+Configured provider fields override client requests. Chat Completions retains
+other client provider preferences; the Responses and Anthropic bridges retain
+only the fields supported by their translators. `only` must contain at least one nonempty provider
+slug; `allow_fallbacks` defaults to false. Invalid policy objects fail before
+an upstream request is sent. The policy applies only to the exact
+`openrouter.ai` host. With the configuration above, an unavailable Z.ai endpoint
+returns an error instead of switching to another provider. Models sharing this
+backend share its restriction; LLooM alias fallback rules remain separate.
+See [OpenRouter provider routing](https://openrouter.ai/docs/guides/routing/provider-selection).
+
 ## Security Defaults
 
 | Setting                         | Default                                | Meaning                                                                                           |
@@ -318,3 +343,12 @@ The current repository includes a static `lloom-host` development server that se
 Setup composes initialization, backend setup, recipe setup, generated clients, and client integration writes into one audited plan. It does not bypass the lower-level safety gates: dry-run is the default, and real execution requires explicit `--apply --yes`. Bootstrap remains the lower-level backend/model/client phase for an existing config.
 
 The default generated gateway port is `8100`; selected backend runtimes occupy the default backend range beginning at `8201`. `setup --port` and `setup --backend-port-range` retarget the generated provider URL, backend base URLs, runtime ports, health URLs, and warmup URLs together so custom port layouts remain internally consistent.
+
+
+### Dashboard memory map
+
+The Models view divides one machine’s physical memory into running backends, system/other applications, and available space. Protected headroom is an overlay within available memory, not additional capacity. Hover, keyboard focus, and selection preview the incoming model’s configured peak estimate against that machine’s live memory and safety reserve. These interactions are read-only. Normal inference prepares the model through admission; optional manual preparation uses the same guarded admission endpoint.
+
+Dashboard and node-status requests opt into passive memory attribution. A five-second shared cache reads process memory and loopback listeners with bounded subprocesses. Overlapping process trees are counted once; Docker PIDs are used only on Linux. Known local Ollama and LLooM audio servers also expose their loaded-model lists through bounded, read-only requests. A healthy server with an empty model cache is not a resident model. Routing and admission status do not request this dashboard sampling.
+
+On macOS, an optional bounded Python helper reads Darwin’s physical-footprint accounting, which includes charged Metal and compressed memory. If unavailable, attribution falls back to RSS. Process memory is approximate and does not describe all shared or cached allocations. Unmeasured running backends use visibly marked estimates. If attribution exceeds measured host use, blocks are reconciled to that measured total. The host’s available-memory reading remains authoritative. Remote nodes use their own observations and safety policy; distributed capacity is never pooled into a single fit promise. Preview estimates never override admission or the live memory safety guard.
diff --git a/docs/browser-experience.md b/docs/browser-experience.md
new file mode 100644
index 0000000..c0ca77e
--- /dev/null
+++ b/docs/browser-experience.md
@@ -0,0 +1,48 @@
+# The LLooM browser experience
+
+LLooM chooses and operates vendor recipes. Recipe vendors own model execution, backend tuning, and distributed execution contracts. The dashboard presents the hardware, installation, readiness, and serving decisions that belong to LLooM.
+
+## First run
+
+After installing the CLI from the repository, run `lloom` in an interactive terminal. With no installed configuration, a loopback setup server opens a browser. Choose Chat & write, Code, Images, or Voice; review the compatible recipe and installation details; then choose **Set up my AI**.
+
+The setup page uses the bundled recipe library. Unsupported hardware or workloads produce an explicit explanation. Unknown download sizes, credentials, or license information stay unknown. A recipe metadata license does not cover the model weights.
+
+Installation runs as a background job in the setup process. Refreshing the page resumes its progress. Keep the terminal open. A failed bootstrap can retry only the exact, unchanged configuration created by that session, within the original 30-minute review window. After a process restart, use `lloom bootstrap --apply --yes` or the installed dashboard to continue. Setup never silently replaces another configuration.
+
+After installation, LLooM starts the gateway and checks a chat request through its normal inference API. Health alone does not mean the model is ready. Media recipes require checking their actual output; the page identifies that remaining step.
+
+`lloom --no-browser`, `lloom onboard`, and `lloom up --go` preserve the CLI paths. JSON, offline, and explicit recipe flags do not unexpectedly launch a browser. `lloom ui` opens an installed gateway.
+
+## Daily use
+
+- **Live** shows clients, gateway nodes, models, and observed traffic. A luminous gateway connects client cards to model rows grouped by physical machine. Active requests drive the light trails. Follow activity focuses the scene on serving models. Detailed topology retains the diagnostic canvas and its manual camera controls. Both views honor reduced motion.
+- **Models** searches and filters the actual gateway catalog. The inspector offers Load, Warm up, Unload, readiness policy, and a small chat trial. Runtime details remain available in a disclosure.
+- **Add model** reviews a vendor recipe or custom model reference before starting a background installation. Plan IDs bind the reviewed input, recipe/backend catalog, and configuration version. Concurrent or stale changes fail visibly. Custom Hugging Face imports require an immutable commit link and publish only after staged acquisition verification. Downloads remain reusable. The new installation API does not accept arbitrary config paths or shell commands.
+- **Machines** shows configured physical nodes and their individual memory readings. Memory is never presented as one interchangeable pool. Distributed model execution requires a compatible vendor recipe.
+- **Clients** supplies the gateway base URL, exact model IDs, an example request, and CLI integration commands. Admin credentials do not belong in client applications.
+- **Settings** retains the advanced recipe, backend, runtime, and setup tools.
+
+Readiness policies express intent: **Auto** loads on demand, **Prefer ready** uses the existing idle residency reconciler, and **Always ready** prevents automatic eviction. All loading still passes through memory admission. Changing readiness does not restart a model or interrupt active work. The API returns a pending job while the current admission completes; the page reports completion or failure. Queued residency starts recheck the saved policy, and pending hard pins protect eviction victims. Use Load when you want to start a cold model immediately.
+
+### Memory protection
+
+Admission counts live host memory use, including other applications, even when the policy specifies only a reserve or an absolute budget. Model estimates are planning inputs, not allocation limits.
+
+Memory protection is enabled by default. A newly started backend is checked before launch and monitored during loading and warmup. If available memory reaches the hard reserve or host utilization reaches the ceiling, LLooM aborts that load, cleans up its processes, and reports the failed threshold. Ordinary Load, forced starts, and disabling automatic eviction do not disable this protection. Automatic retries are blocked until a manual retry or gateway restart. Suspend a model to keep it blocked across restarts.
+
+The installed config accepts `runtimePolicy.memorySafety` with `mode`, `minAvailableMemoryGb`, `maxMemoryUtilization` (a fraction), and `pollIntervalMs`. The normal mode is `"enforce"`. The default ceiling is 90%; the host reserve can impose a stricter limit. Sampling is a userspace safeguard, not a kernel-enforced allocation quota: a backend can allocate between samples.
+
+For deliberate manual experiments, `"mode": "yolo"` disables memory admission and the hard load guard. It leaves authentication, runtime ownership, and maintenance gates in place. The dashboard displays a persistent YOLO warning. Restore `"enforce"` before normal operation; YOLO can exhaust the host's memory.
+
+## Local security
+
+First-run setup binds only to loopback and uses a random session token passed through the URL fragment. It removes the fragment immediately and keeps the token in that tab's session storage. Setup rejects foreign origins, alternate authorities, oversized bodies, and unreviewed apply inputs. Configuration publication cannot overwrite a file created concurrently.
+
+The installed dashboard rejects cross-origin management requests and DNS-rebound loopback authorities. Its page cannot be framed. Remote management writes require both explicit remote-admin enablement and a configured admin key; inference keys alone cannot authorize them. Inference compatibility is unchanged. Local same-user processes remain inside the local trust boundary.
+
+Installation jobs and reviewed plan IDs live in the gateway process. Browser refresh is supported; a gateway restart requires a fresh plan. Existing installer stage state and downloaded files support subsequent retries. The config mutation store detects concurrent writes but is not a cross-process transaction lock.
+
+## Remaining product work
+
+The repository is still the distribution source; a signed single-command installer and published npm package are not part of this change. Nearby discovery and approval-based pairing between independently installed gateways need an authenticated node identity protocol. The Machines page does not invent peers or automatically grant access. Existing configured federation remains available through the CLI and gateway.
diff --git a/docs/clusters.md b/docs/clusters.md
index 44b7d8f..c9b9652 100644
--- a/docs/clusters.md
+++ b/docs/clusters.md
@@ -37,9 +37,7 @@ The declarative form is useful for review or hand editing:
         "labels": { "role": "node", "architecture": "darwin-arm64", "accelerator": "metal" },
         "resources": { "memoryGb": 64 },
         "proxy": {
-          "models": [
-            { "id": "local/qwen", "as": "macbook-local/local/qwen", "kind": "chat", "remoteRuntime": "qwen" }
-          ]
+          "models": [{ "id": "local/qwen", "as": "macbook-local/local/qwen", "kind": "chat", "remoteRuntime": "qwen" }]
         }
       }
     }
@@ -60,7 +58,13 @@ lloom cluster discover --id ennspark-cluster
 lloom cluster discover --id ennspark-cluster --api-key-env LLOOM_ADMIN_API_KEY --apply
 ```
 
-Discovery reads only NVIDIA Sync-marked SSH entries and local interface addresses. It records the physical hostnames as metadata but uses the stable Sync/Tailscale aliases (with `-lan` removed) as node IDs. `lloom profile` includes the detected topology, so `lloom select` can rank exact-size cluster recipes before cluster configuration is applied.
+Discovery groups `NVSyncClusterAlias` entries into stable node IDs, including duplicate IP aliases and multiple rails per peer. Older Sync entries with a named `Host` alias still work. The configured local identity and leader remain unchanged. Local observations appear under `cluster.discovery.links`; they do not prove gateway reachability, bandwidth, or a complete ring.
+
+New peers remain physical inventory under `cluster.discovery.nodes`. Join a peer with `lloom cluster add-node <id> <authenticated-gateway-url> --apply` after its gateway is reachable. Discovery does not create a live gateway endpoint from an SSH address.
+
+`--apply` merges into the raw configuration and validates it before an atomic write. It preserves existing federation nodes, model catalogs, authentication references, custom endpoints, and model placements. An endpoint that names its previous `10.100.*` backend host moves to the observed address while retaining its scheme, port, and path. Other endpoints remain unchanged and receive a diagnostic. Discovery does not change gateway listeners or NCCL settings; verify new endpoints before using them.
+
+Physical membership and model placement are separate. In a three-Spark ring, a model can run on one node while a distributed model uses an explicit two-node subset. A three-node member list is also representable, but TP-3 requires a compatible model architecture and backend launcher; discovery does not infer that compatibility or repartition a loaded model. Resource admission applies to the chosen members. Select the NCCL adapters that connect those members: a TP-2 job must not use adapters whose cables lead to a third node outside its placement.
 
 Use private fabric addresses for node-to-node LLooM and raw backend traffic. `backendHost` is deliberately required when a recipe auto-materializes replicas; LLooM binds the generated backend only to that address, never to every interface.
 
@@ -131,14 +135,16 @@ The equivalent explicit config is:
 
 ```json
 {
-  "models": [{
-    "id": "example/Qwen",
-    "targets": [
-      { "id": "spark-1", "node": "spark-1", "backend": "qwen-spark-1", "runtime": "qwen-spark-1" },
-      { "id": "spark-2", "node": "spark-2", "backend": "qwen-spark-2", "runtime": "qwen-spark-2" },
-      { "id": "spark-2-proxy", "node": "spark-2", "backend": "lloom-node-spark-2", "remoteRuntime": "qwen-spark-2" }
-    ]
-  }]
+  "models": [
+    {
+      "id": "example/Qwen",
+      "targets": [
+        { "id": "spark-1", "node": "spark-1", "backend": "qwen-spark-1", "runtime": "qwen-spark-1" },
+        { "id": "spark-2", "node": "spark-2", "backend": "qwen-spark-2", "runtime": "qwen-spark-2" },
+        { "id": "spark-2-proxy", "node": "spark-2", "backend": "lloom-node-spark-2", "remoteRuntime": "qwen-spark-2" }
+      ]
+    }
+  ]
 }
 ```
 
@@ -188,19 +194,39 @@ Explicit config remains supported:
       "placement": {
         "mode": "distributed",
         "members": [
-          { "node": "spark-1", "runtime": "dsv4-ray-head", "role": "head", "order": 10, "resources": { "memoryGb": 6 } },
-          { "node": "spark-2", "runtime": "dsv4-ray-worker-2", "role": "worker", "order": 20, "resources": { "memoryGb": 54 } },
-          { "node": "spark-1", "runtime": "dsv4-vllm-server", "role": "server", "order": 30, "resources": { "memoryGb": 54 } }
+          {
+            "node": "spark-1",
+            "runtime": "dsv4-ray-head",
+            "role": "head",
+            "order": 10,
+            "resources": { "memoryGb": 6 }
+          },
+          {
+            "node": "spark-2",
+            "runtime": "dsv4-ray-worker-2",
+            "role": "worker",
+            "order": 20,
+            "resources": { "memoryGb": 54 }
+          },
+          {
+            "node": "spark-1",
+            "runtime": "dsv4-vllm-server",
+            "role": "server",
+            "order": 30,
+            "resources": { "memoryGb": 54 }
+          }
         ]
       }
     }
   },
-  "models": [{
-    "id": "deepseek-ai/DSv4Flash",
-    "backend": "dsv4flash",
-    "runtime": "dsv4flash-cluster",
-    "upstreamModel": "deepseek-ai/DSv4Flash"
-  }]
+  "models": [
+    {
+      "id": "deepseek-ai/DSv4Flash",
+      "backend": "dsv4flash",
+      "runtime": "dsv4flash-cluster",
+      "upstreamModel": "deepseek-ai/DSv4Flash"
+    }
+  ]
 }
 ```
 
diff --git a/package-lock.json b/package-lock.json
index 31f9a55..9bba72a 100644
--- a/package-lock.json
+++ b/package-lock.json
@@ -9,7 +9,7 @@
       "version": "0.2.3",
       "license": "MIT",
       "dependencies": {
-        "undici": "^7.28.0"
+        "undici": "^8.10.2"
       },
       "bin": {
         "lloom": "bin/lloom.mjs",
@@ -17,8 +17,8 @@
       },
       "devDependencies": {
         "@eslint/js": "^10.0.1",
-        "eslint": "^10.10.0",
-        "prettier": "^3.9.6"
+        "eslint": "^10.11.0",
+        "prettier": "^3.9.8"
       },
       "engines": {
         "node": ">=20.0.0"
@@ -418,9 +418,9 @@
       }
     },
     "node_modules/eslint": {
-      "version": "10.10.0",
-      "resolved": "https://registry.npmjs.org/eslint/-/eslint-10.10.0.tgz",
-      "integrity": "sha512-NPXn6r5zl4uET1DAVPaOwzX3rut4c0wcmw3dWJAfOsTM5+TogXo0DDjz8pwm/hL8cyVNpHqeK4JpN0NjnyFFNw==",
+      "version": "10.11.0",
+      "resolved": "https://registry.npmjs.org/eslint/-/eslint-10.11.0.tgz",
+      "integrity": "sha512-P7a6UEEqb9G95MYAtqkmsTbVXIYyzIfl6NGOIJk162PaahFxFyeGcrlXYFSiagECg4sEm8IseJdZBKR3rx6MsQ==",
       "dev": true,
       "license": "MIT",
       "workspaces": [
@@ -887,9 +887,9 @@
       }
     },
     "node_modules/prettier": {
-      "version": "3.9.6",
-      "resolved": "https://registry.npmjs.org/prettier/-/prettier-3.9.6.tgz",
-      "integrity": "sha512-OpN0zzVdiaiAhxpuuj5efpIS4sY9j7bY6uR5mnj5yPzGkdkjNKSJeUThPb60Jw29QuAZgA4o+/iB49kFiaBX6g==",
+      "version": "3.9.8",
+      "resolved": "https://registry.npmjs.org/prettier/-/prettier-3.9.8.tgz",
+      "integrity": "sha512-WRFq3Wn3WId7LLROfMLdH7xaFr2jR62wU8nLO6rQUOLOxNZUviyJQs1M0iIhLexSFy+L+w0ch66wtoO2jRjG0A==",
       "dev": true,
       "license": "MIT",
       "bin": {
@@ -969,12 +969,12 @@
       }
     },
     "node_modules/undici": {
-      "version": "7.29.0",
-      "resolved": "https://registry.npmjs.org/undici/-/undici-7.29.0.tgz",
-      "integrity": "sha512-IDxfleLmmbSskfWSUATiN1nfn2rDuvnMOqb5CWR92iIfojA0Ud+ulOAAEQ57LPr9rWmsreUyf5lwyao+7GNNVw==",
+      "version": "8.10.2",
+      "resolved": "https://registry.npmjs.org/undici/-/undici-8.10.2.tgz",
+      "integrity": "sha512-/y4/bH9YNU5hi9NIrpOuvGXFcxrj3CMrV+/AYpowAYTpHn8gX/XPFjNy766FPoYY0miQhdW977JFWKGNhBdwyQ==",
       "license": "MIT",
       "engines": {
-        "node": ">=20.18.1"
+        "node": ">=22.19.0"
       }
     },
     "node_modules/uri-js": {
diff --git a/package.json b/package.json
index 2df4fe7..ea65903 100644
--- a/package.json
+++ b/package.json
@@ -64,8 +64,8 @@
     "check": "npm run check:cluster && npm run check:js && npm run check:entity-stagger && npm run check:python && npm run check:shell && npm run check:spark-deploy && npm run check:workers",
     "check:cluster": "node --check src/cluster.mjs && node --check test/cluster.test.mjs",
     "check:entity-stagger": "node --check scripts/entity-stagger-benchmark.mjs && node --check test/entity-stagger-benchmark.test.mjs",
-    "check:js": "node --check bin/lloom.mjs && node --check bin/lloom-host.mjs && node --check src/backend-catalog.mjs && node --check src/benchmarks.mjs && node --check src/bootstrap.mjs && node --check src/client-integrations.mjs && node --check src/community-client.mjs && node --check src/community-site.mjs && node --check src/config.mjs && node --check src/dashboard.mjs && node --check src/doctor.mjs && node --check src/host-memory.mjs && node --check src/host-server.mjs && node --check src/init.mjs && node --check src/installer.mjs && node --check src/interchange.mjs && node --check src/machine-profile.mjs && node --check src/managed-environment.mjs && node --check src/model-acquisition.mjs && node --check src/model-intake.mjs && node --check src/model-removal.mjs && node --check src/onboarding.mjs && node --check src/recipe-index.mjs && node --check src/recipe-pack-export.mjs && node --check src/recipe-pack.mjs && node --check src/registry.mjs && node --check src/route-control.mjs && node --check src/recipes.mjs && node --check src/resource-fit.mjs && node --check src/runtime-manager.mjs && node --check src/runtime-policy.mjs && node --check src/server.mjs && node --check src/setup.mjs && node --check src/setup-status.mjs && node --check src/process-control.mjs && node --check src/security.mjs && node --check src/tts-catalog.mjs && node --check src/voice-profiles.mjs && node --check src/protocol/text.mjs && node --check src/protocol/reasoning-normalize.mjs && node --check src/protocol/reasoning-effort.mjs && node --check src/protocol/responses.mjs && node --check src/protocol/anthropic.mjs && node --check src/protocol/sse.mjs && node --check src/protocol/stream-anthropic.mjs && node --check src/protocol/stream-responses.mjs && node --check src/protocol/index.mjs && node --check scripts/apply-spark-route-catalog.mjs && node --check scripts/capture-live-metrics.mjs && node --check scripts/merge-live-metrics.mjs && node --check scripts/run-qualitative-chat-benchmark.mjs && node --check scripts/run-local-decision-benchmark.mjs && node --check scripts/check-community-deploy.mjs && node --check scripts/generate-client-configs.mjs && node --check scripts/check-interchange.mjs && node --check scripts/check-package.mjs && node --check scripts/release-lib.mjs && node --check scripts/build-release.mjs && node --check scripts/deploy-spark.mjs && node --check test/audio-generations.test.mjs && node --check test/community-host.test.mjs && node --check test/dashboard-status.test.mjs && node --check test/ds4fv-recipe.test.mjs && node --check test/dspark-hotfixes.test.mjs && node --check test/glm53-exl3-recipe.test.mjs && node --check test/host-memory.test.mjs && node --check test/metrics-persistence.test.mjs && node --check test/model-acquisition.test.mjs && node --check test/model-failover.test.mjs && node --check test/qwen38-vllm-recipe.test.mjs && node --check test/reasoning-normalize.test.mjs && node --check test/reasoning-effort.test.mjs && node --check test/resource-fit.test.mjs && node --check test/route-control.test.mjs && node --check test/spark-route-catalog.test.mjs && node --check test/runtime-watchdog.test.mjs && node --check test/tts-catalog.test.mjs && node --check test/voice-profiles.test.mjs && node --check test/chatterbox-recipe.test.mjs && node --check test/hear-recipe.test.mjs && node --check test/smoke.mjs && node --check src/runtime-supervisor.mjs && node --check test/runtime-supervisor.test.mjs && node --check test/comfyui-media.test.mjs && node --check src/runtime-residency.mjs",
-    "check:python": "python3 -m py_compile backends/dspark-vllm/apply-patch-pack.py backends/dspark-vllm/packs/miaai-dsv4flash-d1b76251-defaults/patches/*.py backends/dspark-vllm/packs/miaai-ds4fv-f5665e8/patches/*.py backends/dspark-vllm/packs/miaai-ds4fv-f5665e8/patches/vision_exp/*.py backends/qwen38-vllm/apply-ple-fp8-patch.py backends/qwen38-sglang/*.py backends/glm53-exl3/*.py backends/mlx-audio/lloom_audio_server.py backends/hear/lloom_hear_server.py backends/chatterbox/lloom_chatterbox_server.py patches/apply_mtplx_longctx_fix.py backends/comfyui-media/bridge/*.py backends/comfyui-media/graphs/*.py backends/comfyui-media/build/*.py backends/comfyui-media/install.py backends/qwen-image-diffusers/*.py backends/hear/test/test_hear_security.py",
+    "check:js": "node --check bin/lloom.mjs && node --check bin/lloom-host.mjs && node --check src/backend-catalog.mjs && node --check src/benchmarks.mjs && node --check src/bootstrap.mjs && node --check src/client-integrations.mjs && node --check src/community-client.mjs && node --check src/community-site.mjs && node --check src/config.mjs && node --check src/dashboard.mjs && node --check src/doctor.mjs && node --check src/host-memory.mjs && node --check src/host-server.mjs && node --check src/init.mjs && node --check src/installer.mjs && node --check src/interchange.mjs && node --check src/machine-profile.mjs && node --check src/managed-environment.mjs && node --check src/model-acquisition.mjs && node --check src/model-intake.mjs && node --check src/model-removal.mjs && node --check src/onboarding.mjs && node --check src/recipe-index.mjs && node --check src/recipe-pack-export.mjs && node --check src/recipe-pack.mjs && node --check src/registry.mjs && node --check src/route-control.mjs && node --check src/recipes.mjs && node --check src/resource-fit.mjs && node --check src/runtime-manager.mjs && node --check src/runtime-policy.mjs && node --check src/server.mjs && node --check src/setup.mjs && node --check src/setup-status.mjs && node --check src/process-control.mjs && node --check src/security.mjs && node --check src/tts-catalog.mjs && node --check src/voice-profiles.mjs && node --check src/protocol/text.mjs && node --check src/protocol/reasoning-normalize.mjs && node --check src/protocol/reasoning-effort.mjs && node --check src/protocol/responses.mjs && node --check src/protocol/anthropic.mjs && node --check src/protocol/sse.mjs && node --check src/protocol/stream-anthropic.mjs && node --check src/protocol/stream-responses.mjs && node --check src/protocol/index.mjs && node --check scripts/apply-spark-route-catalog.mjs && node --check scripts/capture-live-metrics.mjs && node --check scripts/merge-live-metrics.mjs && node --check scripts/run-qualitative-chat-benchmark.mjs && node --check scripts/run-local-decision-benchmark.mjs && node --check scripts/check-community-deploy.mjs && node --check scripts/generate-client-configs.mjs && node --check scripts/check-interchange.mjs && node --check scripts/check-package.mjs && node --check scripts/release-lib.mjs && node --check scripts/build-release.mjs && node --check scripts/deploy-spark.mjs && node --check test/audio-generations.test.mjs && node --check test/community-host.test.mjs && node --check test/dashboard-status.test.mjs && node --check test/ds4fv-recipe.test.mjs && node --check test/dspark-hotfixes.test.mjs && node --check test/glm53-exl3-recipe.test.mjs && node --check test/host-memory.test.mjs && node --check test/metrics-persistence.test.mjs && node --check test/model-acquisition.test.mjs && node --check test/model-failover.test.mjs && node --check test/qwen38-vllm-recipe.test.mjs && node --check test/reasoning-normalize.test.mjs && node --check test/reasoning-effort.test.mjs && node --check test/resource-fit.test.mjs && node --check test/route-control.test.mjs && node --check test/spark-route-catalog.test.mjs && node --check test/runtime-watchdog.test.mjs && node --check test/tts-catalog.test.mjs && node --check test/voice-profiles.test.mjs && node --check test/chatterbox-recipe.test.mjs && node --check test/hear-recipe.test.mjs && node --check test/smoke.mjs && node --check src/runtime-supervisor.mjs && node --check test/runtime-supervisor.test.mjs && node --check test/comfyui-media.test.mjs && node --check src/runtime-residency.mjs && node --check src/dashboard-scene.mjs && node --check src/dashboard-presence.mjs && node --check src/dashboard-presence-client.mjs && node --check src/first-run.mjs && node --check src/first-run-page.mjs && node --check src/browser-setup.mjs && node --check src/runtime-preferences.mjs && node --check src/installation-jobs.mjs && node --check src/model-installation.mjs && node --check src/dashboard-installation.mjs && node --check src/runtime-memory-safety.mjs && node --check src/dashboard-memory.mjs && node --check src/runtime-memory-usage.mjs && node --check src/config-profiles.mjs && node --check test/config-profiles.test.mjs",
+    "check:python": "python3 -m py_compile backends/dspark-vllm/apply-patch-pack.py backends/dspark-vllm/packs/miaai-dsv4flash-d1b76251-defaults/patches/*.py backends/dspark-vllm/packs/miaai-ds4fv-f5665e8/patches/*.py backends/dspark-vllm/packs/miaai-ds4fv-f5665e8/patches/vision_exp/*.py backends/qwen38-vllm/apply-ple-fp8-patch.py backends/qwen38-sglang/*.py backends/glm53-exl3/*.py backends/mlx-audio/lloom_audio_server.py backends/hear/lloom_hear_server.py backends/chatterbox/lloom_chatterbox_server.py patches/apply_mtplx_longctx_fix.py backends/comfyui-media/bridge/*.py backends/comfyui-media/graphs/*.py backends/comfyui-media/build/*.py backends/comfyui-media/install.py backends/qwen-image-diffusers/*.py backends/hear/test/test_hear_security.py src/darwin-memory-usage.py",
     "check:shell": "bash -n backends/dspark-vllm/entrypoint.sh backends/dspark-vllm/runtime/build-hardened-image.sh backends/dspark-vllm/packs/miaai-dsv4flash-d1b76251-defaults/patches/*.sh backends/dspark-vllm/packs/miaai-ds4fv-f5665e8/apply-runtime.sh backends/dspark-vllm/packs/miaai-ds4fv-f5665e8/patches/*.sh backends/qwen38-vllm/entrypoint.sh backends/qwen38-sglang/entrypoint.sh backends/glm53-exl3/entrypoint.sh backends/mlx-audio/install.sh backends/hear/install.sh backends/chatterbox/install.sh clients/examples/lloom-claude clients/examples/lloom-codex clients/examples/lloom-hermes clients/examples/lloom-opencode clients/examples/lloom-zero scripts/remote-install-spark.sh scripts/remote-stage-spark-worker.sh",
     "check:spark-deploy": "node --check test/spark-deploy.test.mjs && node test/spark-deploy.test.mjs",
     "generate:clients": "node scripts/generate-client-configs.mjs",
@@ -73,7 +73,7 @@
     "community:check": "node scripts/check-community-deploy.mjs",
     "package:check": "node scripts/check-package.mjs",
     "test:adoption": "node test/resource-fit.test.mjs && node test/model-acquisition.test.mjs && node test/dashboard-status.test.mjs",
-    "test:unit": "node test/runtime-health-ownership.test.mjs && node test/long-prefill-transport.test.mjs && node test/sparkglm-managed.test.mjs && node test/protocol.test.mjs && node test/protocol-stream.test.mjs && node test/stream-metrics.test.mjs && node test/responses-codex.test.mjs && node test/security.test.mjs && node test/community-host.test.mjs && node test/ds4fv-recipe.test.mjs && node test/dspark-hotfixes.test.mjs && node test/glm53-exl3-recipe.test.mjs && node test/host-memory.test.mjs && node test/runtime-policy.test.mjs && node --test test/rate-limit.test.mjs test/runtime-queue.test.mjs test/federated-priority.test.mjs test/runtime-supervisor.test.mjs && node test/interchange.test.mjs && node test/server-resilience.test.mjs && node test/metrics-persistence.test.mjs && node test/model-failover.test.mjs && node test/qwen38-vllm-recipe.test.mjs && node test/reasoning-normalize.test.mjs && node test/reasoning-effort.test.mjs && node test/route-control.test.mjs && node test/spark-route-catalog.test.mjs && node test/chat-lane-canary.test.mjs && node test/runtime-watchdog.test.mjs && node test/tts-catalog.test.mjs && node test/voice-profiles.test.mjs && node test/chatterbox-recipe.test.mjs && node test/hear-recipe.test.mjs && npm run test:adoption && node --test test/performance-sampler.test.mjs && node --test test/audio-generations.test.mjs && node test/video-providers.test.mjs && node --test test/web-functions.test.mjs && node --test test/config-mutation.test.mjs && node --test test/model-maintenance.test.mjs && node --test test/model-maintenance-control.test.mjs && node --test test/route-maintenance-concurrency.test.mjs && node --test test/model-maintenance-integration.test.mjs && npm run test:workers && node test/comfyui-media.test.mjs && node test/qwen-image-diffusers.test.mjs && node test/runtime-residency.test.mjs",
+    "test:unit": "node test/runtime-health-ownership.test.mjs && node test/long-prefill-transport.test.mjs && node test/sparkglm-managed.test.mjs && node test/protocol.test.mjs && node test/protocol-stream.test.mjs && node test/stream-metrics.test.mjs && node test/responses-codex.test.mjs && node test/security.test.mjs && node test/community-host.test.mjs && node test/openrouter-provider.test.mjs && node test/ds4fv-recipe.test.mjs && node test/dspark-hotfixes.test.mjs && node test/glm53-exl3-recipe.test.mjs && node test/host-memory.test.mjs && node test/config-profiles.test.mjs && node test/route-control.test.mjs && node test/runtime-policy.test.mjs && node --test test/rate-limit.test.mjs test/runtime-queue.test.mjs test/federated-priority.test.mjs test/runtime-supervisor.test.mjs && node test/interchange.test.mjs && node test/server-resilience.test.mjs && node test/metrics-persistence.test.mjs && node test/model-failover.test.mjs && node test/qwen38-vllm-recipe.test.mjs && node test/reasoning-normalize.test.mjs && node test/reasoning-effort.test.mjs && node test/route-control.test.mjs && node test/spark-route-catalog.test.mjs && node test/chat-lane-canary.test.mjs && node test/runtime-watchdog.test.mjs && node test/tts-catalog.test.mjs && node test/voice-profiles.test.mjs && node test/chatterbox-recipe.test.mjs && node test/hear-recipe.test.mjs && npm run test:adoption && node --test test/performance-sampler.test.mjs && node --test test/audio-generations.test.mjs && node test/video-providers.test.mjs && node --test test/web-functions.test.mjs && node --test test/config-mutation.test.mjs && node --test test/model-maintenance.test.mjs && node --test test/model-maintenance-control.test.mjs && node --test test/route-maintenance-concurrency.test.mjs && node --test test/model-maintenance-integration.test.mjs && npm run test:workers && node test/comfyui-media.test.mjs && node test/qwen-image-diffusers.test.mjs && node test/runtime-residency.test.mjs && npm run test:ux && npm run test:memory-safety && node --test test/dashboard-memory.test.mjs test/runtime-memory-usage.test.mjs",
     "test:cluster": "node test/cluster.test.mjs",
     "test": "npm run test:cluster && npm run test:unit && npm run test:entity-stagger && npm run smoke && npm run community:check",
     "test:entity-stagger": "node test/entity-stagger-benchmark.test.mjs",
@@ -84,7 +84,9 @@
     "test:workers": "node --test test/research-workers.test.mjs test/research-worker-process.test.mjs",
     "check:workers": "node --check clients/examples/research-workers/runner.mjs && node --check clients/examples/research-workers/process.mjs && node --check clients/examples/research-workers/process-guard.mjs",
     "test:comfyui": "python3 scripts/test-comfyui-media.py",
-    "test:hear": "python3 -m unittest discover -s backends/hear/test -v"
+    "test:hear": "python3 -m unittest discover -s backends/hear/test -v",
+    "test:ux": "node --test test/first-run.test.mjs test/runtime-preferences.test.mjs test/admin-browser-security.test.mjs test/installation-jobs.test.mjs",
+    "test:memory-safety": "node --test test/runtime-memory-safety.test.mjs"
   },
   "engines": {
     "node": ">=20.0.0"
@@ -92,10 +94,10 @@
   "license": "MIT",
   "devDependencies": {
     "@eslint/js": "^10.0.1",
-    "eslint": "^10.10.0",
-    "prettier": "^3.9.6"
+    "eslint": "^10.11.0",
+    "prettier": "^3.9.8"
   },
   "dependencies": {
-    "undici": "^7.28.0"
+    "undici": "^8.10.2"
   }
 }
diff --git a/src/bootstrap.mjs b/src/bootstrap.mjs
index 05d0375..144d940 100644
--- a/src/bootstrap.mjs
+++ b/src/bootstrap.mjs
@@ -4,7 +4,7 @@ import {
   selectIntegrationArtifacts,
   writeGeneratedIntegrationArtifacts
 } from './client-integrations.mjs';
-import { applyBackend, applyRecipe, defaultInstallStatePathFor } from './installer.mjs';
+import { applyBackend, applyRecipe, pinDownloadCommands, defaultInstallStatePathFor } from './installer.mjs';
 import {
   backendIds,
   defaultBackendVariables,
@@ -175,17 +175,25 @@ export async function createBootstrapPlan(
       name: recipe.name,
       backendId: recipe.backend?.id
     },
+    // Frozen executable evidence: the exact backend/recipe step plans plus the
+    // selected recipe and modelRoot. Retries and reviewed-plan execution must
+    // replay these untouched rather than re-profiling or re-selecting.
+    reviewedRecipe: recipe,
+    reviewedBackend: backend,
+    modelRoot: selectedModelRoot,
     backend: await planBackend(backend, {
       variables: backendVariables,
       checkCommands: true
     }),
-    recipe: planRecipe(recipe, config, {
-      modelRoot: selectedModelRoot,
-      backendIds: backendIds(catalog),
-      benchmarkEvidence,
-      benchmarksRoot: selectedBenchmarksRoot,
-      benchmarkValidationErrors: benchmarkErrors
-    }),
+    recipe: await pinDownloadCommands(
+      planRecipe(recipe, config, {
+        modelRoot: selectedModelRoot,
+        backendIds: backendIds(catalog),
+        benchmarkEvidence,
+        benchmarksRoot: selectedBenchmarksRoot,
+        benchmarkValidationErrors: benchmarkErrors
+      })
+    ),
     benchmarks: {
       root: selectedBenchmarksRoot,
       validationErrors: benchmarkErrors,
@@ -212,6 +220,7 @@ export async function applyBootstrap(
     generatedRoot,
     backendVariables = defaultBackendVariables(process.env),
     _benchmarkDocuments = [],
+    reviewedPlan,
     recipesRoot,
     recipeDocuments = [],
     backendCatalogPath,
@@ -223,11 +232,27 @@ export async function applyBootstrap(
     throw new Error('Refusing to bootstrap without yes=true. Re-run with --yes after reviewing the dry-run plan.');
   }
 
-  const profile = await profileMachine();
-  const recipes = [...recipeDocuments, ...(await loadRecipes(recipesRoot))];
-  const recipe = await selectRecipe({ recipeId, recipes, profile, recipesRoot });
-  const catalog = await loadBackendCatalog(backendCatalogPath);
-  const backend = getBackend(catalog, recipe.backend?.id);
+  // A reviewed plan carries the exact backend/recipe plans that were shown to
+  // and approved by the user. When supplied, reuse the frozen evidence so a
+  // later hardware/catalog change cannot silently swap in different commands.
+  const reviewed = reviewedPlan ?? null;
+  let recipe;
+  let backend;
+  let backendPlan = null;
+  if (reviewed) {
+    if (!reviewed.reviewedRecipe || !reviewed.reviewedBackend || !reviewed.backend?.steps || !reviewed.recipe?.steps)
+      throw new Error('Reviewed bootstrap plan is incomplete. Review a fresh plan.');
+    recipe = reviewed.reviewedRecipe;
+    backend = reviewed.reviewedBackend;
+    backendPlan = reviewed.backend;
+  } else {
+    const profile = await profileMachine();
+    const recipes = [...recipeDocuments, ...(await loadRecipes(recipesRoot))];
+    recipe = await selectRecipe({ recipeId, recipes, profile, recipesRoot });
+    const catalog = await loadBackendCatalog(backendCatalogPath);
+    backend = getBackend(catalog, recipe.backend?.id);
+  }
+  if (!recipe) throw new Error('Bootstrap requires a reviewed recipe or recipeId.');
   if (!backend) throw new Error(`Recipe ${recipe.id} references unknown backend ${recipe.backend?.id}`);
 
   const registry = createRegistry(config);
@@ -237,7 +262,8 @@ export async function applyBootstrap(
     yes,
     statePath,
     variables: backendVariables,
-    env: commandEnv
+    env: commandEnv,
+    ...(backendPlan ? { reviewedPlan: backendPlan } : {})
   });
   backendResult.summary = phaseSummary('backend', backendResult);
 
@@ -249,7 +275,10 @@ export async function applyBootstrap(
         env: commandEnv,
         onProgress,
         stdio,
-        ...(modelRoot ? { modelRoot } : {})
+        ...((reviewed?.modelRoot ?? modelRoot) ? { modelRoot: reviewed?.modelRoot ?? modelRoot } : {}),
+        ...((reviewed?.recipePlan ?? (reviewed?.recipe?.steps ? reviewed.recipe : null))
+          ? { reviewedPlan: reviewed.recipePlan ?? reviewed.recipe }
+          : {})
       })
     : blockedPhase('recipe', 'backend phase failed', { dryRun });
   recipeResult.summary ??= phaseSummary('recipe', recipeResult);
diff --git a/src/browser-setup.mjs b/src/browser-setup.mjs
new file mode 100644
index 0000000..d27e787
--- /dev/null
+++ b/src/browser-setup.mjs
@@ -0,0 +1,219 @@
+import fs from 'node:fs/promises';
+import { createHash } from 'node:crypto';
+import { createFirstRunServer } from './first-run.mjs';
+import { createOnboardingPlan } from './onboarding.mjs';
+import { applySetup } from './setup.mjs';
+import { applyBootstrap } from './bootstrap.mjs';
+import { profileMachine, rankRecipes } from './machine-profile.mjs';
+import { loadRecipes, loadRecipeById } from './recipes.mjs';
+import { loadBackendCatalog } from './backend-catalog.mjs';
+import { loadConfig } from './config.mjs';
+
+const fingerprint = (value) => createHash('sha256').update(JSON.stringify(value)).digest('hex');
+export function recipeSupportsWorkload(recipe, workload) {
+  const values = new Set([
+    ...(recipe.capabilities || []),
+    ...(recipe.models || []).flatMap((model) => [model.kind, ...(model.capabilities || [])])
+  ]);
+  if (workload === 'images') return [...values].some((value) => /^image/.test(value));
+  if (workload === 'voice') return [...values].some((value) => /^audio|speech|transcription|tts|stt/.test(value));
+  const chat = ['chat', 'responses', 'anthropic-messages'].some((value) => values.has(value));
+  return workload === 'code'
+    ? chat && (values.has('tools') || (recipe.keywords || []).some((value) => /cod/.test(value)))
+    : chat;
+}
+
+export function createBrowserSetup(config, options, { gatewayStarter, serverOptions = {} } = {}) {
+  const evidence = async (recipeId) => ({
+    recipe: await loadRecipeById(recipeId, options.recipesRoot),
+    backendCatalog: await loadBackendCatalog(options.backendCatalogPath)
+  });
+  return createFirstRunServer({
+    ...serverOptions,
+    gatewayStarter,
+    async planBuilder({ workloadId, recipeId }) {
+      const [recipes, profile] = await Promise.all([loadRecipes(options.recipesRoot), profileMachine()]);
+      const ranked = await rankRecipes(
+        recipes.filter((recipe) => recipeSupportsWorkload(recipe, workloadId)),
+        profile,
+        { checkCommands: true }
+      );
+      const candidates = ranked.filter((candidate) => candidate.selectable);
+      const selected = recipeId ? candidates.find((candidate) => candidate.recipeId === recipeId) : candidates[0];
+      if (!selected)
+        throw new Error(
+          'No compatible ' +
+            workloadId +
+            ' recipe is available for this hardware. Choose another use or add a vendor recipe through the CLI.'
+        );
+      const selectedOptions = {
+        ...options,
+        recipeId: selected.recipeId,
+        offline: true,
+        start: false,
+        includeRuntimes: false
+      };
+      const plan = await createOnboardingPlan(config, selectedOptions);
+      const source = await evidence(selected.recipeId);
+      plan.browserSetup = { options: selectedOptions, fingerprint: fingerprint(source) };
+      const recipe = source.recipe;
+      const option = (candidate) => ({
+        id: candidate.recipeId,
+        name: candidate.name,
+        reason: candidate.reasons.length
+          ? candidate.reasons.join('; ')
+          : 'Compatible with ' + (profile.cpuBrand || profile.platformId) + '.',
+        memoryRequiredGb: candidate.memoryRequiredGb,
+        downloadSizeBytes: null,
+        license:
+          recipes.find((item) => item.id === candidate.recipeId)?.license?.name ||
+          recipes.find((item) => item.id === candidate.recipeId)?.license?.id ||
+          null,
+        credentials: null
+      });
+      return {
+        plan,
+        view: {
+          workloadId,
+          selected: option(selected),
+          alternatives: candidates
+            .filter((c) => c.recipeId !== selected.recipeId)
+            .slice(0, 8)
+            .map(option),
+          machine: {
+            name: profile.cpuBrand || profile.platformId,
+            platformId: profile.platformId,
+            totalMemoryGb: profile.totalMemoryGb,
+            accelerators: profile.accelerators || []
+          },
+          paths: { configPath: plan.configPath, modelRoot: plan.modelRoot },
+          ports: plan.ports,
+          review: {
+            stages: plan.stages,
+            doctorCommands: plan.next?.doctor,
+            packages: (plan.setup?.phases?.bootstrap?.backend?.steps || []).map((step) => step.label || step.id),
+            models: (recipe.models || []).map((model) => model.gatewayModel || model.model)
+          }
+        }
+      };
+    },
+    applyRunner: createReviewedSetupRunner(config, { evidence }),
+    async gatewayProbe(report, plan, gateway) {
+      const installed = await loadConfig(report.configPath);
+      const url = gateway?.url || report.dashboardUrl;
+      const key = installed.security?.apiKeys?.[0] || installed.security?.adminApiKeys?.[0];
+      const headers = { 'content-type': 'application/json', ...(key ? { authorization: 'Bearer ' + key } : {}) };
+      const healthReply = await fetch(new URL('/health', url), { signal: AbortSignal.timeout(15000) });
+      const healthData = healthReply.ok ? await healthReply.json() : null;
+      const health = healthData?.ok === true && (!gateway?.pid || healthData.pid === gateway.pid);
+      if (!health)
+        return {
+          healthy: false,
+          inferenceVerified: false,
+          endpoint: url,
+          detail: 'Installed, but the gateway did not pass its health check.'
+        };
+      const model = installed.models.find(
+        (item) =>
+          (item.kind || 'chat') === 'chat' &&
+          (plan.setup?.phases?.init?.config?.models || []).some((candidate) => candidate.id === item.id)
+      );
+      if (!model)
+        return {
+          healthy: true,
+          inferenceVerified: false,
+          endpoint: url,
+          detail: 'Gateway is running. Try the installed media model to verify its output.'
+        };
+      try {
+        const response = await fetch(new URL('/v1/chat/completions', url), {
+          method: 'POST',
+          headers,
+          body: JSON.stringify({
+            model: model.id,
+            messages: [{ role: 'user', content: 'Reply with Ready.' }],
+            max_tokens: 16,
+            stream: false
+          }),
+          signal: AbortSignal.timeout(180000)
+        });
+        const body = await response.json();
+        const verified =
+          response.ok &&
+          Array.isArray(body.choices) &&
+          body.choices.some((choice) => typeof choice.message?.content === 'string' && choice.message.content.trim());
+        return {
+          healthy: true,
+          inferenceVerified: verified,
+          endpoint: url,
+          detail: verified
+            ? 'A model answered through your gateway.'
+            : body.error?.message || 'Gateway is running, but the model did not return a text response.'
+        };
+      } catch (error) {
+        return {
+          healthy: true,
+          inferenceVerified: false,
+          endpoint: url,
+          detail: 'Gateway is running. Model verification needs attention: ' + error.message
+        };
+      }
+    }
+  });
+}
+
+export function createReviewedSetupRunner(config, { evidence, apply = applySetup, bootstrap = applyBootstrap }) {
+  let ownedConfig = null;
+  return async (plan, { onProgress }) => {
+    const details = plan.browserSetup;
+    if (!details || fingerprint(await evidence(plan.selectedRecipe.id)) !== details.fingerprint)
+      throw new Error('The recipe or backend catalog changed. Review a fresh plan.');
+    // A first-run wizard must never silently replace a newly created or
+    // concurrently installed configuration.
+    let existing;
+    try {
+      existing = JSON.parse(await fs.readFile(plan.configPath, 'utf8'));
+    } catch (error) {
+      if (error.code !== 'ENOENT') throw error;
+    }
+    if (existing) {
+      if (
+        !ownedConfig ||
+        ownedConfig.path !== plan.configPath ||
+        ownedConfig.recipeId !== plan.selectedRecipe.id ||
+        fingerprint(existing) !== ownedConfig.fingerprint
+      )
+        throw new Error('A configuration now exists or changed. Open LLooM to manage it; setup will not overwrite it.');
+      if (fingerprint(plan.setup.phases.init.config) !== ownedConfig.fingerprint)
+        throw new Error(
+          'The installation plan changed after configuration was created. Continue with lloom bootstrap --apply --yes to resume the installed recipe.'
+        );
+      // Retry only the bootstrap for the exact configuration this session
+      // successfully created. Never replace it after a partial download.
+      const installed = await loadConfig(plan.configPath);
+      const result = await bootstrap(installed, {
+        ...details.options,
+        modelRoot: plan.modelRoot,
+        // Retry replays the exact reviewed bootstrap evidence; never re-plan.
+        reviewedPlan: plan.setup?.phases?.bootstrap,
+        dryRun: false,
+        yes: true,
+        onProgress
+      });
+      return { ...result, configPath: plan.configPath, dashboardUrl: plan.dashboardUrl };
+    }
+    const result = await apply(config, {
+      ...details.options,
+      reviewedPlan: plan.setup,
+      dryRun: false,
+      yes: true,
+      start: false,
+      onProgress,
+      exclusiveConfig: true,
+      onConfigWritten(file, value) {
+        ownedConfig = { path: file, recipeId: plan.selectedRecipe.id, fingerprint: fingerprint(value) };
+      }
+    });
+    return { ...result, dashboardUrl: plan.dashboardUrl };
+  };
+}
diff --git a/src/cluster.mjs b/src/cluster.mjs
index 726f2ff..c18598a 100644
--- a/src/cluster.mjs
+++ b/src/cluster.mjs
@@ -1,7 +1,9 @@
 import os from 'node:os';
 import fs from 'node:fs/promises';
 import path from 'node:path';
-import { Agent as UndiciAgent } from 'undici';
+// Dispatcher and fetch must come from the same undici copy: the package's
+// v8 Agent speaks a dispatcher contract Node 22's internal undici v6 rejects.
+import { Agent as UndiciAgent, fetch as undiciFetch } from 'undici';
 import { runCommand } from './process-control.mjs';
 
 // Node's built-in fetch aborts a request that has not produced response headers
@@ -226,20 +228,36 @@ export function parseNvidiaSyncSshConfig(source = '') {
   let entry = null;
   for (const rawLine of String(source).split(/\r?\n/)) {
     const line = rawLine.trim();
-    const host = line.match(/^Host\s+([^*?!\s]+)$/i);
-    if (host) {
-      if (entry?.createdBySync && entry.hostname) entries.push(entry);
-      entry = { alias: host[1], createdBySync: false };
+    // A Host/Match keyword always closes the current block. Wildcard, negated,
+    // and multi-host patterns are not usable aliases, and Match introduces a
+    // conditional scope whose body is not part of the previous Host block, so
+    // the current entry must be finalized and cleared rather than merged forward.
+    const blockStart = line.match(/^(Host|Match)\s+(.*)$/i);
+    if (blockStart) {
+      if (entry) entries.push(entry);
+      entry = null;
+      const alias = blockStart[1].toLowerCase() === 'host' ? blockStart[2].trim() : '';
+      if (alias && /^[^\s*?!]+$/.test(alias)) entry = { alias, createdBySync: false };
       continue;
     }
     if (!entry) continue;
-    if (/^###\s*CreatedBy:\s*NVIDIA Sync$/i.test(line)) entry.createdBySync = true;
+    // Comments carry NVIDIA Sync provenance. The marker is checked before the
+    // generic comment guard, and every other comment line is ignored.
+    const marker = line.match(/^#+\s*NVSyncClusterAlias\s*:\s*(\S+)\s*$/i);
+    if (marker) {
+      entry.clusterAlias = marker[1];
+      continue;
+    }
+    if (/^#/.test(line)) {
+      if (/^#+\s*CreatedBy:\s*NVIDIA Sync\s*$/i.test(line)) entry.createdBySync = true;
+      continue;
+    }
     const hostname = line.match(/^Hostname\s+(\S+)$/i);
     if (hostname) entry.hostname = hostname[1];
     const user = line.match(/^User\s+(\S+)$/i);
     if (user) entry.user = user[1];
   }
-  if (entry?.createdBySync && entry.hostname) entries.push(entry);
+  if (entry) entries.push(entry);
   return entries;
 }
 
@@ -276,52 +294,348 @@ function friendlyNodeId(alias) {
   return String(alias).replace(/-lan$/i, '');
 }
 
+function compareIpv4(left, right) {
+  const leftParts = String(left).split('.').map(Number);
+  const rightParts = String(right).split('.').map(Number);
+  const valid = (parts) =>
+    parts.length === 4 && parts.every((value) => Number.isInteger(value) && value >= 0 && value <= 255);
+  const leftValid = valid(leftParts);
+  const rightValid = valid(rightParts);
+  if (leftValid && rightValid) {
+    for (let index = 0; index < 4; index += 1) {
+      if (leftParts[index] !== rightParts[index]) return leftParts[index] - rightParts[index];
+    }
+    return 0;
+  }
+  if (leftValid !== rightValid) return leftValid ? -1 : 1;
+  return String(left).localeCompare(String(right));
+}
+
+function isAddressAlias(value) {
+  return /^\d{1,3}(\.\d{1,3}){3}$/.test(String(value ?? ''));
+}
+
+function syncPeerMatches(peers, localAddresses, localAddressSet = new Set()) {
+  const matches = [];
+  const ambiguous = [];
+  for (const original of peers) {
+    if (!original?.createdBySync || !original.hostname) continue;
+    const clusterAlias = original.clusterAlias ?? (!isAddressAlias(original.alias) ? original.alias : null);
+    if (!clusterAlias || !/^[a-z0-9][a-z0-9._-]*$/i.test(clusterAlias)) continue;
+    const peer = { ...original, clusterAlias };
+    const privateAddress = /^10\.100\./.test(peer.hostname) ? peer.hostname : null;
+    if (!privateAddress) continue;
+    if (localAddressSet.has(privateAddress)) {
+      ambiguous.push({
+        peerAddress: privateAddress,
+        sshAlias: peer.alias ?? null,
+        clusterAlias: peer.clusterAlias,
+        reason: 'local-address',
+        message: `ignored Sync alias ${peer.clusterAlias}: ${privateAddress} is one of this node's own addresses`
+      });
+      continue;
+    }
+    const candidates = [
+      ...new Map(
+        localAddresses
+          .filter(
+            (address) =>
+              address.address?.startsWith('10.100.') && sameSubnet(address.address, privateAddress, address.prefixlen)
+          )
+          .map((address) => [address.interface ?? address.address, address])
+      ).values()
+    ];
+    const candidateAddresses = candidates.map((address) => address.address);
+    if (candidates.length > 1) {
+      ambiguous.push({
+        peerAddress: privateAddress,
+        sshAlias: peer.alias ?? null,
+        clusterAlias: peer.clusterAlias,
+        localAddresses: candidateAddresses.sort(compareIpv4),
+        reason: 'ambiguous-local-interface',
+        message:
+          `ignored Sync alias ${peer.clusterAlias}: peer address ${privateAddress} matches multiple local ` +
+          `interfaces (${candidates
+            .map((address) => address.interface ?? address.address)
+            .sort()
+            .join(', ')}); ` +
+          `refusing to guess a rail`
+      });
+      continue;
+    }
+    const local = candidates[0];
+    if (!local) continue;
+    matches.push({ peer, local, privateAddress });
+  }
+  return { matches, ambiguous };
+}
+
+function nodeRailsForPeers(matches) {
+  const nodes = new Map();
+  for (const match of matches) {
+    const id = friendlyNodeId(match.peer.clusterAlias);
+    let existing = nodes.get(id);
+    if (!existing) {
+      existing = { id, alias: match.peer.alias, sshAlias: null, sshUser: null, rails: [] };
+      nodes.set(id, existing);
+    }
+    if (!existing.rails.some((candidate) => candidate.peerAddress === match.privateAddress)) {
+      existing.rails.push({
+        interface: match.local.interface,
+        localInterface: match.local.interface,
+        localAddress: match.local.address,
+        peerAddress: match.privateAddress,
+        prefixlen: match.local.prefixlen,
+        sshAlias: match.peer.clusterAlias,
+        sshUser: match.peer.user ?? null
+      });
+    }
+    if (!isAddressAlias(match.peer.alias)) {
+      existing.sshAlias ??= match.peer.alias;
+      existing.sshUser ??= match.peer.user ?? null;
+    } else {
+      existing.sshUser ??= match.peer.user ?? null;
+    }
+  }
+  return nodes;
+}
+
 export function buildNvidiaSyncDiscovery({ peers = [], localAddresses = [], localNodeId, hostname } = {}) {
-  const matches = peers
-    .map((peer) => {
-      const local = localAddresses.find(
-        (address) =>
-          address.address?.startsWith('10.100.') && sameSubnet(address.address, peer.hostname, address.prefixlen)
-      );
-      return local ? { peer, local } : null;
-    })
-    .filter(Boolean);
+  const localAddressSet = new Set(localAddresses.map((address) => address.address).filter(Boolean));
+  const { matches, ambiguous } = syncPeerMatches(peers, localAddresses, localAddressSet);
   if (!matches.length) return null;
-  const preferred = [...matches].sort(
-    (left, right) => Number(right.local.address.split('.')[2]) - Number(left.local.address.split('.')[2])
-  )[0];
   const id = localNodeId || hostname || os.hostname();
-  const peerId = friendlyNodeId(preferred.peer.alias);
-  const nodeIds = [id, peerId].sort();
+  if (matches.some((match) => friendlyNodeId(match.peer.clusterAlias) === id)) {
+    throw new Error(`Sync peer alias conflicts with local node id ${id}`);
+  }
+  const discovered = [...nodeRailsForPeers(matches).values()].sort(
+    (left, right) =>
+      compareIpv4(left.rails[0].peerAddress, right.rails[0].peerAddress) || left.id.localeCompare(right.id)
+  );
+
+  const localPrimary = [...matches].sort((left, right) => compareIpv4(left.local.address, right.local.address))[0]
+    .local;
+  const links = [
+    ...new Map(
+      matches.map((match) => [
+        `${friendlyNodeId(match.peer.clusterAlias)}|${match.privateAddress}`,
+        {
+          nodeId: id,
+          peerNodeId: friendlyNodeId(match.peer.clusterAlias),
+          localInterface: match.local.interface,
+          localAddress: match.local.address,
+          peerAddress: match.privateAddress,
+          prefixlen: match.local.prefixlen
+        }
+      ])
+    ).values()
+  ].sort(
+    (left, right) =>
+      compareIpv4(left.peerAddress, right.peerAddress) ||
+      compareIpv4(left.localAddress, right.localAddress) ||
+      left.peerNodeId.localeCompare(right.peerNodeId)
+  );
+
+  const nodes = {
+    [id]: {
+      id,
+      local: true,
+      backendHost: localPrimary.address,
+      fabricInterface: localPrimary.interface
+    }
+  };
+  for (const node of discovered) {
+    const rails = [...node.rails].sort((left, right) => compareIpv4(left.peerAddress, right.peerAddress));
+    const primary = rails[0];
+    nodes[node.id] = {
+      id: node.id,
+      local: false,
+      backendHost: primary.peerAddress,
+      ...(node.sshAlias ? { sshAlias: node.sshAlias } : {}),
+      sshUser: node.sshUser,
+      rails
+    };
+  }
+
+  const familyIds = Object.keys(nodes).sort();
+  const topology = Object.keys(nodes).length === 2 ? 'direct' : 'local-adjacency';
   return {
     detected: true,
     provider: 'nvidia-sync',
-    topology: 'direct',
-    nodeCount: nodeIds.length,
+    topology,
+    nodeCount: familyIds.length,
     nodeId: id,
-    leaderNode: nodeIds[0],
+    localNode: id,
     fabric: {
-      interface: preferred.local.interface,
-      localAddress: preferred.local.address,
-      peerAddress: preferred.peer.hostname,
-      prefixlen: preferred.local.prefixlen
+      interface: localPrimary.interface,
+      localAddress: localPrimary.address,
+      peerAddress: matches.find((match) => match.local.address === localPrimary.address).privateAddress,
+      prefixlen: localPrimary.prefixlen
     },
-    nodes: {
-      [id]: {
-        id,
-        local: true,
-        backendHost: preferred.local.address,
-        fabricInterface: preferred.local.interface
+    verified: false,
+    evidence: 'local-interface-and-ssh-config',
+    links,
+    nodes,
+    ...(ambiguous.length ? { ambiguousPeers: ambiguous } : {})
+  };
+}
+
+export function nvidiaSyncDiscoverySummary(discovery, { currentConfig = null } = {}) {
+  if (!discovery?.detected) return null;
+  const localId = discovery.nodeId;
+  const configuredNodes = asObject(currentConfig?.cluster?.nodes);
+  const configuredNodeIds = Object.keys(configuredNodes);
+  const discoveredIds = Object.keys(discovery.nodes).sort();
+  const unresolvedPeers = Object.values(discovery.nodes)
+    .filter((node) => !node.local && !configuredNodes[node.id])
+    .map((node) => node.id);
+  const customEndpoints = Object.entries(configuredNodes)
+    .filter(([nodeId, node]) => nodeId !== localId && node?.endpoint)
+    .filter(([nodeId, node]) => {
+      const discovered = discovery.nodes[nodeId];
+      if (!discovered?.backendHost) return false;
+      return !endpointHostMatches(node.endpoint, discovered.backendHost);
+    })
+    .map(([nodeId, node]) => ({ nodeId, endpoint: node.endpoint, backendHost: node.backendHost ?? null }));
+  const diagnostics = [];
+  if (configuredNodeIds.length && !configuredNodes[localId]) {
+    diagnostics.push(
+      `discovered local node id ${localId} is not the configured cluster.nodeId (${configuredNodeIds.join(', ')}); ` +
+        `confirm cluster.nodeId before applying so the local node is not duplicated`
+    );
+  }
+  for (const nodeId of unresolvedPeers) {
+    diagnostics.push(`discovered peer ${nodeId} is not configured yet`);
+  }
+  for (const entry of customEndpoints) {
+    diagnostics.push(
+      `Verify the listener for ${entry.nodeId} before moving its configured endpoint onto the observed fabric`
+    );
+  }
+  for (const ambiguous of discovery.ambiguousPeers ?? []) {
+    diagnostics.push(ambiguous.message);
+  }
+  return {
+    provider: discovery.provider,
+    topology: discovery.topology,
+    nodeId: discovery.nodeId,
+    localNode: discovery.localNode ?? discovery.nodeId,
+    leaderNode: currentConfig?.cluster?.leaderNode ?? null,
+    nodeCount: discovery.nodeCount,
+    discoveredNodes: discoveredIds,
+    discoveredPeers: discoveredIds.filter((nodeId) => nodeId !== localId),
+    observedLinks: discovery.links.length,
+    configuredNodes: configuredNodeIds.sort(),
+    newNodes: discoveredIds.filter((nodeId) => !configuredNodes[nodeId] && nodeId !== localId),
+    unresolvedPeers,
+    customEndpoints,
+    mergeRequired: configuredNodeIds.length > 0,
+    diagnostics,
+    ...(diagnostics.length ? { diagnostic: diagnostics[0] } : {})
+  };
+}
+
+function endpointHostname(endpoint) {
+  const value = String(endpoint ?? '').trim();
+  if (!value) return null;
+  try {
+    return new URL(value).hostname || null;
+  } catch {
+    return null;
+  }
+}
+
+function endpointHostMatches(endpoint, host) {
+  if (!host) return false;
+  const hostname = endpointHostname(endpoint);
+  return Boolean(hostname) && hostname === String(host);
+}
+
+export function mergeNvidiaSyncClusterDiscovery({ discovery, cluster = {}, id, apiKeyEnv, localNodeId } = {}) {
+  if (!discovery?.detected || discovery.provider !== 'nvidia-sync') {
+    throw new Error('NVIDIA Sync cluster was not detected');
+  }
+  const current = asObject(cluster);
+  const nodeId = localNodeId ?? current.nodeId ?? discovery.nodeId;
+  if (current.nodeId && !String(current.nodeId).includes('${') && current.nodeId !== nodeId) {
+    throw new Error('Configured local node identity conflicts with discovery');
+  }
+  if (Object.keys(asObject(current.nodes)).length && !current.nodes[nodeId]) {
+    throw new Error(
+      `Configured node map does not contain local node ${nodeId}; set cluster.nodeId before applying discovery`
+    );
+  }
+  const leaderNode = current.leaderNode;
+  const nodes = { ...asObject(current.nodes) };
+  const diagnostics = [];
+  for (const [observedId, observed] of Object.entries(discovery.nodes)) {
+    const targetId = observedId === discovery.nodeId ? nodeId : observedId;
+    if (targetId !== nodeId && !Object.hasOwn(nodes, targetId)) {
+      diagnostics.push({
+        level: 'unconfigured-peer',
+        nodeId: targetId,
+        message: `Observed ${targetId}; use cluster add-node with its authenticated gateway URL to join it.`
+      });
+      continue;
+    }
+    const existing = asObject(nodes[targetId]);
+    let endpoint = existing.endpoint;
+    let backendHost = existing.backendHost ?? observed.backendHost;
+    if (
+      existing.endpoint &&
+      /^10\.100\./.test(existing.backendHost ?? '') &&
+      endpointHostMatches(existing.endpoint, existing.backendHost)
+    ) {
+      backendHost = observed.backendHost;
+      const url = new URL(existing.endpoint);
+      url.hostname = observed.backendHost;
+      endpoint = url.toString().replace(/\/$/, existing.endpoint.endsWith('/') ? '/' : '');
+    } else if (existing.endpoint && !endpointHostMatches(existing.endpoint, observed.backendHost)) {
+      diagnostics.push({
+        level: 'migration-required',
+        nodeId: targetId,
+        message: `Preserved endpoint for ${targetId}; observed fabric address ${observed.backendHost}. Verify its gateway listener before changing the endpoint.`
+      });
+    }
+    nodes[targetId] = {
+      ...existing,
+      name: existing.name ?? targetId,
+      ...(endpoint ? { endpoint } : {}),
+      backendHost,
+      fabricAddress: observed.backendHost,
+      ...(observed.fabricInterface ? { fabricInterface: observed.fabricInterface } : {}),
+      ...(observed.sshAlias ? { sshAlias: observed.sshAlias } : {}),
+      ...(observed.rails ? { rails: observed.rails } : {}),
+      labels: {
+        provider: 'nvidia-sync',
+        role: leaderNode ? (targetId === leaderNode ? 'leader' : 'worker') : 'node',
+        hardware: 'dgx-spark',
+        ...asObject(existing.labels)
       },
-      [peerId]: {
-        id: peerId,
-        local: false,
-        backendHost: preferred.peer.hostname,
-        fabricInterface: preferred.local.interface,
-        sshAlias: preferred.peer.alias,
-        sshUser: preferred.peer.user ?? null
+      resources: {
+        memoryGb: 128,
+        accelerators: ['cuda', 'nvidia-gpu', 'blackwell', 'gb10'],
+        ...asObject(existing.resources)
       }
-    }
+    };
+  }
+  return {
+    ...current,
+    id: id ?? current.id ?? `${nodeId}-cluster`,
+    provider: discovery.provider,
+    topology: current.topology ?? discovery.topology,
+    nodeId: current.nodeId ?? nodeId,
+    ...(leaderNode ? { leaderNode } : {}),
+    ...(apiKeyEnv ? { apiKeyEnv } : Object.keys(current).length ? {} : { apiKeyEnv: 'LLOOM_CLUSTER_KEY' }),
+    discovery: {
+      provider: discovery.provider,
+      evidence: discovery.evidence,
+      verified: false,
+      nodes: discovery.nodes,
+      links: (discovery.links ?? []).map((link) => ({ ...link, nodeId }))
+    },
+    nodes,
+    diagnostics
   };
 }
 
@@ -336,7 +650,7 @@ async function tailscaleSelfName() {
   }
 }
 
-export async function detectNvidiaSyncCluster({ home = process.env.HOME, hostname = os.hostname() } = {}) {
+export async function detectNvidiaSyncCluster({ home = process.env.HOME, hostname = os.hostname(), localNodeId } = {}) {
   if (process.platform !== 'linux' || !home) return null;
   let sshConfig;
   try {
@@ -357,42 +671,14 @@ export async function detectNvidiaSyncCluster({ home = process.env.HOME, hostnam
   return buildNvidiaSyncDiscovery({
     peers,
     localAddresses,
-    localNodeId: (await tailscaleSelfName()) ?? hostname,
+    localNodeId: localNodeId ?? (await tailscaleSelfName()) ?? hostname,
     hostname
   });
 }
 
-export function nvidiaSyncClusterConfig(discovery, { id, port = 8100, apiKeyEnv = 'LLOOM_CLUSTER_KEY' } = {}) {
-  if (!discovery?.detected || discovery.provider !== 'nvidia-sync') {
-    throw new Error('NVIDIA Sync cluster was not detected');
-  }
-  const clusterId = id || `${discovery.leaderNode}-cluster`;
-  return {
-    id: clusterId,
-    provider: discovery.provider,
-    topology: discovery.topology,
-    nodeId: discovery.nodeId,
-    leaderNode: discovery.leaderNode,
-    apiKeyEnv,
-    nodes: Object.fromEntries(
-      Object.entries(discovery.nodes).map(([nodeId, node]) => [
-        nodeId,
-        {
-          name: nodeId,
-          endpoint: `http://${node.backendHost}:${port}`,
-          backendHost: node.backendHost,
-          fabricInterface: node.fabricInterface,
-          ...(node.sshAlias ? { sshAlias: node.sshAlias } : {}),
-          labels: {
-            provider: 'nvidia-sync',
-            role: nodeId === discovery.leaderNode ? 'leader' : 'worker',
-            hardware: 'dgx-spark'
-          },
-          resources: { memoryGb: 128, accelerators: ['cuda', 'nvidia-gpu', 'blackwell', 'gb10'] }
-        }
-      ])
-    )
-  };
+// Pure discovery planning; remote gateways join through authenticated add-node.
+export function nvidiaSyncClusterConfig(discovery, options = {}) {
+  return mergeNvidiaSyncClusterDiscovery({ discovery, ...options });
 }
 
 export function runtimePlacement(runtime, config, env = process.env) {
@@ -602,7 +888,7 @@ export function validateClusterConfig(config, env = process.env) {
 export class ClusterCoordinator {
   constructor(
     config,
-    { env = process.env, fetchImpl = fetch, logger = console, telemetry = null, profile = null, models = null } = {}
+    { env = process.env, fetchImpl = undiciFetch, logger = console, telemetry = null, profile = null, models = null } = {}
   ) {
     this.config = config;
     this.env = env;
@@ -682,7 +968,7 @@ export class ClusterCoordinator {
     }
   }
 
-  async localNodeStatus({ runtimeStatus = null } = {}) {
+  async localNodeStatus({ runtimeStatus = null, includeMemoryUsage = false } = {}) {
     const profile = typeof this.profile === 'function' ? await this.profile() : this.profile;
     const models = typeof this.models === 'function' ? await this.models() : this.models;
     return {
@@ -704,7 +990,9 @@ export class ClusterCoordinator {
       telemetry: this.telemetry ? await this.telemetry.snapshot() : null,
       runtimeManager:
         runtimeStatus ??
-        (this.runtimeManager ? await this.runtimeManager.status({ localOnly: true }) : { runtimes: {} })
+        (this.runtimeManager
+          ? await this.runtimeManager.status({ localOnly: true, includeMemoryUsage })
+          : { runtimes: {} })
     };
   }
 
@@ -717,7 +1005,7 @@ export class ClusterCoordinator {
     if (!refresh && cached?.pending) return cached.pending;
     const pending = (async () => {
       try {
-        const result = await this.requestNode(nodeId, '/gateway/node');
+        const result = await this.requestNode(nodeId, '/gateway/node?memoryUsage=1');
         return {
           ...result.node,
           id: nodeId,
diff --git a/src/config-profiles.mjs b/src/config-profiles.mjs
new file mode 100644
index 0000000..16b052c
--- /dev/null
+++ b/src/config-profiles.mjs
@@ -0,0 +1,280 @@
+// Named fleet profiles: one file describes routing + residency for the whole
+// deployment; applying it is a single atomic config swap.
+//
+// A profile lives in <configDir>/profiles/<name>.json and may set:
+//   routes.<aliasId>    route profile name (alias.routeProfiles key) or a
+//                       plain model/alias id to pin as the sole member
+//   residency.<runtime> 'always' | 'preferred' | 'auto' (keep-warm roster)
+//   defaults            optional top-level defaults override (chatModel, ...)
+// Applying validates the composed config completely before the atomic write
+// (mutateConfigSource validates the staged file), so a swap is all-or-nothing.
+import { promises as fs } from 'node:fs';
+import path from 'node:path';
+import { mutateConfigSource } from './config-mutation.mjs';
+import { loadConfig } from './config.mjs';
+
+const RESIDENCY = new Set(['always', 'preferred', 'auto']);
+const MAX_PROFILE_BYTES = 256 * 1024;
+
+function fail(message, statusCode = 400) {
+  return Object.assign(new Error(message), { statusCode });
+}
+
+function object(value) {
+  return value && typeof value === 'object' && !Array.isArray(value) ? value : null;
+}
+
+function safeName(name) {
+  return typeof name === 'string' && /^[a-z0-9][a-z0-9._-]{0,63}$/i.test(name) ? name : null;
+}
+
+function profilesDir(config) {
+  if (!config.sourcePath) throw fail('Fleet profiles need a file-backed LLooM config.', 409);
+  return path.join(path.dirname(path.resolve(config.sourcePath)), 'profiles');
+}
+
+// Route profile semantics copied from route-control: a profile's `members`
+// (legacy `target`/`fallbacks`) becomes the alias's complete member list.
+export function resolveRouteTarget(alias, target) {
+  if (!object(alias)) throw fail(`Unknown route alias: ${target && object(target) ? '' : target}`);
+  const profiles = object(alias.routeProfiles) ?? {};
+  if (typeof target === 'string' && profiles[target]) {
+    const profile = profiles[target];
+    const members = Array.isArray(profile.members) ? profile.members : null;
+    if (!members?.length) throw fail(`Route profile ${target} has no members.`);
+    const optionalMembers = Array.isArray(profile.optionalMembers) ? profile.optionalMembers : [];
+    return { activeRoute: target, members, optionalMembers };
+  }
+  if (typeof target !== 'string' || !target.trim()) throw fail('Route target must be an id or profile name.');
+  const id = target.trim();
+  const known = (candidate) =>
+    candidate === id || (object(alias.members)?.includes ?? (() => false)).call(alias.members, id);
+  if (!known(id) && !Array.isArray(alias.members))
+    throw fail(`Alias has no route profile or member named ${id}.`);
+  return { activeRoute: null, members: [id], optionalMembers: [] };
+}
+
+export function normalizeProfileDocument(raw, name) {
+  const doc = object(raw);
+  if (!doc) throw fail(`Profile ${name} must be a JSON object.`);
+  const allowed = new Set(['name', 'description', 'routes', 'residency', 'defaults']);
+  const unknown = Object.keys(doc).filter((key) => !allowed.has(key));
+  if (unknown.length) throw fail(`Profile ${name} has unsupported sections: ${unknown.join(', ')}.`);
+  const routes = {};
+  for (const [aliasId, target] of Object.entries(object(doc.routes) ?? {})) {
+    if (typeof aliasId !== 'string' || !aliasId.trim() || aliasId.length > 200)
+      throw fail(`Profile ${name} has an invalid alias id.`);
+    if (typeof target !== 'string' || !target.trim() || target.length > 500)
+      throw fail(`Profile ${name}: route for ${aliasId} must be an id or profile name.`);
+    routes[aliasId] = target.trim();
+  }
+  const residency = {};
+  for (const [runtimeId, policy] of Object.entries(object(doc.residency) ?? {})) {
+    if (typeof runtimeId !== 'string' || !runtimeId.trim() || runtimeId.length > 200)
+      throw fail(`Profile ${name} has an invalid runtime id.`);
+    if (!RESIDENCY.has(policy)) throw fail(`Profile ${name}: residency for ${runtimeId} must be always, preferred, or auto.`);
+    residency[runtimeId.trim()] = policy;
+  }
+  const defaults = object(doc.defaults);
+  if (defaults) {
+    for (const [key, value] of Object.entries(defaults)) {
+      if (typeof value !== 'string' || value.length > 500) throw fail(`Profile ${name}: defaults.${key} must be a short string.`);
+    }
+  }
+  return {
+    name: typeof doc.name === 'string' ? doc.name : name,
+    description: typeof doc.description === 'string' ? doc.description.slice(0, 500) : '',
+    routes,
+    residency,
+    defaults: defaults ? { ...defaults } : null
+  };
+}
+
+export function profilePaths(config) {
+  return { dir: profilesDir(config), fileFor: (name) => path.join(profilesDir(config), name + '.json') };
+}
+
+export async function listProfiles(config) {
+  const dir = profilesDir(config);
+  let names = [];
+  try {
+    names = (await fs.readdir(dir)).filter((name) => name.endsWith('.json')).map((name) => name.slice(0, -5));
+  } catch (error) {
+    if (error.code !== 'ENOENT') throw error;
+  }
+  names.sort((left, right) => left.localeCompare(right));
+  const active = await activeProfileName(config);
+  const profiles = [];
+  for (const name of names) {
+    if (!safeName(name)) continue;
+    try {
+      profiles.push(await readProfile(config, name, active));
+    } catch {
+      profiles.push({ name, error: 'unreadable profile' });
+    }
+  }
+  return { dir, active, profiles };
+}
+
+export async function readProfile(config, name, activeName = null) {
+  const safe = safeName(name);
+  if (!safe) throw fail('Invalid profile name.');
+  const file = path.join(profilesDir(config), safe + '.json');
+  const raw = await fs.readFile(file, 'utf8');
+  if (raw.length > MAX_PROFILE_BYTES) throw fail('Profile file is too large.', 413);
+  const doc = normalizeProfileDocument(JSON.parse(raw), safe);
+  return { ...doc, file, active: activeName != null ? safe === activeName : undefined };
+}
+
+async function readRawProfile(config, name) {
+  const safe = safeName(name);
+  if (!safe) throw fail('Invalid profile name.');
+  const file = path.join(profilesDir(config), safe + '.json');
+  const raw = await fs.readFile(file, 'utf8');
+  if (raw.length > MAX_PROFILE_BYTES) throw fail('Profile file is too large.', 413);
+  return { safe, doc: JSON.parse(raw), file };
+}
+
+// The active profile name is recorded in the config under fleet.activeProfile
+// when applied; hand-written configs simply have no marker.
+async function activeProfileName(config) {
+  const raw = JSON.parse(await fs.readFile(path.resolve(config.sourcePath), 'utf8'));
+  const marker = object(raw.fleet)?.activeProfile;
+  return safeName(marker) ? marker : null;
+}
+
+export function planProfileChanges(config, doc) {
+  const changes = { routes: [], residency: [], defaults: [], unchanged: [] };
+  const aliases = object(config.aliases) ?? {};
+  for (const [aliasId, target] of Object.entries(doc.routes)) {
+    const alias = aliases[aliasId];
+    if (!object(alias)) throw fail(`Config has no alias ${aliasId}.`, 409);
+    const resolved = resolveRouteTarget(alias, target);
+    const same =
+      JSON.stringify(alias.members ?? []) === JSON.stringify(resolved.members) &&
+      JSON.stringify(alias.optionalMembers ?? []) === JSON.stringify(resolved.optionalMembers) &&
+      (alias.activeRoute ?? null) === resolved.activeRoute;
+    if (same) changes.unchanged.push({ kind: 'route', id: aliasId, value: target });
+    else changes.routes.push({ id: aliasId, from: alias.activeRoute ?? alias.members, to: target, ...resolved });
+  }
+  const runtimes = object(config.runtimes) ?? {};
+  for (const [runtimeId, policy] of Object.entries(doc.residency)) {
+    const runtime = runtimes[runtimeId];
+    if (!object(runtime)) throw fail(`Config has no runtime ${runtimeId}.`, 409);
+    const current = runtime.keepWarm === true ? 'always' : runtime.preferredWarm === true ? 'preferred' : 'auto';
+    if (current === policy) changes.unchanged.push({ kind: 'residency', id: runtimeId, value: policy });
+    else changes.residency.push({ id: runtimeId, from: current, to: policy });
+  }
+  for (const [key, value] of Object.entries(doc.defaults ?? {})) {
+    const current = object(config.defaults)?.[key];
+    if (current === value) changes.unchanged.push({ kind: 'default', id: key, value });
+    else changes.defaults.push({ id: key, from: current ?? null, to: value });
+  }
+  return changes;
+}
+
+// Compose the profile onto the raw source. Pure and synchronous so
+// mutateConfigSource can stage + validate the exact candidate.
+export function composeProfile(raw, doc, name) {
+  const fleet = object(raw.fleet) ?? {};
+  raw.fleet = { ...fleet, activeProfile: name };
+  const aliases = object(raw.aliases) ?? {};
+  for (const [aliasId, target] of Object.entries(doc.routes)) {
+    const alias = object(aliases[aliasId]);
+    if (!alias) throw fail(`Config has no alias ${aliasId}.`, 409);
+    const resolved = resolveRouteTarget({ ...alias }, target);
+    alias.members = resolved.members;
+    if (resolved.optionalMembers.length) alias.optionalMembers = resolved.optionalMembers;
+    else delete alias.optionalMembers;
+    if (resolved.activeRoute) alias.activeRoute = resolved.activeRoute;
+    else delete alias.activeRoute;
+    // A fresh route invalidates stale per-member suspensions.
+    delete alias.suspendedMembers;
+    aliases[aliasId] = alias;
+  }
+  raw.aliases = aliases;
+  const runtimes = object(raw.runtimes) ?? {};
+  for (const [runtimeId, policy] of Object.entries(doc.residency)) {
+    const runtime = object(runtimes[runtimeId]);
+    if (!runtime) throw fail(`Config has no runtime ${runtimeId}.`, 409);
+    runtime.keepWarm = policy === 'always';
+    runtime.preferredWarm = policy === 'preferred';
+    runtimes[runtimeId] = runtime;
+  }
+  raw.runtimes = runtimes;
+  if (doc.defaults) raw.defaults = { ...(object(raw.defaults) ?? {}), ...doc.defaults };
+  return raw;
+}
+
+export function createFleetProfileController({ getConfig, reload, env = process.env }) {
+  async function apply(name, { yes = false } = {}) {
+    if (yes !== true) throw fail('Review the profile and confirm with yes: true.');
+    if (!getConfig().sourcePath) throw fail('This gateway has no writable installed configuration.', 409);
+    const { safe, doc } = await readRawProfile(getConfig(), name);
+    const profile = normalizeProfileDocument(doc, safe);
+    let outcome = null;
+    await mutateConfigSource(getConfig(), (raw) => {
+      // Plan against raw source state for accurate reporting, then compose.
+      outcome = planProfileChanges({ ...raw, sourcePath: getConfig().sourcePath }, profile);
+      composeProfile(raw, profile, safe);
+    });
+    reload?.();
+    return { profile: safe, ...outcome };
+  }
+
+  async function save(name, { description = '', overwrite = false, yes = false } = {}) {
+    if (yes !== true) throw fail('Confirm capturing the current configuration with yes: true.');
+    const safe = safeName(name);
+    if (!safe) throw fail('Profile names use letters, numbers, dots, dashes, underscores.');
+    const config = getConfig();
+    if (!config.sourcePath) throw fail('This gateway has no file-backed configuration.', 409);
+    const dir = profilesDir(config);
+    await fs.mkdir(dir, { recursive: true });
+    const file = path.join(dir, safe + '.json');
+    if (!overwrite) {
+      try {
+        await fs.access(file);
+        throw fail(`Profile ${safe} already exists; pass overwrite: true to replace it.`, 409);
+      } catch (error) {
+        if (error.statusCode === 409) throw error;
+        if (error.code !== 'ENOENT') throw error;
+      }
+    }
+    const source = await loadConfig(config.sourcePath);
+    const routes = {};
+    for (const [aliasId, alias] of Object.entries(object(source.aliases) ?? {})) {
+      if (Array.isArray(alias.members) && (alias.routeProfiles || alias.activeRoute)) {
+        routes[aliasId] = alias.activeRoute ?? alias.members[0];
+      }
+    }
+    const residency = {};
+    for (const [runtimeId, runtime] of Object.entries(object(source.runtimes) ?? {})) {
+      if (runtime.keepWarm === true) residency[runtimeId] = 'always';
+      else if (runtime.preferredWarm === true) residency[runtimeId] = 'preferred';
+    }
+    const doc = {
+      name: safe,
+      description: String(description).slice(0, 500),
+      routes,
+      residency,
+      defaults: object(source.defaults) ? { ...object(source.defaults) } : undefined
+    };
+    for (const key of Object.keys(doc)) if (doc[key] == null) delete doc[key];
+    const tmp = file + '.tmp-' + process.pid;
+    await fs.writeFile(tmp, JSON.stringify(doc, null, 2) + '\n');
+    await fs.rename(tmp, file);
+    return { profile: safe, file, routes: Object.keys(routes).length, residency: Object.keys(residency).length };
+  }
+
+  return {
+    list: () => listProfiles(getConfig()),
+    read: (name) => readProfile(getConfig(), name),
+    plan: async (name) => {
+      const { safe, doc } = await readRawProfile(getConfig(), name);
+      const profile = normalizeProfileDocument(doc, safe);
+      return { profile: safe, ...planProfileChanges(getConfig(), profile) };
+    },
+    apply,
+    save
+  };
+}
diff --git a/src/config.mjs b/src/config.mjs
index f496a2f..9ed7402 100644
--- a/src/config.mjs
+++ b/src/config.mjs
@@ -370,6 +370,30 @@ function validateConfig(config, sourcePath, env) {
     }
   }
 
+  const memorySafety = config.runtimePolicy?.memorySafety;
+  if (memorySafety != null) {
+    if (typeof memorySafety !== 'object' || Array.isArray(memorySafety)) {
+      errors.push('runtimePolicy.memorySafety must be an object');
+    } else {
+      if (memorySafety.mode != null && !['enforce', 'yolo'].includes(memorySafety.mode)) {
+        errors.push('runtimePolicy.memorySafety.mode must be enforce or yolo');
+      }
+      for (const key of ['minAvailableMemoryGb', 'maxMemoryUtilization', 'pollIntervalMs']) {
+        const value = memorySafety[key];
+        if (value == null) continue;
+        if (
+          typeof value !== 'number' ||
+          !Number.isFinite(value) ||
+          value <= 0 ||
+          (key === 'maxMemoryUtilization' && value >= 1) ||
+          (key === 'pollIntervalMs' && (value < 50 || value > 1000))
+        ) {
+          errors.push(`runtimePolicy.memorySafety.${key} is outside its safe range`);
+        }
+      }
+    }
+  }
+
   const distributedMemberGroups = new Map();
   for (const [runtimeId, runtime] of Object.entries(config.runtimes ?? {})) {
     if (runtime?.placement?.mode !== 'distributed') continue;
diff --git a/src/darwin-memory-usage.py b/src/darwin-memory-usage.py
new file mode 100644
index 0000000..5944cc9
--- /dev/null
+++ b/src/darwin-memory-usage.py
@@ -0,0 +1,41 @@
+"""Read Darwin's per-process physical footprint; never inspect process contents.
+
+The rusage_info_v2 ABI is defined by Apple's sys/resource.h (macOS 10.9+).
+Physical footprint includes charged memory that RSS omits, notably Metal and
+compressed allocations. A failed per-PID observation stays absent.
+"""
+import ctypes
+import json
+import sys
+
+
+class RusageInfoV2(ctypes.Structure):
+    _fields_ = [('uuid', ctypes.c_uint8 * 16)] + [
+        (name, ctypes.c_uint64) for name in (
+            'user_time', 'system_time', 'pkg_idle_wkups', 'interrupt_wkups',
+            'pageins', 'wired_size', 'resident_size', 'phys_footprint',
+            'proc_start_abstime', 'proc_exit_abstime', 'child_user_time',
+            'child_system_time', 'child_pkg_idle_wkups', 'child_interrupt_wkups',
+            'child_pageins', 'child_elapsed_abstime', 'diskio_bytesread',
+            'diskio_byteswritten'
+        )
+    ]
+
+
+def main():
+    libproc = ctypes.CDLL('/usr/lib/libproc.dylib')
+    libproc.proc_pid_rusage.argtypes = [ctypes.c_int, ctypes.c_int, ctypes.c_void_p]
+    libproc.proc_pid_rusage.restype = ctypes.c_int
+    result = {}
+    for value in sys.argv[1:]:
+        pid = int(value)
+        if not 0 < pid < 2 ** 31:
+            continue
+        usage = RusageInfoV2()
+        if libproc.proc_pid_rusage(pid, 2, ctypes.byref(usage)) == 0:
+            result[str(pid)] = usage.phys_footprint
+    print(json.dumps(result))
+
+
+if __name__ == '__main__':
+    main()
diff --git a/src/dashboard-installation.mjs b/src/dashboard-installation.mjs
new file mode 100644
index 0000000..367b83f
--- /dev/null
+++ b/src/dashboard-installation.mjs
@@ -0,0 +1,155 @@
+import { createHash } from 'node:crypto';
+import fs from 'node:fs/promises';
+import { createInstallationJobs } from './installation-jobs.mjs';
+import { createModelImportPlan } from './model-intake.mjs';
+import { createSetupPlan, applySetup } from './setup.mjs';
+import { loadRecipeById } from './recipes.mjs';
+import { loadBackendCatalog, getBackend, planBackend } from './backend-catalog.mjs';
+import { loadConfig } from './config.mjs';
+import { mutateConfigSource } from './config-mutation.mjs';
+import { pinDownloadCommands } from './installer.mjs';
+import { installImportedModelAssets } from './model-installation.mjs';
+import { validateAcquisitionStep } from './model-acquisition.mjs';
+
+const digest = (value) => createHash('sha256').update(JSON.stringify(value)).digest('hex');
+const invalid = (message) => Object.assign(new Error(message), { statusCode: 409 });
+export function createDashboardInstallation({ getConfig, reload, env = process.env }) {
+  const source = async () => {
+    if (!getConfig().sourcePath) throw invalid('Installations need a writable LLooM configuration.');
+    return loadConfig(getConfig().sourcePath);
+  };
+  const evidence = async (recipeId) => ({
+    catalog: await loadBackendCatalog(),
+    recipe: recipeId ? await loadRecipeById(recipeId) : null
+  });
+  return createInstallationJobs({
+    async plan(input) {
+      if (!input || Object.keys(input).some((key) => !['recipeId', 'modelRef', 'backend', 'name'].includes(key)))
+        throw invalid('Choose a vendor recipe or model reference.');
+      if (Boolean(input.recipeId) === Boolean(input.modelRef))
+        throw invalid('Choose exactly one recipe or model reference.');
+      for (const value of Object.values(input))
+        if (typeof value !== 'string' || value.length > 2000) throw invalid('Invalid model reference.');
+      const config = await source();
+      const baseline = digest(JSON.parse(await fs.readFile(config.sourcePath, 'utf8')));
+      const sources = await evidence(input.recipeId);
+      const options = {
+        ...input,
+        configPath: config.sourcePath,
+        additive: true,
+        offline: true,
+        start: false,
+        includeRuntimes: false
+      };
+      const plan = input.recipeId ? await createSetupPlan(config, options) : createModelImportPlan(config, options);
+      if (
+        !input.recipeId &&
+        plan.reference.type === 'huggingface' &&
+        !/^[a-f0-9]{40,64}$/i.test(plan.reference.revision || '')
+      )
+        throw invalid(
+          'Use a Hugging Face file or repository link pinned to a commit, such as /tree/<commit>. Branches and latest references cannot bind a reviewed installation.'
+        );
+      if (!input.recipeId && plan.additions.runtimeId) {
+        const backend = getBackend(sources.catalog, plan.inference.backend);
+        if (!backend) throw invalid('Unknown vendor backend.');
+        plan.backendPlan = await planBackend(backend);
+      }
+      if (!input.recipeId && plan.download?.command?.[0] === 'hf') {
+        const command = plan.download.command;
+        const pinned = await pinDownloadCommands(
+          {
+            steps: [
+              {
+                action: 'download-model',
+                provider: 'huggingface',
+                model: command[2],
+                revision: plan.reference.revision,
+                include: plan.reference.filePath ? [plan.reference.filePath] : [],
+                destination: command[command.indexOf('--local-dir') + 1],
+                command,
+                commands: [command]
+              }
+            ]
+          },
+          { env, requireAvailable: true }
+        );
+        const errors = validateAcquisitionStep(pinned.steps[0]);
+        if (plan.reference.filePath?.startsWith('-'))
+          errors.push('Model file names cannot begin with a command option.');
+        if (errors.length) throw invalid(errors.join('; '));
+        plan.download.acquisition = pinned.steps[0];
+        plan.download.command = pinned.steps[0].command;
+        delete plan.download.shellCommand;
+      }
+      return {
+        input,
+        options,
+        plan,
+        baseline,
+        sources,
+        digest: digest(sources),
+        view: {
+          kind: input.recipeId ? 'recipe' : 'model',
+          summary:
+            'Installs the vendor backend and model files, then adds the model to this gateway. Loading uses normal memory admission.',
+          details: input.recipeId
+            ? {
+                recipe: plan.selectedRecipe,
+                configPath: plan.configPath,
+                modelRoot: plan.modelRoot,
+                ports: plan.ports,
+                backend: plan.phases.bootstrap.backend,
+                models: plan.phases.bootstrap.recipe,
+                integrations: plan.phases.bootstrap.integrations
+              }
+            : {
+                reference: plan.reference,
+                backend: plan.inference,
+                backendInstallation: plan.backendPlan,
+                additions: plan.additions,
+                download: plan.download,
+                configPath: plan.configPath
+              }
+        }
+      };
+    },
+    async apply(prepared, onProgress) {
+      if (digest(await evidence(prepared.input.recipeId)) !== prepared.digest)
+        throw invalid('The vendor recipe changed. Review a fresh plan.');
+      const config = await source();
+      if (digest(JSON.parse(await fs.readFile(config.sourcePath, 'utf8'))) !== prepared.baseline)
+        throw invalid('Your configuration changed. Review a fresh installation plan.');
+      const writeConfig = async (file, value) => {
+        if (file !== config.sourcePath) throw invalid('The installation destination changed.');
+        await mutateConfigSource(config, (raw) => {
+          if (digest(raw) !== prepared.baseline)
+            throw invalid(
+              'Your configuration changed during installation. Files are retained; review a fresh plan to continue.'
+            );
+          for (const key of Object.keys(raw)) delete raw[key];
+          Object.assign(raw, value);
+        });
+      };
+      let result;
+      if (prepared.input.recipeId) {
+        result = await applySetup(config, {
+          ...prepared.options,
+          reviewedPlan: prepared.plan,
+          dryRun: false,
+          yes: true,
+          onProgress,
+          writeConfig
+        });
+      } else {
+        await installImportedModelAssets(prepared.plan, { onProgress, backendCatalog: prepared.sources.catalog });
+        await writeConfig(config.sourcePath, prepared.plan.config);
+        result = { ok: true };
+      }
+      // Even failed bootstrap can have written configuration. Surface reload
+      // failures, and make the installed/failed state visible in the dashboard.
+      await reload();
+      return result;
+    }
+  });
+}
diff --git a/src/dashboard-memory.mjs b/src/dashboard-memory.mjs
new file mode 100644
index 0000000..125e8fc
--- /dev/null
+++ b/src/dashboard-memory.mjs
@@ -0,0 +1,433 @@
+function buildMemoryMap({ node = null, runtimes = {}, models = [], previewModelId = null, memorySafety = null } = {}) {
+  if (node?.local !== true && node?.runtimeManager?.runtimes) {
+    runtimes = { ...runtimes };
+    for (const [id, observed] of Object.entries(node.runtimeManager.runtimes)) {
+      runtimes[id] = { ...runtimes[id], ...observed, node: node.id, remote: false };
+    }
+  }
+  const GiB = 1024 * 1024 * 1024;
+
+  function numberOrNull(value) {
+    return typeof value === 'number' && Number.isFinite(value) ? value : null;
+  }
+
+  function nonNegativeOrNull(value) {
+    const number = numberOrNull(value);
+    return number !== null && number >= 0 ? number : null;
+  }
+
+  function bounded(value, minimum, maximum) {
+    const number = numberOrNull(value);
+    return number === null || number < minimum || number > maximum ? null : number;
+  }
+
+  function positiveOrNull(value) {
+    const number = numberOrNull(value);
+    return number !== null && number > 0 ? number : null;
+  }
+
+  function sameNode(runtime, currentNode) {
+    if (
+      !runtime ||
+      runtime.remote === true ||
+      runtime.distributed === true ||
+      runtime.placement?.mode === 'distributed'
+    )
+      return false;
+    const placementNode = runtimeNode(runtime);
+    if (placementNode) return placementNode === currentNode;
+    return node?.local === true;
+  }
+
+  function runtimeNode(runtime) {
+    return runtime?.node ?? runtime?.placement?.node ?? null;
+  }
+
+  function safetyReserve(policy, total) {
+    if (!policy) return null;
+    if (policy.mode === 'yolo') return 0;
+    const minAvailable = nonNegativeOrNull(policy.minAvailableMemoryGb);
+    const maxUtilization = nonNegativeOrNull(policy.maxMemoryUtilization);
+    if (minAvailable === null || maxUtilization === null || maxUtilization > 1) return null;
+    return Math.min(total, Math.max(minAvailable * GiB, total * (1 - maxUtilization)));
+  }
+
+  function colorIndex(value) {
+    const text = String(value ?? '');
+    let hash = 2166136261;
+    for (let index = 0; index < text.length; index += 1) {
+      hash ^= text.charCodeAt(index);
+      hash = (hash * 16777619) >>> 0;
+    }
+    return hash % 8;
+  }
+
+  function targetNodes(model) {
+    return (Array.isArray(model?.targets) ? model.targets : [])
+      .map((target) => {
+        return typeof target === 'object' ? (target?.node ?? null) : null;
+      })
+      .filter(Boolean);
+  }
+
+  function modelRuntimeId(model) {
+    if (model?.runtime) return model.runtime;
+    const ids = uniqueSorted(
+      (model?.targets ?? [])
+        .filter((target) => !target.node || target.node === nodeId)
+        .map((target) => target.remoteRuntime ?? target.runtime)
+    );
+    return ids.length === 1 ? ids[0] : null;
+  }
+
+  function uniqueSorted(values) {
+    return [...new Set(values.filter(Boolean))].sort((left, right) => String(left).localeCompare(String(right)));
+  }
+
+  function roundPercent(value) {
+    return Number.isFinite(value) ? Math.round(value * 1e6) / 1e6 : null;
+  }
+
+  const nodeId = node?.id ?? null;
+  const telemetryMemory = node?.telemetry?.memory;
+  const totalBytes = nonNegativeOrNull(telemetryMemory?.totalBytes);
+  const availableBytes = telemetryMemory ? bounded(telemetryMemory.availableBytes, 0, totalBytes) : null;
+  const reportedUsedBytes = telemetryMemory ? bounded(telemetryMemory.usedBytes, 0, totalBytes) : null;
+  const hasCapacity = typeof totalBytes === 'number' && totalBytes > 0;
+  const invalidAvailable = telemetryMemory?.availableBytes != null && availableBytes === null;
+  const invalidUsed = telemetryMemory?.usedBytes != null && reportedUsedBytes === null;
+  const hasMemoryFacts =
+    hasCapacity && !invalidAvailable && !invalidUsed && (availableBytes !== null || reportedUsedBytes !== null);
+  const known =
+    (node?.reachable === true || (node?.local === true && node?.reachable !== false)) && hasCapacity && hasMemoryFacts;
+  let usedBytes;
+  let availableMemoryBytes;
+  if (hasCapacity && availableBytes !== null) {
+    availableMemoryBytes = availableBytes;
+    usedBytes = totalBytes - availableBytes;
+  } else if (hasCapacity && reportedUsedBytes !== null) {
+    usedBytes = reportedUsedBytes;
+    availableMemoryBytes = totalBytes - usedBytes;
+  } else {
+    usedBytes = null;
+    availableMemoryBytes = null;
+  }
+  const activePolicy = node?.local === true ? memorySafety : node?.runtimeManager?.memorySafety;
+  const reserveBytes = hasMemoryFacts ? safetyReserve(activePolicy, totalBytes) : null;
+  const usableBytes = hasMemoryFacts && reserveBytes !== null ? Math.max(0, availableMemoryBytes - reserveBytes) : null;
+
+  if (!known) {
+    return {
+      known: false,
+      totalBytes: hasCapacity ? totalBytes : null,
+      usedBytes: null,
+      availableBytes: hasCapacity ? availableBytes : null,
+      reserveBytes: null,
+      usableBytes: null,
+      nodeId,
+      segments: [],
+      attributionNote: 'Memory telemetry is unavailable; no allocation is inferred.',
+      preview: null
+    };
+  }
+
+  const runtimeIds = Object.keys(runtimes ?? {}).sort((left, right) => left.localeCompare(right));
+  const includedRuntimeIds = [];
+  for (const runtimeId of runtimeIds) {
+    const runtime = runtimes[runtimeId];
+    const status = String(runtime?.status ?? '').toLowerCase();
+    const active = ['running', 'healthy', 'starting', 'warming', 'external', 'stopping', 'draining'].includes(status);
+    if (!active || runtime?.paused === true) continue;
+    if (!sameNode(runtime, nodeId)) continue;
+    includedRuntimeIds.push(runtimeId);
+  }
+
+  const runtimeGroups = new Map();
+  for (const runtimeId of includedRuntimeIds) {
+    const runtime = runtimes[runtimeId];
+    const usage = runtime?.memoryUsage;
+    const groupId = typeof usage?.groupId === 'string' && usage.groupId ? usage.groupId : `runtime:${runtimeId}`;
+    if (!runtimeGroups.has(groupId)) {
+      runtimeGroups.set(groupId, { runtimeIds: [], usage: null });
+    }
+    const group = runtimeGroups.get(groupId);
+    group.runtimeIds.push(runtimeId);
+    if (!group.usage) group.usage = usage ?? null;
+  }
+
+  const attributed = [];
+  for (const [groupId, group] of runtimeGroups) {
+    const primaryRuntimeId = group.runtimeIds[0];
+    const runtime = runtimes[primaryRuntimeId];
+    const usages = group.runtimeIds.map((runtimeId) => runtimes[runtimeId]?.memoryUsage).filter(Boolean);
+    const rssValues = usages.map((usage) => nonNegativeOrNull(usage.residentBytes)).filter((value) => value !== null);
+    const rssBytes = rssValues.length ? Math.max(...rssValues) : null;
+    const estimateValues = group.runtimeIds
+      .map((runtimeId) => positiveOrNull(runtimes[runtimeId]?.memoryGb))
+      .filter((value) => value !== null)
+      .map((value) => value * GiB);
+    const bytes = rssBytes !== null ? rssBytes : estimateValues.length ? Math.max(...estimateValues) : null;
+    if (bytes === null) continue;
+    const modelIds = new Set();
+    for (const runtimeId of group.runtimeIds) {
+      for (const model of models ?? []) {
+        if (modelRuntimeId(model) === runtimeId) {
+          if (model.id) modelIds.add(model.id);
+        }
+      }
+    }
+    const names = uniqueSorted([...modelIds].map((id) => models.find((model) => model.id === id)?.name || id));
+    const label = names.length
+      ? names.slice(0, 2).join(' + ') + (names.length > 2 ? ' +' + (names.length - 2) : '')
+      : (runtime?.id ?? primaryRuntimeId);
+    const estimated = rssBytes === null;
+    attributed.push({
+      id: groupId,
+      kind: 'runtime',
+      label,
+      bytes: Math.round(bytes),
+      percent: roundPercent((bytes / totalBytes) * 100),
+      runtimeId: primaryRuntimeId,
+      modelIds: uniqueSorted([...modelIds]),
+      estimated
+    });
+  }
+
+  let reconciliationScale = 1;
+  let attributedBytes = attributed.reduce((sum, segment) => sum + segment.bytes, 0);
+  if (attributedBytes > usedBytes) {
+    reconciliationScale = usedBytes / attributedBytes;
+    for (const segment of attributed) {
+      segment.bytes = Math.round(segment.bytes * reconciliationScale);
+      segment.estimated = true;
+    }
+    attributedBytes = attributed.reduce((sum, segment) => sum + segment.bytes, 0);
+  }
+  attributed.sort((left, right) => String(left.runtimeId).localeCompare(String(right.runtimeId)));
+  const colors = new Set();
+  for (const segment of attributed) {
+    let index = colorIndex(segment.runtimeId);
+    if (colors.size < 8) while (colors.has(index)) index = (index + 1) % 8;
+    colors.add(index);
+    segment.colorIndex = index;
+    segment.percent = roundPercent((segment.bytes / totalBytes) * 100);
+  }
+
+  if (attributedBytes > usedBytes && attributed.length) {
+    const excess = attributedBytes - usedBytes;
+    const largest = attributed.reduce((left, right) => (right.bytes > left.bytes ? right : left));
+    largest.bytes = Math.max(0, largest.bytes - excess);
+    attributedBytes = attributed.reduce((sum, segment) => sum + segment.bytes, 0);
+  }
+  for (const segment of attributed) segment.percent = roundPercent((segment.bytes / totalBytes) * 100);
+  const systemBytes = Math.max(0, usedBytes - attributedBytes);
+
+  const segments = [...attributed];
+  if (systemBytes > 0 || usedBytes === 0) {
+    segments.push({
+      id: 'system',
+      kind: 'system',
+      label: 'System & other apps',
+      bytes: systemBytes,
+      percent: roundPercent((systemBytes / totalBytes) * 100),
+      runtimeId: null,
+      modelIds: [],
+      estimated: false,
+      colorIndex: 8
+    });
+  }
+  const availableSegmentBytes = totalBytes - usedBytes;
+  segments.push({
+    id: 'available',
+    kind: 'available',
+    label: 'Available',
+    bytes: availableSegmentBytes,
+    percent: roundPercent((availableSegmentBytes / totalBytes) * 100),
+    runtimeId: null,
+    modelIds: [],
+    estimated: false,
+    colorIndex: 9
+  });
+
+  const attributionParts = ['Live process memory is approximate. Shared models use one block. Previews are estimates.'];
+  if (reconciliationScale !== 1)
+    attributionParts.push('App measurements overlap; blocks are adjusted to match total memory in use.');
+  if (
+    [...runtimeGroups.values()].some((group) =>
+      group.runtimeIds.every((runtimeId) => {
+        const usage = runtimes[runtimeId]?.memoryUsage;
+        return !usage || nonNegativeOrNull(usage.residentBytes) === null;
+      })
+    )
+  )
+    attributionParts.push('Striped blocks use an estimate until a live reading is available.');
+
+  const preview = previewModelId ? previewModel(previewModelId) : null;
+  function previewModel(modelId) {
+    const model = (models ?? []).find((item) => item?.id === modelId);
+    if (!model) return null;
+    const runtimeId = modelRuntimeId(model);
+    const runtime = runtimeId ? runtimes[runtimeId] : null;
+    const label = model.name ?? model.id;
+    const base = {
+      modelId,
+      label,
+      runtimeId,
+      status: 'unknown',
+      additionalBytes: null,
+      remainingBytes: null,
+      projectedUsedBytes: null,
+      percent: null,
+      nodeId,
+      shared: false,
+      message: ''
+    };
+    const targets = targetNodes(model);
+    const otherTargets = targets.filter((target) => target !== nodeId);
+    const allTargetsAreOther = targets.length > 0 && otherTargets.length === targets.length;
+    const runtimePlacementNode = runtime ? runtimeNode(runtime) : null;
+    const runtimeIsOtherNode =
+      (runtime?.remote === true && runtimePlacementNode === null) ||
+      (runtimePlacementNode !== null && runtimePlacementNode !== nodeId);
+    if (runtimeIsOtherNode || allTargetsAreOther) {
+      return {
+        ...base,
+        status: 'other-node',
+        additionalBytes: 0,
+        projectedUsedBytes: usedBytes,
+        remainingBytes: availableMemoryBytes,
+        percent: roundPercent((usedBytes / totalBytes) * 100),
+        nodeId: targets[0] ?? runtimePlacementNode,
+        message: `Runs on node ${targets[0] ?? runtimePlacementNode ?? 'another node'}; no impact on this machine.`
+      };
+    }
+    if (runtime?.distributed === true || runtime?.placement?.mode === 'distributed' || new Set(targets).size > 1) {
+      return { ...base, status: 'unknown', shared: true, message: 'Needs room on each machine.' };
+    }
+    if (
+      runtime?.maintenance ||
+      runtime?.enabled === false ||
+      runtime?.paused === true ||
+      runtime?.status === 'paused' ||
+      model.paused === true
+    ) {
+      return { ...base, status: 'paused', message: 'Paused; automatic starts are disabled.' };
+    }
+    if (!runtimeId && (model.federated || targets.length > 0))
+      return { ...base, status: 'unknown', message: 'Waiting for model memory from this machine.' };
+    if (!runtimeId) {
+      return {
+        ...base,
+        status: 'external',
+        additionalBytes: 0,
+        projectedUsedBytes: usedBytes,
+        remainingBytes: availableMemoryBytes,
+        percent: roundPercent((usedBytes / totalBytes) * 100),
+        message: 'No local model load is expected.'
+      };
+    }
+    if (!runtime) {
+      return { ...base, status: 'unknown', message: 'No runtime status is available.' };
+    }
+    const usage = runtime.memoryUsage;
+    const loadedIds = Array.isArray(usage?.loadedModelIds) ? usage.loadedModelIds : null;
+    const wantedId = model.upstreamModel ?? model.id;
+    const residentConfirmed =
+      runtime.healthy === true &&
+      usage?.residencyKnown === true &&
+      loadedIds !== null &&
+      (loadedIds.includes(wantedId) || loadedIds.includes(model.id));
+    const commandParts = String(runtime.command ?? '')
+      .trim()
+      .split(/\s+/);
+    const commandBase = commandParts[0]?.split(/[\\/]/).pop() ?? '';
+    const lazyBackend = [commandBase, ...(runtime.args ?? [])].some((part) =>
+      /(?:^|[/\\])(ollama|lloom-audio-server|lloom_audio_server(?:\.py)?)$/.test(String(part))
+    );
+    const sharedBackend = usage?.sharedRuntimeIds?.length > 1 || lazyBackend;
+    const modelsOnRuntime = (models ?? []).filter((item) => modelRuntimeId(item) === runtimeId);
+    const runtimeIsSharedGroup = runtimeGroups.get(usage?.groupId ?? `runtime:${runtimeId}`)?.runtimeIds.length > 1;
+    if (residentConfirmed) {
+      return {
+        ...base,
+        status: 'resident',
+        additionalBytes: 0,
+        projectedUsedBytes: usedBytes,
+        remainingBytes: availableMemoryBytes,
+        percent: roundPercent((usedBytes / totalBytes) * 100),
+        shared: sharedBackend || runtimeIsSharedGroup,
+        message:
+          sharedBackend || runtimeIsSharedGroup
+            ? 'Already available; shared backend footprint can change.'
+            : 'Already available'
+      };
+    }
+    if (sharedBackend && runtime.healthy === true && !(usage?.residencyKnown && loadedIds !== null)) {
+      return { ...base, status: 'unknown', shared: true, message: 'Shared backend residency is unknown.' };
+    }
+    const memoryGb = positiveOrNull(runtime.memoryGb);
+    if (memoryGb === null) {
+      return { ...base, status: 'unknown', message: 'Cold start estimate is unknown.' };
+    }
+    // A healthy server may load weights lazily. Only confirmed model-cache
+    // observations promise residency; otherwise forecast growth toward its peak.
+    const measured =
+      !sharedBackend && !runtimeIsSharedGroup && modelsOnRuntime.length === 1 && runtime.healthy === true
+        ? nonNegativeOrNull(usage?.residentBytes)
+        : null;
+    const incoming = Math.max(0, memoryGb * GiB - (measured ?? 0));
+    const remaining = availableMemoryBytes - incoming;
+    const reserve = reserveBytes ?? null;
+    if (reserve === null) {
+      return {
+        ...base,
+        status: 'unknown',
+        additionalBytes: incoming,
+        remainingBytes: remaining,
+        projectedUsedBytes: usedBytes + incoming,
+        percent: roundPercent((incoming / totalBytes) * 100),
+        shared: sharedBackend,
+        message: 'Memory reserve is unknown; no fit is guaranteed.'
+      };
+    }
+    if (remaining <= reserve) {
+      return {
+        ...base,
+        status: 'blocked',
+        additionalBytes: incoming,
+        remainingBytes: remaining,
+        projectedUsedBytes: usedBytes + incoming,
+        percent: roundPercent((incoming / totalBytes) * 100),
+        shared: sharedBackend,
+        message: 'Does not fit within the reserve.'
+      };
+    }
+    const tightLimit = reserve + Math.max(GiB, totalBytes * 0.02);
+    const tight = remaining <= tightLimit;
+    return {
+      ...base,
+      status: tight ? 'tight' : 'fits',
+      additionalBytes: incoming,
+      remainingBytes: remaining,
+      projectedUsedBytes: usedBytes + incoming,
+      percent: roundPercent((incoming / totalBytes) * 100),
+      shared: sharedBackend,
+      message: tight ? 'Expected to fit, with little reserve headroom.' : 'Expected to fit'
+    };
+  }
+
+  return {
+    known: true,
+    totalBytes,
+    usedBytes,
+    availableBytes: availableMemoryBytes,
+    reserveBytes,
+    usableBytes,
+    nodeId,
+    segments,
+    attributionNote: attributionParts.join(' '),
+    preview
+  };
+}
+
+export { buildMemoryMap };
diff --git a/src/dashboard-presence-client.mjs b/src/dashboard-presence-client.mjs
new file mode 100644
index 0000000..5289a96
--- /dev/null
+++ b/src/dashboard-presence-client.mjs
@@ -0,0 +1,319 @@
+// Runs inside the dashboard's existing script, sharing its authenticated API
+// helpers and observed state. It never invents topology or hardware readings.
+export const presenceScript = String.raw`
+    let presenceView = "live";
+    let presenceImport = null;
+    let presencePlanVersion = 0;
+    let presenceLastFocus = null;
+    let presenceBusy = false;
+    let presenceRenderKey = "";
+    let presenceMachineKey = "";
+    let presenceInstallTimer = null;
+    let presenceInstallSeen = null;
+    const presencePolicyNames = { auto:"Automatic", preferred:"Prefer instant replies", always:"Keep ready" };
+    function presenceNotice(message, error = false) {
+      const toast = $("#presence-toast");
+      toast.querySelector("span").textContent = message;
+      toast.dataset.error = String(error);
+      toast.hidden = false;
+    }
+    function presenceSetView(name) {
+      if (!["live","models","machines","clients","settings"].includes(name)) name = "live";
+      presenceView = name;
+      document.querySelectorAll(".presence-nav [data-view]").forEach(button => {
+        if (button.dataset.view === name) button.setAttribute("aria-current","page");
+        else button.removeAttribute("aria-current");
+      });
+      document.querySelectorAll("[data-presence-panel]").forEach(panel => panel.hidden = panel.dataset.presencePanel !== name);
+      for (const id of ["models","machines","clients"]) $("#view-" + id).hidden = id !== name;
+      $(".operations-dock").open = name === "settings";
+      closeModelInspector(); closeNodeInspector();
+      if (name === "clients") presenceLoadIntegrations();
+      try { history.replaceState(null,"","#" + name); } catch {}
+      renderPresence();
+      window.scrollTo({top:0,behavior:"instant"});
+    }
+    function presenceRuntime(model) {
+      return model.runtime ? state.status?.runtimeManager?.runtimes?.[model.runtime] : null;
+    }
+    function presencePolicy(runtime) {
+      return runtime?.keepWarm ? "always" : runtime?.preferredWarm ? "preferred" : "auto";
+    }
+    function presenceModelResident(model) {
+      const rt=presenceRuntime(model),usage=rt?.memoryUsage;
+      if(!rt?.healthy)return false;
+      if(usage?.residencyKnown&&Array.isArray(usage.loadedModelIds))return usage.loadedModelIds.some(id=>id===model.id||id===model.upstreamModel);
+      return false;
+    }
+    function presenceModelLabel(model) {
+      const runtime = presenceRuntime(model);
+      if (model.alias) return "Route";
+      if (!model.runtime) return model.federated ? "Shared model" : "External provider";
+      if (runtime?.maintenance) return "Paused";
+      const transition={starting:"Getting ready",warming:"Getting ready",queued:"Waiting for room",stopping:"Freeing memory",draining:"Finishing work",failed:"Needs attention",unreachable:"Unavailable",disabled:"Disabled"}[runtime?.status];
+      if(transition)return transition;
+      if (runtime?.activeRequests > 0 && runtime?.healthy) return "Serving";
+      if (presenceModelResident(model)) return "Ready to use";
+      return "Starts when needed";
+    }
+    function renderPresenceModels() {
+      const search = $("#presence-search").value.trim().toLowerCase();
+      const kind = $("#presence-kind").value;
+      const models = (state.physicalModels || []).filter(model =>
+        (!search || (model.name + " " + model.id).toLowerCase().includes(search)) &&
+        (!kind || (model.kind || "chat").startsWith(kind))
+      ).sort((a,b) => Number(Boolean(b.runtime)) - Number(Boolean(a.runtime)) || Number(Boolean(presenceRuntime(b)?.healthy)) - Number(Boolean(presenceRuntime(a)?.healthy)));
+      $("#presence-model-count").textContent = models.length + (models.length === 1 ? " model" : " models");
+      const key = JSON.stringify(models.map(model => {const rt=presenceRuntime(model);return [model.id,model.name,model.kind,model.contextWindow,model.runtime,presenceModelLabel(model),rt?.memoryGb,rt?.node,rt?.keepWarm,rt?.preferredWarm];}));
+      if (key === presenceRenderKey) return;
+      presenceRenderKey = key;
+      $("#presence-models").innerHTML = models.map(model => {
+        const rt = presenceRuntime(model), policy = presencePolicy(rt);
+        const location = rt?.node || (model.targets || []).map(t => t.node).filter(Boolean).join(", ") || (model.runtime ? "This machine" : "Upstream");
+        return '<article class="presence-card"><div class="presence-card-top"><span class="presence-eyebrow">' + escapeHtml(({chat:"Chat & code",audio_speech:"Speech",audio_transcription:"Transcription",audio_generation:"Music & audio",embedding:"Embeddings",image:"Images",video:"Video"})[model.kind] || model.kind || "Chat") + '</span><span class="presence-status" data-ready="' + String(Boolean(rt?.healthy || rt?.activeRequests)) + '">' + escapeHtml(presenceModelLabel(model)) + '</span></div><h3>' + escapeHtml(model.name || model.id) + '</h3><p>' + escapeHtml(model.id) + '</p><dl><dt>Runs on</dt><dd>' + escapeHtml(location) + '</dd><dt>Memory estimate</dt><dd>' + (rt?.memoryGb != null ? escapeHtml(rt.memoryGb) + ' GB' : 'Not reported') + '</dd><dt>Availability</dt><dd>' + escapeHtml(rt ? presencePolicyNames[policy] : 'Managed upstream') + '</dd></dl><div class="actions"><button type="button" data-presence-model="' + escapeHtml(model.id) + '">Manage model</button></div></article>';
+      }).join("") || '<div class="empty">' + (state.models.length ? 'No models match these filters.' : 'No configured models yet. Add a model to get started.') + '</div>';
+    }
+    function renderPresenceMachines() {
+      const nodes = Object.values(state.status?.cluster?.nodes || {});
+      const rows = nodes.length ? nodes : (state.topologySummary?.host ? [{id:"local",name:"This machine",local:true,telemetry:state.topologySummary.host}] : []);
+      const key=JSON.stringify(rows.map(node=>[node.id,node.name,node.local,node.reachable,node.profile?.cpuBrand,node.telemetry?.memory?.usedBytes,node.telemetry?.memory?.availableBytes,node.telemetry?.memory?.totalBytes]));
+      if(key===presenceMachineKey)return;
+      presenceMachineKey=key;
+      $("#presence-machines").innerHTML = rows.map(node => {
+        const memory = node.telemetry?.memory || {};
+        const total = Number(memory.totalBytes);
+        const available = Number(memory.availableBytes);
+        const used = Number.isFinite(available) ? total - available : Number(memory.usedBytes);
+        const hasMemory = total > 0 && Number.isFinite(used);
+        return '<article class="presence-card"><div class="presence-card-top"><span class="presence-eyebrow">' + (node.local ? 'This machine' : 'Configured peer') + '</span><span class="presence-status" data-ready="' + String(node.reachable !== false) + '">' + (node.reachable === false ? 'Unavailable' : 'Connected') + '</span></div><h3>' + escapeHtml(node.name || node.id) + '</h3><p>' + escapeHtml(node.profile?.cpuBrand || node.telemetry?.cpu?.model || node.profile?.platformId || 'Hardware details unavailable') + '</p>' + (hasMemory ? '<div class="presence-memory" aria-label="' + escapeHtml(formatBytes(used) + ' of ' + formatBytes(total) + ' memory used') + '"><span style="width:' + Math.max(0,Math.min(100,used/total*100)) + '%"></span></div><p>' + escapeHtml(formatBytes(Number.isFinite(available) ? available : used)) + (Number.isFinite(available) ? ' available / ' : ' used / ') + escapeHtml(formatBytes(total)) + '</p>' : '<p>Memory reading unavailable</p>') + '<div class="actions"><button type="button" data-presence-node="' + escapeHtml(node.id) + '"' + (state.status?.cluster?.nodes?.[node.id] ? '' : ' disabled') + '>Machine details</button></div></article>';
+      }).join("") || '<div class="empty">Waiting for hardware telemetry.</div>';
+    }
+    function presenceClientExample() {
+      const model = $("#presence-client-model").value;
+      const selected=(state.physicalModels||[]).find(m=>m.id===model),kind=selected?.kind||'chat';
+      const templates={chat:['/v1/chat/completions',{model,messages:[{role:'user',content:'Hello'}]}],embedding:['/v1/embeddings',{model,input:'Text to search'}],image:['/v1/images/generations',{model,prompt:'A quiet mountain lake'}],audio_speech:['/v1/audio/speech',{model,input:'Hello there',voice:'default'}],audio_generation:['/v1/audio/generations',{model,prompt:'Gentle piano'}],video:['/v1/videos/generations',{model,prompt:'A quiet mountain lake'}]};
+      if(kind==='audio_transcription'){$("#presence-client-example").textContent="POST "+endpoint+"/v1/audio/transcriptions\nAuthorization: Bearer YOUR_LLOOM_KEY\nMultipart form: model="+model+", file=YOUR_AUDIO_FILE";return;}
+      const [route,body]=templates[kind]||templates.chat;
+      $("#presence-client-example").textContent = "POST " + endpoint + route + "\nContent-Type: application/json\nAuthorization: Bearer YOUR_LLOOM_KEY\n\n" + JSON.stringify(body,null,2);
+    }
+    function renderPresenceClients() {
+      $("#presence-client-url").value = endpoint + "/v1";
+      const select = $("#presence-client-model"), selected = select.value;
+      const models = state.physicalModels || [];
+      const key = models.map(m=>m.id).join("\n");
+      if (select.dataset.models !== key) {
+        select.innerHTML = models.map(m=>'<option value="' + escapeHtml(m.id) + '">' + escapeHtml(m.name || m.id) + '</option>').join("");
+        if (models.some(m=>m.id === selected)) select.value = selected;
+        select.dataset.models = key;
+      }
+      $("#presence-client-auth").textContent = state.security?.authRequired ? "Use a configured inference key. Keep your admin key out of client applications." : "This loopback gateway allows local clients without a key. Remote access requires authenticated configuration.";
+      presenceClientExample();
+    }
+    function renderPresence() {
+      if (!$("#presence-models")) return;
+      const safety = state.status?.runtimeManager?.memorySafety;
+      const warning = $("#presence-memory-safety");
+      if(warning) warning.hidden = safety?.mode !== "yolo";
+      renderPresenceModels(); renderPresenceMachines(); renderPresenceClients(); renderPresencePolicy();
+    }
+    function renderPresencePolicy() {
+      const model = state.physicalModels.find(m=>m.id === state.selectedModelId);
+      const runtime = model && presenceRuntime(model);
+      if(model) $("#model-inspector-state span:last-child").textContent=presenceModelLabel(model);
+      $("#presence-policy").hidden = !model;
+      $("#presence-trial").hidden = !model || (model.kind || "chat") !== "chat";
+      for (const button of document.querySelectorAll("[data-residency]")) {
+        button.disabled = presenceBusy || !runtime || runtime.enabled === false || runtime.management === "external" || runtime.remote === true || Boolean(runtime.maintenance);
+        button.setAttribute("aria-pressed",String(Boolean(runtime) && presencePolicy(runtime) === button.dataset.residency));
+      }
+      const managed=runtime&&runtime.enabled!==false&&runtime.management!=="external"&&!runtime.remote&&!runtime.distributed&&!runtime.maintenance;
+      const transitioning=runtime&&['starting','warming','queued','draining','stopping'].includes(runtime.status);
+      for(const id of ['model-start','model-stop'])$("#"+id).disabled=presenceBusy||!managed||transitioning;
+      $("#model-start").textContent=transitioning?"Getting ready…":"Make ready now";
+      $("#model-stop").disabled||=Boolean(runtime?.activeRequests||runtime?.queuedRequests||runtime?.keepWarm);
+      $("#presence-send").disabled=presenceBusy||Boolean(runtime?.maintenance)||runtime?.enabled===false;
+      $("#presence-availability").textContent=!runtime?"Available through your connected provider.":runtime.maintenance?"Paused. This model is protected from starting.":transitioning?"LLooM is getting this ready for you.":runtime.status==='failed'?"This model needs attention. Open details to see what happened.":"Just use it. LLooM prepares this automatically when your app asks.";
+      $("#presence-policy-hint").textContent = !runtime?"Your provider manages availability.":presencePolicy(runtime)==="always"?"Kept ready for quick replies. Choose Automatic if you want LLooM to reclaim its memory.":presencePolicy(runtime)==="preferred"?"Stays ready when there is room. LLooM makes space when another model needs it.":"Recommended: LLooM gets this ready when needed. Your downloaded files stay on disk.";
+      $("#presence-model-error").textContent=runtime?.lastError||"";
+    }
+    async function presenceLoadIntegrations() {
+      try {
+        const manifest = await getJson("/gateway/integrations");
+        $("#presence-integrations").innerHTML = (manifest.clients || []).map(client=>'<details><summary>' + escapeHtml(client.name || client.id) + '</summary><pre>' + escapeHtml("lloom integrate " + client.id) + '</pre><p class="muted">Preview first. Add --apply --yes only after reviewing the generated configuration.</p></details>').join("") || "No client profiles available.";
+      } catch(error) { $("#presence-integrations").textContent = error.message; }
+    }
+    function presenceInvalidatePlan() {
+      presenceImport = null; presencePlanVersion++;
+      $("#presence-add-review").hidden = true; $("#presence-add-apply").hidden = true;
+      $("#presence-add-plan").hidden = false;
+    }
+    function presenceOpenAdd() {
+      presenceLastFocus = document.activeElement;
+      presenceInvalidatePlan();
+      $("#presence-add-error").textContent = "";
+      const selected = state.library?.selected;
+      $("#presence-recommendation").innerHTML = selected
+        ? '<div class="presence-card"><span class="presence-eyebrow">Recommended for this hardware</span><h3>' + escapeHtml(selected.name || selected.recipeId) + '</h3><p>' + escapeHtml((selected.reasons || []).join(" · ") || "Matched by the local recipe library.") + '</p><button type="button" id="presence-use-recipe">Review recommended setup</button></div>'
+        : '<p class="muted">No compatible recipe recommendation is available yet.</p>';
+      $("#presence-add").hidden = false; $("#presence-model-ref").focus();
+    }
+    function presenceCloseAdd() {
+      if (presenceBusy) return;
+      $("#presence-add").hidden = true; presenceLastFocus?.focus();
+    }
+    async function presenceReviewRecipe() {
+      const selected = state.library?.selected;
+      if (!selected) return;
+      const recipeId = selected.recipeId;
+      const version = presencePlanVersion;
+      const plan = await postJson("/gateway/installations/plan",{recipeId});
+      if(version!==presencePlanVersion)return;
+      presenceImport = {planId:plan.planId};
+      presenceShowPlan(plan);
+    }
+    function presenceShowPlan(plan) {
+      $("#presence-plan-json").textContent = JSON.stringify(plan.details || plan,null,2);
+      $("#presence-plan-summary").textContent = plan.summary || "Review the packages, model files, and configuration this installation will use.";
+      $("#presence-add-review").hidden = false;
+      $("#presence-add-apply").hidden = plan.ok === false;
+      $("#presence-add-plan").hidden = plan.ok !== false;
+    }
+    async function presenceRun(button, action) {
+      if (presenceBusy) return;
+      presenceBusy = true;
+      if (button) button.disabled = true;
+      renderPresencePolicy();
+      try { return await action(); }
+      catch(error) { presenceNotice(error.message,true); $("#presence-add-error").textContent = error.message; }
+      finally { presenceBusy = false; if(button) button.disabled=false; renderPresencePolicy(); }
+    }
+    async function presencePollInstallation() {
+      clearTimeout(presenceInstallTimer);
+      try {
+        const {job}=await getJson("/gateway/installations");
+        if(!job)return;
+        if(job.status==="running") {
+          presenceNotice(job.detail);
+          presenceInstallTimer=setTimeout(presencePollInstallation,1500);
+        } else if(presenceInstallSeen!==job.id) {
+          presenceInstallSeen=job.id;
+          presenceNotice(job.error || job.detail,job.status==="failed");
+          await refresh();
+        }
+      } catch(error) { presenceNotice("Could not read installation progress: "+error.message,true); }
+    }
+    // Keep the model inspector available from every view, rather than clipping
+    // it when the canvas is hidden.
+    document.body.append($("#model-inspector"),$("#node-inspector"));
+    const inspectorBody=$("#model-inspector .model-inspector-body");
+    const availability=document.createElement("p");availability.id="presence-availability";availability.className="presence-availability";inspectorBody.prepend(availability);
+    const policy = document.createElement("section");policy.id = "presence-policy";
+    policy.innerHTML = '<h3>When should this stay ready?</h3><div class="presence-readiness">' + Object.entries(presencePolicyNames).map(([id,name])=>'<button type="button" data-residency="' + id + '" aria-pressed="false">' + name + '</button>').join("") + '</div><p id="presence-policy-hint" class="muted"></p>';
+    const trial = document.createElement("section");trial.id = "presence-trial"; trial.className = "presence-trial";
+    trial.innerHTML = '<label for="presence-prompt">Ask anything</label><textarea id="presence-prompt" placeholder="What can I help you with?" maxlength="8000"></textarea><button id="presence-send" class="primary" type="button">Send message</button><pre id="presence-answer" aria-live="polite"></pre>';
+    inspectorBody.append(trial);
+    const connect=document.createElement("button");connect.type="button";connect.id="presence-connect";connect.textContent="Connect an app";inspectorBody.append(connect);
+    connect.addEventListener("click",()=>{const id=state.selectedModelId;presenceSetView("clients");if([...$("#presence-client-model").options].some(option=>option.value===id))$("#presence-client-model").value=id;presenceClientExample();});
+    const technical=document.createElement("details");technical.className="presence-advanced";
+    technical.innerHTML='<summary>Options & details</summary><p id="presence-model-error" role="status"></p>';
+    technical.append(policy,$("#model-inspector .model-inspector-actions"),$("#model-inspector-details"),$("#model-inspector-tags"));inspectorBody.append(technical);
+    $("#model-start").textContent="Make ready now";$("#model-warm").remove();$("#model-stop").textContent="Free up memory";
+    const oldRenderModels = renderModels;
+    renderModels = function() { oldRenderModels(); renderPresence(); };
+    const oldRenderInspector = renderModelInspector;
+    renderModelInspector = function() { oldRenderInspector(); renderPresencePolicy(); };
+    const oldShowOutput = showOutput;
+    showOutput = function(value) { oldShowOutput(value); presenceNotice(value?.error?.message || value?.error || "Operation complete. Details are in Settings.",Boolean(value?.error)); };
+    const keyField = $("#api-key").closest("label");
+    keyField.style.cssText = "";
+    const settingsHeader = document.createElement("div"); settingsHeader.className="presence-heading";
+    settingsHeader.innerHTML='<div><h2>Settings &amp; details.</h2><p>Authentication, recipes, runtimes, and installation plans.</p></div>';
+    $(".operations-content").prepend(settingsHeader,keyField);
+    $("#api-key").addEventListener("change",()=>{ sessionStorage.setItem("lloom_api_key",$("#api-key").value.trim()); localStorage.removeItem("lloom_api_key"); refresh(); });
+    $(".presence-nav").addEventListener("click",event=>{const button=event.target.closest("[data-view]"); if(button) presenceSetView(button.dataset.view);});
+    $("#presence-toast button").addEventListener("click",()=>$("#presence-toast").hidden=true);
+    $("#presence-search").addEventListener("input",()=>renderPresenceModels());
+    $("#presence-kind").addEventListener("change",()=>renderPresenceModels());
+    $("#presence-client-model").addEventListener("change",presenceClientExample);
+    $("#presence-copy-url").addEventListener("click",event=>presenceRun(event.currentTarget,async()=>{await navigator.clipboard.writeText(endpoint+"/v1");presenceNotice("Gateway URL copied.");}));
+    $("#presence-copy-example").addEventListener("click",event=>presenceRun(event.currentTarget,async()=>{await navigator.clipboard.writeText($("#presence-client-example").textContent);presenceNotice("Example request copied. Replace the key placeholder in your client.");}));
+    $("#presence-add-close").addEventListener("click",presenceCloseAdd);
+    $("#presence-add-form").addEventListener("input",presenceInvalidatePlan);
+    $("#presence-add-form").addEventListener("submit",event=>{
+      event.preventDefault(); const button=$("#presence-add-plan");
+      presenceRun(button,async()=>{
+        const version=presencePlanVersion;
+        const input={modelRef:$("#presence-model-ref").value.trim(),backend:$("#presence-model-backend").value.trim()||undefined,name:$("#presence-model-name").value.trim()||undefined};
+        const plan=await postJson("/gateway/installations/plan",input);
+        if(version!==presencePlanVersion)return;
+        presenceImport={planId:plan.planId}; presenceShowPlan(plan);
+      });
+    });
+    $("#presence-add-apply").addEventListener("click",event=>presenceRun(event.currentTarget,async()=>{
+      if(!presenceImport || !ensureAdminKeyIfNeeded()) return;
+      await postJson("/gateway/installations",{...presenceImport,yes:true});
+      presenceBusy=false; presenceCloseAdd(); presencePollInstallation();
+    }));
+    document.addEventListener("click",event=>{
+      const add=event.target.closest("[data-add-model]");
+      if(add) presenceOpenAdd();
+      const model=event.target.closest("[data-presence-model]");
+      if(model) {openModelInspector(model.dataset.presenceModel);$("#model-inspector-close").focus();}
+      const machine=event.target.closest("[data-presence-node]");
+      if(machine) openNodeInspector(machine.dataset.presenceNode);
+      if(event.target.closest("#presence-use-recipe")) presenceRun(event.target.closest("button"),presenceReviewRecipe);
+      const residency=event.target.closest("[data-residency]");
+      if(residency) presenceRun(residency,async()=>{
+        const model=state.physicalModels.find(m=>m.id===state.selectedModelId);
+        if(!model?.runtime || !ensureAdminKeyIfNeeded()) return;
+        const path="/gateway/runtimes/"+encodeURIComponent(model.runtime)+"/residency";
+        const accepted=await postJson(path,{policy:residency.dataset.residency,yes:true});
+        presenceNotice("Preference saved. LLooM will take care of it.");
+        const poll=async()=>{
+          try {
+            const {job}=await getJson(path);
+            if(!job || job.id!==accepted.id)return;
+            if(job.status==='pending'){setTimeout(poll,1500);return;}
+            await refresh();
+            presenceNotice(job.status==='failed'?"Could not update availability: "+job.error:job.policy==='auto'?"LLooM will prepare this when your app needs it.":"Availability updated: "+presencePolicyNames[job.policy]+".",job.status==='failed');
+          }catch(error){presenceNotice("Could not check readiness: "+error.message,true);}
+        };
+        setTimeout(poll,500);
+      });
+    });
+    // Capture legacy runtime actions once, add bounded busy/error handling, and
+    // keep preparation behind the normal admission and safety checks.
+    document.addEventListener("click",event=>{
+      const button=event.target.closest("button[data-runtime]");
+      if(!button) return;
+      event.stopImmediatePropagation();
+      if(button.disabled || !button.dataset.runtime || !ensureAdminKeyIfNeeded()) return;
+      presenceRun(button,async()=>{
+        const load=button.dataset.action==="start";
+        const action=load?"admit":button.dataset.action;
+        const result=await postJson("/gateway/runtimes/"+encodeURIComponent(button.dataset.runtime)+"/"+action,load?{apply:true,yes:true,force:false,warmup:true}:{});
+        oldShowOutput(result);await refresh();presenceNotice(load?"Ready for your apps.":"Memory released. The model stays installed.");
+      });
+    },true);
+    $("#presence-send").addEventListener("click",event=>presenceRun(event.currentTarget,async()=>{
+      const prompt=$("#presence-prompt").value.trim(), model=state.selectedModelId;
+      if(!prompt || !model) return;
+      $("#presence-answer").textContent="Waiting for "+model+"…";
+      try {
+        const result=await postJson("/v1/chat/completions",{model,messages:[{role:"user",content:prompt}],max_tokens:256,stream:false});
+        $("#presence-answer").textContent=result.choices?.[0]?.message?.content || "No text response returned.";
+      } catch(error) { $("#presence-answer").textContent=error.message; throw error; }
+    }));
+    document.addEventListener("keydown",event=>{
+      if(event.key==="Escape") {presenceCloseAdd();closeModelInspector();closeNodeInspector();}
+      const dialog=!$("#presence-add").hidden?$("#presence-add .presence-dialog"):null;
+      if(event.key==="Tab" && dialog) {
+        const items=[...dialog.querySelectorAll('button:not([disabled]),input,select,textarea,summary')].filter(el=>el.getClientRects().length);
+        const first=items[0],last=items.at(-1);
+        if(event.shiftKey && document.activeElement===first){event.preventDefault();last?.focus();}
+        else if(!event.shiftKey && document.activeElement===last){event.preventDefault();first?.focus();}
+      }
+    });
+    presenceSetView(location.hash.slice(1));
+    presencePollInstallation();
+`;
diff --git a/src/dashboard-presence.mjs b/src/dashboard-presence.mjs
new file mode 100644
index 0000000..63a863a
--- /dev/null
+++ b/src/dashboard-presence.mjs
@@ -0,0 +1,139 @@
+export const presenceStyles = `
+    :root { color-scheme:dark; --bg:#080d12; --band:#0d151d; --panel:#111c25; --panel-2:#15232e; --line:#243641; --text:#edf5f8; --muted:#9eb3c0; --accent:#38dff5; --accent-2:#85c8ff; --danger:#fb927e; }
+    body { font-family:Inter,-apple-system,BlinkMacSystemFont,"Segoe UI",sans-serif; font-size:14px; letter-spacing:0; }
+    body > header { margin-left:190px; min-height:72px; padding:16px 28px; background:var(--bg); border-bottom:1px solid var(--line); }
+    body > header .brand { display:none; }
+    .topline { width:100%; justify-content:flex-end; }
+    main { margin-left:190px; width:auto; padding:24px 28px; max-width:none; min-width:0; }
+    main > * { min-width:0; }
+    button.primary { background:#1e4953; border-color:#36717d; }
+    button,input,select,textarea { font-family:inherit; border-radius:9px; }
+    button { text-transform:none; font-size:13px; letter-spacing:0; font-weight:500; }
+    button:focus-visible, a:focus-visible, input:focus-visible, select:focus-visible { outline:2px solid var(--accent); outline-offset:3px; }
+    h1,h2,h3,strong { font-weight:500; }
+    .pill { text-transform:none; letter-spacing:0; font-size:12px; border-radius:20px; }
+    .band,.empty { border-radius:14px; }
+    .band-head h2,label { text-transform:none; letter-spacing:0; }
+    .band-body { padding:22px; }
+    #presence-memory-safety { border:1px solid #b67c37; background:#392915; color:#ffd8a0; border-radius:12px; padding:14px 18px; margin-bottom:20px; }
+    .presence-nav { position:fixed; top:0; bottom:0; left:0; width:190px; padding:26px 18px; background:#0b1219; border-right:1px solid var(--line); display:flex; flex-direction:column; gap:7px; z-index:20; }
+    .presence-brand { padding:0 12px 30px; font-size:24px; letter-spacing:-1px; font-weight:500; }
+    .presence-brand small { display:block; color:var(--muted); font-size:10px; letter-spacing:2px; margin-top:3px; }
+    .presence-nav button { display:flex; align-items:center; gap:12px; text-align:left; background:transparent; border:1px solid transparent; min-height:43px; color:var(--muted); padding:10px 12px; }
+    .presence-nav button[aria-current="page"] { color:var(--text); background:#18303a; border-color:#25515e; }
+    .presence-nav button:last-child { margin-top:auto; }
+    .presence-nav svg { width:18px; height:18px; fill:none; stroke:currentColor; stroke-width:1.6; }
+    .presence-heading { display:flex; justify-content:space-between; align-items:center; flex-wrap:wrap; gap:16px; margin-bottom:24px; }
+    .presence-heading h2 { margin:0 0 7px; font-size:clamp(26px,3vw,38px); letter-spacing:-1.2px; font-weight:500; }
+    .presence-heading p { margin:0; color:var(--muted); line-height:1.6; }
+    .presence-view[hidden], [hidden] { display:none!important; }
+    .presence-toolbar { display:flex; flex-wrap:wrap; gap:12px; align-items:center; margin-bottom:22px; }
+    .presence-toolbar input { min-width:180px; max-width:380px; flex:1; margin:0; }
+    .presence-toolbar select { width:auto; min-width:140px; margin:0; }
+    .presence-grid { display:grid; grid-template-columns:repeat(auto-fill,minmax(265px,1fr)); gap:16px; }
+    .presence-card { background:linear-gradient(140deg,#14222c,#101a23); border:1px solid var(--line); border-radius:16px; padding:22px; display:flex; flex-direction:column; gap:14px; min-width:0; }
+    .presence-card h3 { font-size:17px; margin:0; line-height:1.4; overflow-wrap:anywhere; }
+    .presence-card p { color:var(--muted); margin:0; line-height:1.65; overflow-wrap:anywhere; }
+    .presence-card .actions { margin-top:auto; }
+    .presence-eyebrow { font-size:11px; color:var(--muted); letter-spacing:1.1px; text-transform:uppercase; }
+    .presence-card-top { display:flex; justify-content:space-between; align-items:center; gap:8px; }
+    .presence-status { font-size:12px; color:var(--muted); display:flex; align-items:center; gap:7px; }
+    .presence-status::before { content:""; width:6px; height:6px; background:currentColor; border-radius:50%; flex-shrink:0; }
+    .presence-status[data-ready="true"] { color:var(--accent); }
+    .presence-card dl { display:grid; grid-template-columns:1fr 1fr; gap:7px; margin:0; font-size:12px; }
+    .presence-card dt { color:var(--muted); }
+    .presence-card dd { margin:0; text-align:right; overflow-wrap:anywhere; }
+    .presence-memory { height:5px; background:#263640; border-radius:4px; overflow:hidden; }
+    .presence-memory > span { display:block; height:100%; background:var(--accent); }
+    .presence-section-title { font-size:18px; margin:30px 0 16px; }
+    .presence-message { color:var(--muted); padding:20px 0; line-height:1.7; max-width:800px; }
+    .presence-notice { position:fixed; z-index:60; bottom:20px; left:218px; right:28px; max-width:660px; background:#162b35; color:var(--text); border:1px solid #345767; border-radius:12px; padding:16px 48px 16px 18px; box-shadow:0 10px 35px #0008; overflow-wrap:anywhere; }
+    .presence-notice[data-error="true"] { border-color:#a36050; }
+    .presence-notice button { position:absolute; right:7px; top:7px; background:none; border:0; }
+    .presence-connection { display:grid; grid-template-columns:minmax(0,1.3fr) minmax(260px,1fr); gap:20px; }
+    .presence-connection pre { white-space:pre-wrap; overflow-wrap:anywhere; font-size:12px; background:#091118; border-radius:10px; padding:18px; max-height:360px; overflow:auto; }
+    .presence-connection label { display:block; font-size:12px; margin-top:12px; }
+    .presence-connection select { margin-top:8px; }
+    .presence-readiness { display:flex; flex-wrap:wrap; gap:6px; }
+    .presence-readiness button { flex:1; font-size:12px; padding:9px 7px; }
+    .presence-readiness button[aria-pressed="true"] { background:#20414b; border-color:var(--accent); }
+    .presence-drawer { position:fixed; z-index:45; inset:0; display:grid; place-items:center; padding:20px; background:#02080cc9; }
+    .presence-dialog { width:min(620px,100%); max-height:90vh; overflow:auto; border:1px solid var(--line); background:var(--panel); border-radius:20px; padding:28px; }
+    .presence-dialog h2 { margin-top:0; letter-spacing:-.5px; }
+    .presence-dialog label { display:block; margin:14px 0; }
+    .presence-dialog pre { max-height:230px; overflow:auto; white-space:pre-wrap; overflow-wrap:anywhere; font-size:12px; }
+    .presence-dialog details { margin:16px 0; color:var(--muted); }
+    .presence-dialog .actions { justify-content:flex-end; flex-wrap:wrap; }
+    .presence-trial { margin-top:18px; }
+    .presence-trial textarea { width:100%; padding:12px; background:var(--bg); color:var(--text); border:1px solid var(--line); resize:vertical; min-height:90px; }
+    .presence-trial pre:empty { display:none; }
+    .presence-trial pre { white-space:pre-wrap; font-size:13px; line-height:1.7; }
+    .topology { border:1px solid var(--line); border-radius:18px; min-height:600px; background:#080f15; box-shadow:none; }
+    .topology::before,.topology::after { display:none; }
+    .topology-canvas { min-height:600px; height:calc(100vh - 230px); background:transparent; }
+    .topology-hud { top:16px; left:16px; right:16px; flex-wrap:wrap; gap:8px; }
+    .topology-hud-panel { flex:1 1 220px; }
+    .topology-hud-right { flex-wrap:wrap; max-width:100%; }
+    .topology-hud-panel .mono { font-family:inherit; font-size:12px; }
+    .topology-hud-panel { background:#101b24e8; border-color:var(--line); border-radius:10px; box-shadow:none; }
+    .fabric-title { font-family:inherit; font-size:13px; letter-spacing:0; font-weight:500; }
+    .topology-model-filter,.topology-metrics,.topology-zoom { border-radius:9px; }
+    .fabric-totals { gap:14px; }
+    .fabric-total strong { font-family:inherit; font-weight:500; }
+    .model-inspector { position:fixed; z-index:40; top:88px; right:24px; bottom:24px; max-height:calc(100vh - 112px); width:min(390px,calc(100vw - 32px)); border-radius:18px; background:#101b24; border-color:#345461; box-shadow:0 24px 80px #0008; }
+    .model-inspector:not(.open) { visibility:hidden; }
+    .model-inspector-title { font-weight:500; font-size:22px; }
+    .model-detail-grid { grid-template-columns:1fr 1fr; }
+    .model-detail strong { font-weight:400; }
+    .model-inspector-actions { grid-template-columns:repeat(2,minmax(0,1fr)); }
+    .presence-availability {color:#acc8d6;font-size:13px;line-height:1.65;margin:0 0 14px}
+    .presence-advanced {margin-top:22px;border-top:1px solid var(--line);padding-top:14px}
+    .presence-advanced summary {color:#8faebd;font-size:12px;cursor:pointer}
+    .presence-advanced .model-inspector-actions {margin:16px 0}
+    #presence-connect {width:100%;margin-top:12px}
+    #presence-send {width:100%;margin-top:8px}
+    #presence-model-error:empty {display:none}
+    .operations-dock { margin:0; border-radius:16px; }
+    .operations-dock > summary { display:none; }
+    .operations-content { padding:0; }
+    .operations-content .grid.two { grid-template-columns:1fr; }
+    @media(max-width:1000px) { .presence-nav { width:155px; padding:24px 10px; } body > header, main { margin-left:155px; } main { padding:22px 18px; } .presence-notice { left:175px; } .topology-hud { flex-wrap:wrap; gap:8px; } .topology-hud-right { flex-wrap:wrap; } .topology-hud-panel > .muted { display:none; } .presence-connection { grid-template-columns:1fr; } }
+    @media(max-width:640px) { .presence-nav { position:sticky; width:100%; top:0; bottom:auto; flex-direction:row; padding:8px; gap:4px; border-right:0; border-bottom:1px solid var(--line); } .presence-brand { display:none; } .presence-nav button { flex:1; flex-direction:column; justify-content:center; padding:6px 2px; gap:4px; font-size:11px; } .presence-nav button:last-child { margin-top:0; } body > header, main { margin-left:0; } body > header { padding:12px 16px; min-height:0; } .topline { justify-content:space-between; gap:8px; } #endpoint { max-width:65%; overflow:hidden; text-overflow:ellipsis; } main { padding:20px 12px; } .topology-hud { top:10px; left:10px; right:10px; } .topology-canvas { min-height:550px; height:65vh; } .topology { min-height:550px; } .topology-zoom { left:8px; right:auto; } .presence-heading { margin-bottom:20px; } .presence-grid { grid-template-columns:1fr; } .presence-notice { left:12px; right:12px; bottom:12px; } .model-inspector { top:76px; bottom:12px; right:12px; width:calc(100vw - 24px); max-height:calc(100vh - 88px); } .presence-dialog { padding:20px; } .presence-drawer { padding:12px; } }
+    @media(prefers-reduced-motion:reduce) { *,*::before,*::after { animation:none!important; transition:none!important; scroll-behavior:auto!important; } }
+`;
+
+const icons = {
+  live: '<circle cx="12" cy="12" r="3"/><circle cx="4" cy="5" r="1.5"/><circle cx="20" cy="5" r="1.5"/><circle cx="20" cy="19" r="1.5"/><path d="M5 6l5 4m4 0 5-4m-5 8 5 4"/>',
+  models: '<path d="m12 3 9 5-9 5-9-5 9-5Zm-9 9 9 5 9-5M3 16l9 5 9-5"/>',
+  machines: '<rect x="3" y="4" width="18" height="13" rx="2"/><path d="M8 21h8m-4-4v4"/>',
+  clients: '<path d="m7 5-5 7 5 7m10-14 5 7-5 7M14 4l-4 16"/>',
+  settings: '<path d="M4 7h16M4 17h16"/><circle cx="9" cy="7" r="3"/><circle cx="16" cy="17" r="3"/>'
+};
+export const presenceNav = `<nav class="presence-nav" aria-label="Main navigation"><div class="presence-brand">LLooM<small>BY ENNTITY</small></div>${Object.entries(
+  icons
+)
+  .map(
+    ([id, paths]) =>
+      `<button type="button" data-view="${id}" ${id === 'live' ? 'aria-current="page"' : ''}><svg viewBox="0 0 24 24" aria-hidden="true">${paths}</svg>${id[0].toUpperCase() + id.slice(1)}</button>`
+  )
+  .join('')}</nav>`;
+
+export const presenceViews = `
+    <section id="view-models" class="presence-view" hidden>
+      <div class="presence-heading"><div><h2>Models that fit.</h2><p>Your models, available when you need them.</p></div><button type="button" class="primary" data-add-model>Add model</button></div>
+      <div class="presence-toolbar"><input id="presence-search" type="search" aria-label="Search models" placeholder="Find a model…"><select id="presence-kind" aria-label="Model type"><option value="">All models</option><option value="chat">Chat &amp; code</option><option value="image">Images</option><option value="audio">Voice &amp; audio</option><option value="embedding">Embeddings</option><option value="video">Video</option></select><span id="presence-model-count" class="muted"></span></div>
+      <div id="presence-models" class="presence-grid"></div>
+    </section>
+    <section id="view-machines" class="presence-view" hidden>
+      <div class="presence-heading"><div><h2>Better together.</h2><p>See the hardware behind your models.</p></div></div>
+      <div id="presence-machines" class="presence-grid"></div>
+      <h3 class="presence-section-title">Nearby machines</h3><div class="presence-message">Automatic discovery is coming. Your configured peers are shown above.</div>
+      <div class="presence-message">Shared models use the gateway you already connect to. Running one larger model across machines requires a vendor recipe that supports their hardware and connection.</div>
+    </section>
+    <section id="view-clients" class="presence-view" hidden>
+      <div class="presence-heading"><div><h2>Bring your favorite client.</h2><p>One gateway for the tools you already use.</p></div></div>
+      <div class="presence-connection"><div class="presence-card"><h3>Connect to LLooM</h3><p>Use an OpenAI-compatible endpoint in your application. Models stay behind the same gateway as your hardware changes.</p><label>Base URL<input id="presence-client-url" readonly></label><label>Gateway model<select id="presence-client-model"></select></label><div class="actions"><button id="presence-copy-url" type="button">Copy URL</button><button id="presence-copy-example" type="button">Copy request</button></div><pre id="presence-client-example"></pre></div><div class="presence-card"><h3>Client integrations</h3><p>Generate configuration through the CLI. Review the proposed files before applying changes.</p><div id="presence-integrations"></div><p id="presence-client-auth"></p></div></div>
+    </section>
+    <div id="presence-toast" class="presence-notice" role="status" hidden><span></span><button type="button" aria-label="Dismiss notification">×</button></div>
+    <div id="presence-add" class="presence-drawer" hidden><section class="presence-dialog" role="dialog" aria-modal="true" aria-labelledby="presence-add-title"><h2 id="presence-add-title">Add a model.</h2><p class="muted">Choose a recommended recipe or bring a model reference. LLooM checks the installation before applying it.</p><div id="presence-recommendation"></div><form id="presence-add-form"><label>Model reference<input id="presence-model-ref" required placeholder="Paste a model link or an Ollama tag" aria-describedby="presence-model-ref-hint"></label><p id="presence-model-ref-hint" class="muted">Hugging Face links must select a specific commit. Ollama tags, local paths, and compatible endpoints are also supported.</p><details><summary>Installation details</summary><label>Backend<input id="presence-model-backend" placeholder="Choose automatically"></label><label>Display name<input id="presence-model-name" placeholder="Optional"></label></details><div id="presence-add-review" hidden><p id="presence-plan-summary"></p><details><summary>Review full plan</summary><pre id="presence-plan-json"></pre></details></div><p id="presence-add-error" role="alert"></p><div class="actions"><button id="presence-add-close" type="button">Cancel</button><button type="submit" class="primary" id="presence-add-plan">Review installation</button><button type="button" class="primary" id="presence-add-apply" hidden>Install model</button></div></form></section></div>
+`;
diff --git a/src/dashboard-scene.mjs b/src/dashboard-scene.mjs
new file mode 100644
index 0000000..b9f33d1
--- /dev/null
+++ b/src/dashboard-scene.mjs
@@ -0,0 +1,450 @@
+// The product composition: physical machines around a single gateway, with
+// motion driven by observed requests. The old diagnostic canvas stays available.
+export const sceneStyles = `
+  body { background:radial-gradient(ellipse at 48% 8%,#0a1c24 0,transparent 48%),#050b0f; }
+  body > header { min-height:38px;padding:8px 26px;background:transparent;border:0; }
+  body > header .topline {font-size:11px;opacity:.65}
+  body > header #refresh {padding:4px 10px;font-size:11px}
+  main {padding-top:8px}
+  .presence-nav {background:linear-gradient(180deg,#081218,#070d12);padding-top:24px}
+  .presence-nav button[aria-current] {color:#2be1f6;background:linear-gradient(90deg,#11313c,#102029);position:relative;border-color:#193540}
+  .presence-nav button[aria-current]::before {content:"";position:absolute;left:-9px;top:8px;bottom:8px;width:3px;border-radius:4px;background:#2be1f6;box-shadow:0 0 14px #2be1f666}
+  .presence-brand {display:flex;gap:12px;align-items:center;padding:0 8px 30px;font-weight:600;letter-spacing:-.6px}
+  .presence-brand small {font-weight:400;text-transform:none;letter-spacing:0;font-size:12px}
+  .presence-brand .scene-logo {width:30px;height:40px;filter:drop-shadow(0 0 8px #29dff344)}
+  .presence-heading h2 {font-size:36px;font-weight:600;letter-spacing:-1.1px}
+  .presence-heading {margin-bottom:22px}
+  button.primary,.scene-primary {background:linear-gradient(115deg,#29def3,#26cfee);border:1px solid #64e5f3;color:#03202a;box-shadow:0 5px 23px #0dcced16;font-weight:600}
+  button.primary:hover,.scene-primary:hover {background:#72eafb;box-shadow:0 0 22px #2ddaf32b}
+  .scene-icon {display:inline-flex;align-items:center;justify-content:center;flex-shrink:0}
+  .scene-icon svg {width:22px;height:22px;fill:none;stroke:currentColor;stroke-width:1.5;stroke-linecap:round;stroke-linejoin:round}
+  .scene-layout {display:grid;grid-template-columns:minmax(0,1fr) 310px;gap:14px;align-items:stretch}
+  .scene-diagram,.scene-detail,.scene-analytics,.scene-memory-panel,.scene-network {border:1px solid #24343f;border-radius:16px;background:linear-gradient(135deg,#0b171d88,#060d12b0);box-shadow:inset 0 1px #8ad7f304}
+  .scene-diagram {position:relative;height:clamp(440px,calc(100vh - 320px),620px);overflow:hidden}
+  .scene-columns {position:relative;z-index:1;display:grid;grid-template-columns:minmax(130px,.78fr) minmax(100px,.66fr) minmax(240px,1.42fr);gap:18px;height:100%;padding:22px 20px}
+  .scene-column {min-width:0;min-height:0;display:flex;flex-direction:column}
+  .scene-column > h3 {font-size:14px;margin:0 0 6px;font-weight:500}
+  .scene-column > p {font-size:12px;color:#9fb5c7;margin:0}
+  .scene-clients {display:flex;flex:1;flex-direction:column;justify-content:space-evenly;gap:18px;padding:38px 0}
+  .scene-client {display:flex;gap:12px;align-items:center;padding:17px 13px;border:1px solid #284553;border-radius:13px;background:linear-gradient(135deg,#162933aa,#08151bec);box-shadow:0 12px 24px #0002;min-height:72px}
+  .scene-client .scene-icon {width:34px;height:38px;border-radius:9px;background:linear-gradient(135deg,#263a47,#11212c);color:#c9ecf8}
+  .scene-client strong {display:block;font-size:13px;font-weight:500;overflow-wrap:anywhere}
+  .scene-client small {display:block;margin-top:6px;color:#9cb5c7;font-size:11px;line-height:1.4}
+  .scene-client[data-active="true"] {border-color:#2d7283}
+  .scene-gateway-column {text-align:center}
+  .scene-gateway-wrap {flex:1;display:flex;flex-direction:column;align-items:center;justify-content:center;padding-bottom:28px}
+  .scene-gateway {position:relative;width:98px;height:98px;border-radius:50%;border:2px solid #32def7;background:radial-gradient(circle at 32% 25%,#143642,#03131c 70%);display:grid;place-items:center;box-shadow:0 0 0 7px #0a2c3629,0 0 26px #14d3f22b,inset 0 0 24px #10cbea13;isolation:isolate}
+  .scene-gateway::before,.scene-gateway::after {content:"";position:absolute;inset:-12px;border-radius:50%;border:1px solid #28d4ef18;pointer-events:none}
+  .scene-gateway::after {inset:-28px;border-color:#27d9f208}
+  .scene-gateway[data-active="true"] {animation:scene-breathe 3s ease-in-out infinite;box-shadow:0 0 0 7px #0a2c3629,0 0 38px #14d3f258,inset 0 0 28px #10cbea25}
+  .scene-gateway svg {width:37px;height:48px;filter:drop-shadow(0 0 8px #21dff36b)}
+  .scene-gateway-wrap > strong {font-size:17px;margin-top:18px;font-weight:500}
+  .scene-gateway-wrap > small {font-size:12px;color:#a4c0d2;margin-top:8px;text-align:center}
+  .scene-machines {display:flex;flex-direction:column;justify-content:center;gap:12px;flex:1;min-height:0;overflow:auto;padding-top:18px;justify-content:flex-start}
+  .scene-machine {border:1px solid #29404d;border-radius:13px;background:linear-gradient(125deg,#10222d80,#0a141cdb);padding:13px;min-width:0}
+  .scene-machine-header {display:flex;gap:10px;align-items:center;margin-bottom:12px}
+  .scene-machine-header .scene-icon {color:#b9d8e9}
+  .scene-machine-header strong {display:block;font-size:13px;font-weight:500}
+  .scene-machine-header small {display:block;color:#9bb6c8;font-size:11px;margin-top:4px;line-height:1.4}
+  .scene-machine-header > div:nth-child(2) {flex:1;min-width:0;overflow-wrap:anywhere}
+  .scene-mini-memory {width:76px;text-align:right;flex-shrink:0}
+  .scene-mini-memory > i {height:5px;border-radius:5px;background:#1a303e;display:block;overflow:hidden;margin-bottom:5px}
+  .scene-mini-memory b {height:100%;display:block;border-radius:5px;background:linear-gradient(90deg,#0dc7e7,#4ce5f5);box-shadow:0 0 8px #37d7f85c}
+  .scene-mini-memory small {font-size:10px!important}
+  .scene-model {width:100%;display:flex;align-items:center;gap:10px;margin-top:7px;border:1px solid #203440;border-radius:10px;background:linear-gradient(100deg,#11212a8a,#09151ad9);padding:10px 9px;text-align:left;min-height:43px;box-shadow:inset 0 1px #b5eaff03}
+  .scene-model[data-serving="true"],.scene-model[data-selected="true"] {border-color:#27cbe6;background:linear-gradient(100deg,#07303e,#0a1720);box-shadow:inset 0 0 20px #22dffa07,0 0 15px #19cfea0a}
+  .scene-model > .scene-icon svg {width:18px;height:18px}
+  .scene-model-name {flex:1;min-width:0;white-space:nowrap;overflow:hidden;text-overflow:ellipsis;font-size:12px}
+  .scene-model-state {font-size:10px;color:#a9c2d4;background:#1a2c37;border:1px solid #223d4b;border-radius:15px;padding:4px 8px;white-space:nowrap}
+  [data-serving="true"] > .scene-model-state {color:#49e5ee;border-color:#087d91;background:#06313b}
+  .scene-more {border:0;background:none;font-size:11px;color:#9ec2d4;padding:10px 4px 0;display:block}
+  .scene-links {position:absolute;inset:0;width:100%;height:100%;pointer-events:none;z-index:0;overflow:visible}
+  .scene-detail {min-width:0;padding:22px;height:clamp(440px,calc(100vh - 320px),620px);overflow:auto}
+  .scene-placeholder {height:100%;display:flex;flex-direction:column;justify-content:center;text-align:center;align-items:center;gap:18px}
+  .scene-placeholder .scene-icon {width:72px;height:72px;border:1px solid #28414e;border-radius:50%;color:#7fdfea;background:radial-gradient(circle at 40% 30%,#203d4b,#0b1922)}
+  .scene-placeholder .scene-icon svg {width:32px;height:32px}
+  .scene-placeholder h3 {font-size:22px;margin:0}
+  .scene-placeholder p {font-size:13px;color:#99b2c5;line-height:1.7;margin:0}
+  .scene-detail .model-inspector {position:static;visibility:visible;width:100%;max-height:none;height:auto;border:0;border-radius:0;background:transparent;box-shadow:none;overflow:visible}
+  .scene-detail .model-inspector:not(.open) {display:none}
+  .scene-detail .model-inspector-header {padding:0 0 22px;border-bottom:1px solid #233642}
+  .scene-detail .model-inspector-body {padding:0;overflow:visible}
+  .scene-detail .model-inspector-title {font-size:22px;line-height:1.35}
+  .scene-detail .model-inspector-state {margin-bottom:12px}
+  .scene-detail #presence-policy {padding:20px 0;border-bottom:1px solid #233642}
+  .scene-detail #presence-policy h3 {font-size:14px;margin:0 0 14px}
+  .presence-readiness {gap:0;border:1px solid #2c4757;border-radius:10px;overflow:hidden}
+  .presence-readiness button {border-radius:0;border:0;border-right:1px solid #243c49;padding:12px 4px;background:transparent;font-size:11px;white-space:nowrap}
+  .presence-readiness button:last-child {border:0}
+  .presence-readiness button[aria-pressed="true"] {background:#25d3ee;color:#012333}
+  #presence-policy-hint {font-size:12px;line-height:1.6;margin:16px 0 0}
+  .scene-detail .model-inspector-actions {padding:20px 0;gap:7px}
+  .scene-detail details {border-top:1px solid #233642;padding-top:18px;margin-top:18px}
+  .scene-detail summary {font-size:13px;cursor:pointer}
+  .scene-inspector-memory {padding:18px 0;border-bottom:1px solid #233642;display:flex;justify-content:space-between;gap:12px;font-size:12px;color:#a9c1cf}
+  .scene-inspector-memory strong {color:#e1f2f6;font-weight:400}
+  .scene-analytics {margin-top:14px;display:grid;grid-template-columns:1fr 1fr 1fr;padding:20px 24px;gap:24px}
+  .scene-chart {min-width:0}
+  .scene-chart + .scene-chart {border-left:1px solid #20323e;padding-left:24px}
+  .scene-chart h3 {display:flex;justify-content:space-between;gap:10px;margin:0 0 16px;font-size:13px;font-weight:500}
+  .scene-chart h3 span {font-size:11px;color:#9cbed2;font-weight:400}
+  .scene-chart svg {width:100%;height:80px;overflow:visible}
+  .scene-chart-caption {font-size:10px;color:#6e899d;margin-top:6px}
+  .scene-memory-row {display:grid;grid-template-columns:minmax(70px,1fr) 1.4fr auto;align-items:center;gap:12px;font-size:11px;margin:12px 0;color:#b9cedb}
+  .scene-memory-row > span:first-child {overflow:hidden;text-overflow:ellipsis;white-space:nowrap}
+  .scene-memory-row i {height:7px;border-radius:4px;background:#182f3c;overflow:hidden}
+  .scene-memory-row b {display:block;height:100%;border-radius:4px;background:linear-gradient(90deg,#11ccea,#53e7f6)}
+  .scene-tools {display:flex;align-items:center;justify-content:space-between;gap:12px;margin:17px 0 4px;color:#7998ad;font-size:11px;flex-wrap:wrap}
+  .scene-tools .actions {align-items:center;gap:8px;margin:0}
+  .scene-tools button {padding:8px 12px;font-size:11px;background:#0b171f;border-color:#223a48}
+  .scene-follow {display:flex;gap:8px;align-items:center;font-size:11px;color:#acccdc}
+  .scene-follow input {accent-color:#22d9f3;width:auto;margin:0}
+  .scene-memory-panel {position:relative;z-index:6;background:linear-gradient(150deg,#0a1a2299,#060d12d8);padding:22px 24px 20px}
+  .scene-memory-panel h3 {display:flex;justify-content:space-between;gap:14px;font-size:15px;margin:0 0 6px;align-items:baseline;flex-wrap:wrap}
+  .scene-memory-panel h3 small {font-size:11px;color:#7d99ac;font-weight:400}
+  .scene-memory-panel select {width:auto;max-width:190px;font-size:11px;padding:5px 8px;background:#0a161e}
+  .scene-mem-sub {font-size:11.5px;color:#8ea9bb;margin:0 0 16px;line-height:1.5;max-width:66ch}
+  .scene-mem-head {display:flex;align-items:flex-end;justify-content:space-between;gap:16px;flex-wrap:wrap;margin:0 0 14px}
+  .scene-mem-total {display:flex;align-items:baseline;gap:9px;font-size:26px;font-weight:600;letter-spacing:-.6px;line-height:1}
+  .scene-mem-total span {font-size:12px;font-weight:400;color:#7f9cb0;letter-spacing:0;display:block;line-height:1.15}
+  .scene-mem-total em {font-style:normal;font-size:13px;color:#9fc4d6;font-weight:400}
+  .scene-mem-readouts {display:flex;gap:18px;flex-wrap:wrap;justify-content:flex-end}
+  .scene-mem-readout {text-align:right;min-width:86px}
+  .scene-mem-readout b {display:block;font-size:15px;font-weight:600;letter-spacing:-.2px;line-height:1.15}
+  .scene-mem-readout span {font-size:10px;letter-spacing:.7px;text-transform:uppercase;color:#6f8b9e}
+  .scene-mem-readout.used b {color:#e6f4fa}
+  .scene-mem-readout.free b {color:#5fe6d0}
+  .scene-mem-readout span.scene-mem-dot {display:inline-flex;align-items:center;gap:6px}
+  .scene-mem-readout span.scene-mem-dot::before {content:"";width:6px;height:6px;border-radius:50%;background:currentColor;opacity:.85}
+  .scene-mem-readout.free span.scene-mem-dot {color:#4fd8c4}
+  .scene-mem-readout.used span.scene-mem-dot {color:#8fb2c5}
+  .scene-mem-instrument {position:relative;border:1px solid #1f3a47;border-radius:12px;background:linear-gradient(180deg,#08131a,#060e13);overflow:hidden;isolation:isolate}
+  .scene-mem-ruler {position:relative;height:20px;border-bottom:1px solid #16303d}
+  .scene-mem-ticks {position:absolute;inset:0;display:flex;pointer-events:none}
+  .scene-mem-tick {flex:1 0 0;min-width:0;border-left:1px solid #16303d;position:relative}
+  .scene-mem-tick span {position:absolute;left:6px;top:5px;font-size:9px;color:#5d7c8e;white-space:nowrap}
+  .scene-mem-grid {position:absolute;inset:20px 0 0;pointer-events:none;display:flex;opacity:.5}
+  .scene-mem-grid i {flex:1 0 0;min-width:0;border-left:1px solid #11eaf50a}
+  .scene-mem-bar {position:relative;display:flex;height:142px;cursor:default;margin-top:2px}
+  .scene-mem-bar[data-known="false"] {height:70px}
+  .scene-mem-block {position:relative;flex:0 0 auto;min-width:0;border:0;padding:0;margin:0;background:transparent;color:#dceaf2;font:inherit;text-align:left;overflow:hidden;cursor:pointer;transition:filter .22s ease,opacity .22s ease}
+  .scene-mem-block > .scene-mem-fill {position:absolute;inset:0;background:var(--seg-fill);opacity:.9;transition:opacity .22s ease,box-shadow .22s ease}
+  .scene-mem-block > .scene-mem-rim {position:absolute;inset:0;border-right:1px solid #04121a99;background:linear-gradient(180deg,#ffffff14,#ffffff00 42%,#00000038)}
+  .scene-mem-block > .scene-mem-face {position:relative;z-index:2;display:flex;flex-direction:column;justify-content:space-between;height:100%;padding:11px 12px;gap:6px}
+  .scene-mem-block em {font-style:normal;font-size:10px;letter-spacing:.5px;color:#eaf7fc;text-shadow:0 1px 3px #04121ad9;white-space:nowrap}
+  .scene-mem-block em i {font-style:normal;opacity:.72;margin-left:5px;font-size:9px}
+  .scene-mem-block b {font-size:15px;font-weight:600;letter-spacing:-.2px;text-shadow:0 1px 3px #04121ad9;white-space:nowrap}
+  .scene-mem-block.wide > .scene-mem-face {padding:11px 14px}
+  .scene-mem-block.narrow em,.scene-mem-block.narrow b {display:none}
+  .scene-mem-block[data-kind="available"] {color:#eafffb}
+  .scene-mem-block[data-kind="available"] > .scene-mem-rim {background:linear-gradient(180deg,#ffffff1f,#ffffff00 40%,#0000001f)}
+  .scene-mem-block.system > .scene-mem-fill {background-image:repeating-linear-gradient(135deg,#ffffff10 0 7px,#ffffff00 7px 14px)}
+  .scene-mem-block[data-estimated="true"] > .scene-mem-fill {opacity:.72;background-image:repeating-linear-gradient(115deg,#ffffff12 0 6px,#ffffff00 6px 13px)}
+  .scene-mem-block:hover > .scene-mem-fill,.scene-mem-block[data-preview="true"] > .scene-mem-fill {opacity:1;box-shadow:inset 0 0 30px #ffffff1f}
+  .scene-mem-block[data-selected="true"] > .scene-mem-rim {box-shadow:inset 0 0 0 1px #ffffff4d}
+  .scene-mem-block:focus-visible {outline:2px solid #6ff0ff;outline-offset:-2px}
+  .scene-mem-ghost {position:absolute;top:0;bottom:0;left:0;width:0;pointer-events:none;transition:width .34s cubic-bezier(.22,1,.36,1),left .34s cubic-bezier(.22,1,.36,1)}
+  .scene-mem-ghost > .scene-mem-ghost-fill {position:absolute;inset:0;border:1px dashed #7ce8ffd9;border-left:0;border-radius:0 8px 8px 0;background:repeating-linear-gradient(115deg,#7ce8ff2e 0 6px,#7ce8ff0d 6px 12px);box-shadow:0 0 22px #23dcf61f,inset 0 0 24px #23dcf614}
+  .scene-mem-ghost[data-overflow="true"] > .scene-mem-ghost-fill {border-color:#ffb487ee;background:repeating-linear-gradient(115deg,#ff9d6a3a 0 6px,#ff9d6a12 6px 12px);box-shadow:0 0 22px #ff9d6a26}
+  .scene-mem-ghost[data-mode="resident"] > .scene-mem-ghost-fill,.scene-mem-ghost[data-mode="external"] > .scene-mem-ghost-fill {border-style:solid;border-color:#7ce8ff77;background:#7ce8ff14}
+  .scene-mem-ghost-label {position:absolute;right:8px;bottom:8px;font-size:10px;color:#bdf1ff;background:#062028e0;border:1px solid #2a6273;border-radius:7px;padding:4px 8px;white-space:nowrap;pointer-events:none;box-shadow:0 6px 18px #0006}
+  .scene-mem-ghost[data-overflow="true"] .scene-mem-ghost-label {color:#ffd6bd;border-color:#9b5a39;background:#2a140ce8}
+  .scene-mem-overflow {position:absolute;inset:0;z-index:3;pointer-events:none;opacity:0;transition:opacity .25s ease;background:repeating-linear-gradient(135deg,#ff9d6a1c 0 8px,#ff9d6a00 8px 16px)}
+  .scene-mem-overflow[data-on="true"] {opacity:1}
+  .scene-mem-overflow b {position:absolute;right:8px;top:8px;font-size:10px;font-weight:500;color:#ffd0b4;background:#2a120ae0;border:1px solid #9b5a39;border-radius:7px;padding:4px 8px;white-space:nowrap}
+  .scene-mem-overlay {position:absolute;top:0;bottom:0;pointer-events:none;border-left:1px dashed #48e9c7aa;background:linear-gradient(90deg,#48e9c729,#48e9c700 75%);transition:opacity .22s ease}
+  .scene-mem-overlay b {position:absolute;left:7px;top:8px;font-size:9px;letter-spacing:.5px;color:#8bf0da;white-space:nowrap;text-shadow:0 1px 3px #04121ad9}
+  .scene-mem-reserve {position:absolute;top:0;bottom:0;left:auto;right:0;pointer-events:none}
+  .scene-mem-reserve::before {content:"";position:absolute;top:0;bottom:0;left:0;width:1px;background:linear-gradient(180deg,#ffd79a00,#ffd79acc 18%,#ffd79acc 82%,#ffd79a00)}
+  .scene-mem-reserve b {position:absolute;left:0;bottom:6px;font-size:9px;letter-spacing:.5px;color:#e7bd82;white-space:nowrap;transform:translateX(-100%) translateX(-6px);text-shadow:0 1px 3px #04121ad9}
+  .scene-mem-bar[data-mode="external"] .scene-mem-overlay,.scene-mem-bar[data-mode="unknown"] .scene-mem-overlay {background:linear-gradient(90deg,#8fa8b833,#8fa8b800 75%);border-left-color:#9fb6c4aa}
+  .scene-mem-forecast {margin:13px 0 0;font-size:12px;line-height:1.55;color:#a9c3d3;display:flex;gap:9px;align-items:flex-start}
+  .scene-mem-forecast .scene-mem-forecast-icon {flex-shrink:0;width:8px;height:8px;margin-top:5px;border-radius:50%;background:#7f9aab}
+  .scene-mem-forecast[data-status="fits"] .scene-mem-forecast-icon {background:#3fe3cf;box-shadow:0 0 10px #3fe3cf7a}
+  .scene-mem-forecast[data-status="tight"] .scene-mem-forecast-icon {background:#f5c86a;box-shadow:0 0 10px #f5c86a7a}
+  .scene-mem-forecast[data-status="blocked"] .scene-mem-forecast-icon {background:#ff9d6a;box-shadow:0 0 10px #ff9d6a7a}
+  .scene-mem-forecast[data-status="resident"] .scene-mem-forecast-icon {background:#4fe0f2;box-shadow:0 0 10px #4fe0f27a}
+  .scene-mem-forecast[data-status="external"] .scene-mem-forecast-icon,.scene-mem-forecast[data-status="other-node"] .scene-mem-forecast-icon {background:#9db7c7}
+  .scene-mem-forecast strong {color:#eaf6fb;font-weight:500}
+  .scene-mem-forecast p {margin:0;flex:1}
+  .scene-mem-forecast .scene-mem-hint {color:#7d99ac;font-size:11px;margin-top:4px;display:block}
+  .scene-mem-legend {display:grid;grid-template-columns:repeat(auto-fit,minmax(196px,1fr));gap:6px;margin:15px 0 0}
+  .scene-mem-key {display:flex;align-items:center;gap:10px;width:100%;text-align:left;padding:9px 11px;border:1px solid #1c333f;border-radius:10px;background:#0a151c;color:#d3e4ee;font:inherit;font-size:12px;min-height:38px;transition:border-color .2s ease,background .2s ease,box-shadow .2s ease}
+  .scene-mem-key:hover {border-color:#2c6072;background:#0d1d26}
+  .scene-mem-key[data-selected="true"] {border-color:#3ed4ea;box-shadow:0 0 0 1px #28b9d126,0 6px 20px #0ad0f014}
+  .scene-mem-key[data-preview="true"] {border-color:#2e7f92;background:#0e222c}
+  .scene-mem-key:focus-visible {outline:2px solid #6ff0ff;outline-offset:2px}
+  .scene-mem-swatch {width:12px;height:12px;border-radius:4px;flex-shrink:0;background:var(--seg-fill);box-shadow:0 0 0 1px #ffffff1a}
+  .scene-mem-key-name {flex:1;min-width:0;overflow:hidden;text-overflow:ellipsis;white-space:nowrap}
+  .scene-mem-key-size {font-size:11px;color:#9db9c9;white-space:nowrap}
+  .scene-mem-key[data-estimated="true"] .scene-mem-key-size::after {content:" est.";color:#7d99ac}
+  .scene-mem-key em {font-style:normal;font-size:10px;color:#7d99ac;letter-spacing:.4px}
+  .scene-mem-rel {font-size:10px;color:#7d99ac;padding:4px 7px;border:1px solid #24404d;border-radius:7px;white-space:nowrap}
+  .scene-mem-key[data-preview="true"] .scene-mem-rel {color:#8bf0da;border-color:#2b6f7d}
+  .scene-mem-note {font-size:11px;color:#7d99ac;line-height:1.6;margin:13px 0 0;display:flex;gap:9px;flex-wrap:wrap;align-items:center}
+  .scene-mem-note span.scene-mem-chip {display:inline-flex;align-items:center;gap:6px;border:1px solid #1f3a47;border-radius:20px;padding:4px 10px;background:#08131a}
+  .scene-mem-note span.scene-mem-chip::before {content:"";width:8px;height:8px;border-radius:3px;background:#3d5b6b}
+  .scene-mem-note span.scene-mem-chip.measured::before {background:#2ec9e0}
+  .scene-mem-note span.scene-mem-chip.estimated::before {background:repeating-linear-gradient(115deg,#f3c268 0 3px,#f3c26800 3px 6px)}
+  .scene-mem-empty {padding:22px;font-size:12px;color:#93aebe;line-height:1.6}
+  @keyframes sceneMemBreathe {0%,100%{opacity:.7}50%{opacity:1}}
+  .scene-mem-bar[data-preview="true"] .scene-mem-ghost > .scene-mem-ghost-fill {animation:sceneMemBreathe 2.6s ease-in-out infinite}
+  .scene-mem-sticky {position:sticky;top:12px;align-self:start;margin-bottom:16px}
+  .scene-model-layout {display:grid;grid-template-columns:minmax(0,1fr) 310px;gap:22px}
+  .scene-model-list {display:flex;flex-direction:column;gap:9px}
+  .scene-model-row {display:grid;grid-template-columns:44px minmax(130px,1.4fr) minmax(75px,.7fr) auto;gap:16px;align-items:center;border:1px solid #203541;border-radius:13px;padding:19px;background:linear-gradient(120deg,#111d2490,#09141b80);cursor:pointer;text-align:left;min-height:92px}
+  .scene-model-row[data-selected="true"] {border-color:#29d3eb;background:linear-gradient(100deg,#0a2a37c0,#0c1c2690)}
+  .scene-model-row > .scene-icon {width:44px;height:48px;background:linear-gradient(135deg,#24374688,#101d2a);border:1px solid #2b414d;border-radius:11px;color:#d6e9f4}
+  .scene-model-row > .scene-icon svg {width:25px;height:25px}
+  .scene-model-row h3 {margin:0;font-size:16px;overflow-wrap:anywhere;font-weight:500;line-height:1.35}
+  .scene-model-row p {font-size:12px;color:#96aebe;margin:6px 0 0}
+  .scene-row-status {font-size:12px;color:#a6c2d1;display:flex;align-items:center;gap:7px}
+  .scene-row-status::before {content:"";width:7px;height:7px;border-radius:50%;background:#718b9c;flex-shrink:0}
+  .scene-model-row[data-ready="true"] .scene-row-status::before {background:#44dccf;box-shadow:0 0 8px #44dccf22}
+  .scene-row-policy {font-size:10px;color:#a1bbc9;border:1px solid #28414e;border-radius:8px;padding:7px 9px;white-space:nowrap}
+  .scene-model-tabs {display:flex;align-items:center;gap:25px;border-bottom:1px solid #273b45;margin:0 0 18px}
+  .scene-model-tabs strong {padding:0 0 13px;font-size:16px;border-bottom:3px solid #1ed5f0}
+  .scene-model-tabs button {background:none;border:0;color:#8aa5b7;padding:0 0 15px}
+  .scene-chips {display:flex;gap:8px;flex-wrap:wrap;margin:0 0 18px}
+  .scene-chips button {border-radius:22px;padding:8px 16px;background:#0b141a;font-size:12px;border-color:#34454e}
+  .scene-chips button[aria-pressed="true"] {border-color:#29dcee;background:#0d2c36;color:#55e9f5}
+  .scene-add-capability {display:flex;gap:15px;align-items:center;border:1px dashed #35515d;border-radius:13px;padding:22px;margin-top:18px;background:#09131977}
+  .scene-add-capability > .scene-icon {width:40px;height:40px;border:1px solid #8aa9b8;border-radius:50%}
+  .scene-add-capability strong {font-size:16px;font-weight:500;display:block}
+  .scene-add-capability p {font-size:12px;color:#98b3c4;margin:7px 0 0}
+  .scene-add-capability button {margin-left:auto;padding:11px 16px}
+  .scene-network {display:flex;align-items:center;justify-content:center;gap:0;padding:40px 30px;margin-bottom:30px;min-height:190px;overflow:auto}
+  .scene-device {min-width:110px;text-align:center}
+  .scene-device .scene-icon {display:flex;filter:drop-shadow(0 0 12px #20d3f72b);color:#a7edff;margin-bottom:12px}
+  .scene-device svg {width:88px;height:70px;stroke-width:.65}
+  .scene-device strong {font-size:13px;font-weight:500}
+  .scene-wire {height:1px;min-width:60px;flex:1;max-width:160px;background:linear-gradient(90deg,#22d3f4,#1c8fa5);box-shadow:0 0 9px #20d5ed60;position:relative;margin:0 16px 22px}
+  .scene-wire::before,.scene-wire::after {content:"";position:absolute;top:-3px;width:7px;height:7px;border-radius:50%;background:#28def4;box-shadow:0 0 12px #25d5ee}
+  .scene-wire::after {right:0}
+  .scene-wire span {position:absolute;top:-22px;left:0;right:0;text-align:center;font-size:10px;color:#72dbe9;white-space:nowrap}
+  .scene-machine-list {display:flex;flex-direction:column;gap:12px}
+  .scene-machine-list .presence-card {display:grid;grid-template-columns:54px minmax(120px,1fr) minmax(150px,1fr) auto;gap:24px;align-items:center;padding:23px}
+  .scene-machine-list .presence-card > .scene-icon {color:#aae9f8;filter:drop-shadow(0 0 9px #1cc5df40)}
+  .scene-machine-list .presence-card > .scene-icon svg {width:52px;height:45px;stroke-width:.8}
+  .scene-machine-list .presence-card h3 {font-size:17px}
+  .scene-machine-list .presence-card p {font-size:12px;margin-top:7px}
+  @keyframes scene-breathe {50%{box-shadow:0 0 0 10px #0a2c3630,0 0 46px #14d3f262,inset 0 0 28px #10cbea25}}
+  @media(min-width:1650px) {.scene-layout,.scene-model-layout {grid-template-columns:minmax(0,1fr) 360px}.scene-columns {gap:30px;padding:30px}.scene-diagram,.scene-columns,.scene-detail {min-height:610px}.scene-model-name{font-size:13px}}
+  @media(max-width:1220px) {.scene-layout,.scene-model-layout {grid-template-columns:minmax(0,1fr) 280px;gap:12px}.scene-columns {grid-template-columns:120px 90px minmax(210px,1fr);gap:8px;padding:20px 15px}.scene-gateway {width:78px;height:78px}.scene-detail {padding:18px}.scene-mini-memory{width:58px}.scene-model-row {grid-template-columns:35px minmax(0,1fr) auto;gap:12px;padding:15px}.scene-row-policy{display:none}.scene-model-row > .scene-icon{width:35px;height:40px}.scene-model-row h3{font-size:14px}.scene-model-state{font-size:9px;padding:4px 6px}.scene-model-name{font-size:11px}}
+  @media(max-width:1050px) {.scene-layout,.scene-model-layout {grid-template-columns:1fr}.scene-detail {min-height:0}.scene-detail:has(.scene-placeholder:not([hidden])){display:none}.scene-columns{grid-template-columns:minmax(125px,.8fr) minmax(100px,.8fr) minmax(230px,1.4fr);gap:20px}.scene-analytics{padding:20px;gap:18px}.scene-chart + .scene-chart{padding-left:18px}.scene-machine-list .presence-card{grid-template-columns:44px 1fr auto;gap:18px}.scene-machine-list .presence-card .machine-memory{grid-column:2 / -1;grid-row:2}.scene-model-row {grid-template-columns:40px minmax(0,1fr) auto auto}.scene-row-policy{display:block}}
+  @media(max-width:680px) {.presence-brand{display:none}.presence-nav{padding:8px}.presence-nav button[aria-current]::before{left:12px;right:12px;top:auto;bottom:0;width:auto;height:2px}body > header{display:none}main{padding-top:22px}.presence-heading h2{font-size:29px}.scene-columns{grid-template-columns:1fr 1fr;gap:18px;padding:20px;min-height:0}.scene-gateway-column{grid-column:2;grid-row:1}.scene-client-column{grid-column:1;grid-row:1}.scene-machine-column{grid-column:1/-1}.scene-gateway-wrap{min-height:210px;padding:30px 0 0}.scene-clients{padding:25px 0;gap:12px}.scene-diagram{min-height:0}.scene-machines{padding-top:16px}.scene-analytics{grid-template-columns:1fr;gap:22px}.scene-chart + .scene-chart{border-left:0;border-top:1px solid #20323e;padding:20px 0 0}.scene-chart svg{height:74px}.scene-model-row{grid-template-columns:34px minmax(0,1fr) auto;padding:15px 12px;gap:10px}.scene-model-row .scene-row-policy{display:none}.scene-row-status{font-size:10px}.scene-memory-panel{padding:17px 15px 15px}.scene-mem-sticky{position:static}.scene-mem-bar{height:112px}.scene-mem-bar[data-known="false"]{height:60px}.scene-mem-legend{grid-template-columns:1fr}.scene-mem-readouts{gap:14px}.scene-mem-total{font-size:23px}.scene-mem-readout{min-width:72px}.scene-add-capability{padding:18px;flex-wrap:wrap}.scene-add-capability button{width:100%;margin:0}.scene-network{padding:25px 18px;justify-content:flex-start}.scene-machine-list .presence-card{grid-template-columns:40px 1fr;gap:14px;padding:18px}.scene-machine-list .presence-card .actions{grid-column:1/-1}.scene-machine-list .presence-card .machine-memory{grid-column:1/-1;grid-row:auto}.scene-links{opacity:.85}.scene-model-name{font-size:12px}.scene-model-state{font-size:10px}}
+  @media(max-width:1050px){.scene-detail:has(.model-inspector.open){position:fixed;z-index:80;left:12px;right:12px;bottom:12px;height:auto;max-height:calc(100dvh - 96px);background:linear-gradient(145deg,#122530,#09141c);box-shadow:0 -20px 70px #0009,0 0 0 1px #42616b66;padding:22px}.scene-detail:has(.model-inspector.open) .model-inspector-header{position:sticky;top:-22px;margin-top:-22px;padding-top:22px;background:#10212b;z-index:1}}
+  @media(prefers-reduced-motion:reduce){.scene-gateway{animation:none!important}.scene-links .scene-particle{display:none}.scene-model{transition:none}}
+  .scene-memory-panel {padding:18px 20px;margin-bottom:20px;background:#0b151d}
+  .scene-memory-panel h3 {margin-bottom:14px}.scene-memory-panel h3 small{display:block;color:#67dbe7;margin-top:5px}.scene-memory-panel h3 small[hidden]{display:none}
+  .scene-mem-head {margin-bottom:12px}.scene-mem-total {font-size:22px}.scene-mem-total b {font-weight:500}
+  .scene-mem-bar {height:76px}.scene-mem-block {border-radius:0;transition:width .38s cubic-bezier(.22,1,.36,1),filter .2s,opacity .2s}
+  .scene-mem-face em {max-width:100%;overflow:hidden;text-overflow:ellipsis;font-size:10px}
+  .scene-mem-legend {grid-template-columns:repeat(auto-fit,minmax(145px,1fr));gap:5px;margin-top:10px}
+  .scene-mem-key {padding:7px 9px;font-size:11px;min-height:44px;gap:7px}.scene-mem-key-size {font-size:10px}
+  .scene-mem-note {font-size:10px;margin:9px 0 0}.scene-mem-forecast {margin-top:10px;min-height:36px;font-size:12px}
+  .scene-mem-ruler {overflow:hidden}.scene-mem-tick {position:absolute;top:0;bottom:0;width:0}.scene-mem-tick:last-child span {left:auto;right:5px}
+  .scene-mem-reserve {right:0;background:repeating-linear-gradient(120deg,#edbe6c0d 0 5px,transparent 5px 10px)}
+  .scene-mem-reserve b {transform:none;left:6px;bottom:6px;font-size:8px;white-space:normal;line-height:1.2}
+  .scene-mem-ghost {z-index:3}.scene-mem-reserve {z-index:4}
+  .scene-model-row[data-preview="true"],.scene-model[data-preview="true"] {border-color:#53cada;box-shadow:inset 0 0 25px #36d9e507,0 0 20px #31b6ce0b}
+  .scene-row-footprint {color:#7dbbc8;font-size:11px;white-space:nowrap}
+  @media(min-width:1051px){.scene-memory-panel{position:sticky;top:12px;z-index:7}#scene-model-detail{align-self:start;position:sticky;top:12px;max-height:calc(100dvh - 24px);height:auto;min-height:440px}.scene-model-row{scroll-margin-top:390px}}
+  @media(max-width:680px){.scene-memory-panel{padding:16px 12px}.scene-mem-bar{height:64px}.scene-mem-legend{grid-template-columns:repeat(2,minmax(0,1fr))}.scene-mem-key{min-width:0}.scene-mem-head{gap:8px}.scene-mem-readouts{gap:10px}.scene-mem-total{font-size:20px}.scene-mem-forecast{font-size:11px}.scene-mem-reserve b{font-size:7px}.scene-mem-block.compact .scene-mem-face{display:none}.scene-row-footprint{display:block;margin-top:4px}}
+  @media(prefers-reduced-motion:reduce){.scene-mem-block,.scene-mem-ghost{transition:none!important}.scene-mem-ghost-fill{animation:none!important}}
+
+`;
+
+export const sceneScript = String.raw`
+    const scenePaths={
+      model:'<path d="m12 2 9 5-9 5-9-5 9-5Zm-9 5v10l9 5 9-5V7M12 12v10"/>',
+      chat:'<path d="M21 11.5a8.5 8.5 0 0 1-8.5 8.5H4l-2 2v-9.5A8.5 8.5 0 1 1 21 11.5Z"/><path d="M7 9h9M7 13h6"/>',
+      code:'<path d="m7 6-5 6 5 6m10-12 5 6-5 6M14 3l-4 18"/>',
+      image:'<rect x="3" y="3" width="18" height="18" rx="3"/><circle cx="8" cy="8" r="1.4"/><path d="m3 17 5-5 4 4 3-4 6 6"/>',
+      audio:'<path d="M3 10v4m4-8v12m5-16v20m5-16v12m4-8v4"/>',
+      laptop:'<rect x="4" y="3" width="16" height="14" rx="1.5"/><path d="M4 17 1 20h22l-3-3M9 20h6"/>',
+      server:'<path d="m6 3 11-2 3 2v18l-3 2-11-2V3Zm0 0 3 2 11-2M9 5v17"/><path d="M12 16v2m3-2v2"/>',
+      cloud:'<path d="M6 19a5 5 0 0 1-1-9.9A7 7 0 0 1 18.7 8 5.5 5.5 0 0 1-.2 11H6Z"/>',
+      terminal:'<path d="m4 7 5 5-5 5m8 0h8"/>',
+      plus:'<path d="M12 4v16M4 12h16"/>'
+    };
+    function sceneIcon(kind){return '<span class="scene-icon" aria-hidden="true"><svg viewBox="0 0 24 24">'+(scenePaths[kind]||scenePaths.model)+'</svg></span>';}
+    const sceneLogo='<svg class="scene-logo" viewBox="0 0 36 48" fill="none" aria-hidden="true"><path d="M6 3h8v33h17v9H6V3Z" stroke="#57e6f7" stroke-width="2"/><path d="M10 7v33h17" stroke="#1c7486" stroke-width="1"/></svg>';
+    function sceneKind(model){return ({chat:'Chat & code',embedding:'Search & retrieval',audio_speech:'Speech',audio_transcription:'Transcription',audio_generation:'Music & audio',image:'Generate & edit',video:'Video'})[model.kind] || model.kind || 'Chat & code';}
+    function sceneModelIcon(model){return (model.kind||'').startsWith('audio')?'audio':(model.kind||'').startsWith('image')?'image':model.kind==='chat'?'model':'code';}
+    function sceneMemory(node){
+      const memory=node?.telemetry?.memory || {};
+      const total=Number(memory.totalBytes), available=Number(memory.availableBytes), rawUsed=Number(memory.usedBytes);
+      const used=memory.availableBytes!=null&&Number.isFinite(available)?total-available:memory.usedBytes!=null?rawUsed:NaN;
+      return {total,used,available,known:total>0&&Number.isFinite(used),percent:Math.max(0,Math.min(100,used/total*100))};
+    }
+    function sceneNodes(){const nodes=Object.values(state.status?.cluster?.nodes || {});return nodes.length?nodes:state.topologySummary?.host?[{id:'local',local:true,name:'This machine',telemetry:state.topologySummary.host}]:[];}
+    function sceneNodeName(node){return node.local?'This machine':node.name || node.id;}
+    function sceneMiniMemory(node){const m=sceneMemory(node);return m.known?'<div class="scene-mini-memory"><i><b style="width:'+m.percent+'%"></b></i><small>'+escapeHtml(formatBytes(m.used))+' / '+escapeHtml(formatBytes(m.total))+'</small></div>':'';}
+    function sceneDeviceIcon(node){return node.local || /apple|mac/i.test(node.profile?.platformId || node.profile?.cpuBrand || '')?'laptop':'server';}
+    document.querySelector('.presence-brand').innerHTML=sceneLogo+'<span>LLooM<small>by Enntity</small></span>';
+    const scene=document.createElement('section');scene.id='presence-scene';scene.dataset.presencePanel='live';
+    scene.innerHTML='<div class="scene-layout"><div class="scene-diagram"><svg id="scene-links" class="scene-links" aria-hidden="true"></svg><div class="scene-columns"><div class="scene-column scene-client-column"><h3>Clients</h3><p id="scene-client-count">Waiting for telemetry</p><div id="scene-clients" class="scene-clients"></div></div><div class="scene-column scene-gateway-column"><h3>LLooM Gateway</h3><p>Routes and balances</p><div class="scene-gateway-wrap"><div id="scene-gateway" class="scene-gateway">'+sceneLogo+'</div><strong>Gateway</strong><small id="scene-active">Connecting…</small></div></div><div class="scene-column scene-machine-column"><h3>Models on your machines</h3><p id="scene-machine-count"></p><div id="scene-machines" class="scene-machines"></div></div></div></div><aside id="scene-live-detail" class="scene-detail"><div class="scene-placeholder">'+sceneIcon('model')+'<h3>Your AI, together.</h3><p>Choose a model to try it or connect an app.</p></div></aside></div><div class="scene-analytics"><div class="scene-chart"><h3>Requests <span id="scene-requests-label"></span></h3><svg id="scene-request-chart" aria-label="Observed active requests over this session"></svg><div class="scene-chart-caption">Observed during this session</div></div><div class="scene-chart"><h3>Response time <span id="scene-latency-label"></span></h3><svg id="scene-latency-chart" aria-label="Observed completed request durations"></svg><div class="scene-chart-caption">Recent completed requests · includes generation time</div></div><div class="scene-chart"><h3>Machine memory <span>Used / total</span></h3><div id="scene-memory-list"></div></div></div><div class="scene-tools"><span id="scene-health">Waiting for gateway telemetry</span><div class="actions"><label class="scene-follow"><input type="checkbox" id="scene-follow">Follow activity</label><button type="button" id="scene-diagnostic">Detailed topology</button></div></div>';
+    $('.topology').before(scene);
+    $('.topology').dataset.presencePanel='diagnostic';$('.topology').hidden=true;
+    const modelLeft=document.createElement('div');modelLeft.className='scene-model-main';
+    const modelLayout=document.createElement('div');modelLayout.className='scene-model-layout';$('#view-models').append(modelLayout);modelLayout.append(modelLeft);
+    const memoryPanel=document.createElement('section');memoryPanel.className='scene-memory-panel';memoryPanel.innerHTML='<h3><span>Room for your AI <small id="scene-memory-context" hidden></small></span><select id="scene-memory-machine" aria-label="Memory by machine"></select></h3><div id="scene-memory-bar"></div><div id="scene-memory-forecast" class="scene-mem-forecast" role="status"><span class="scene-mem-forecast-icon" aria-hidden="true"></span><p>Point to a model to preview its memory.</p></div>';
+    modelLeft.append(memoryPanel);
+    const modelTabs=document.createElement('div');modelTabs.className='scene-model-tabs';modelTabs.innerHTML='<strong>Installed</strong><button type="button" data-add-model>Discover</button>';
+    modelLeft.append(modelTabs,$('.presence-toolbar'));
+    const chips=document.createElement('div');chips.className='scene-chips';chips.innerHTML=[['','All'],['chat','Chat & code'],['image','Images'],['audio','Voice'],['embedding','Search'],['video','Video']].map(([id,name])=>'<button type="button" data-scene-kind="'+id+'" aria-pressed="'+String(!id)+'">'+name+'</button>').join('');
+    modelLeft.append(chips,$('#presence-models'));$('#presence-models').className='scene-model-list';$('#presence-kind').hidden=true;
+    const capability=document.createElement('div');capability.className='scene-add-capability';capability.innerHTML=sceneIcon('plus')+'<div><strong>Add a capability</strong><p>Get more done with another model.</p></div><button class="primary" type="button" data-add-model>Find a model</button>';modelLeft.append(capability);
+    const modelDetail=document.createElement('aside');modelDetail.id='scene-model-detail';modelDetail.className='scene-detail';modelDetail.innerHTML='<div class="scene-placeholder">'+sceneIcon('model')+'<h3>Ready when you are.</h3><p>Choose a model to try it or connect an app. LLooM takes care of getting it ready.</p></div>';
+    modelLayout.append(modelLeft,modelDetail);$('#view-models').append(modelLayout);
+    const network=document.createElement('div');network.id='scene-network';network.className='scene-network';$('#presence-machines').before(network);$('#presence-machines').className='scene-machine-list';
+    const inspectorMemory=document.createElement('div');inspectorMemory.className='scene-inspector-memory';inspectorMemory.id='scene-inspector-memory';$('#presence-availability').after(inspectorMemory);
+    let sceneClientKey='',sceneMachineKey='',sceneModelKey='',sceneNodeKey='',sceneLinkKey='',sceneSamples=[],sceneSampleAt=0;
+    let sceneMemoryNode=null,sceneFollowing=false;
+    let sceneMemPointer=null,sceneMemFocus=null,sceneMemShape='',sceneMemPaint=false;
+    const sceneMemColors=['#25bbd8','#53d5b6','#709eec','#a58ceb','#e1b176','#dc93bd','#79c5ce','#a0bb78'];
+    function sceneMemColor(segment){return segment.kind==='system'?'#304955':segment.kind==='available'?'#123337':sceneMemColors[Math.abs(segment.colorIndex||0)%sceneMemColors.length];}
+    function sceneMemModelNode(id){
+      const model=(state.physicalModels||[]).find(m=>m.id===id),rt=model&&presenceRuntime(model);
+      if(!model?.runtime)return model?.targets?.find(target=>target.node)?.node||null;
+      return rt?.node||rt?.placement?.node||(!rt?.remote?sceneNodes().find(n=>n.local)?.id:null);
+    }
+    function sceneMemSnapshot(id=sceneMemPointer||sceneMemFocus||state.selectedModelId,ownNode=false){
+      const nodes=sceneNodes(),follow=sceneMemPointer||sceneMemFocus;
+      const target=ownNode?sceneMemModelNode(id):follow&&sceneMemModelNode(follow);
+      const node=nodes.find(n=>n.id===(target||sceneMemoryNode))||nodes.find(n=>n.local)||nodes[0];
+      return buildMemoryMap({node,runtimes:state.status?.runtimeManager?.runtimes||{},models:state.physicalModels||[],previewModelId:id,memorySafety:state.status?.runtimeManager?.memorySafety});
+    }
+    function sceneMemFormat(value){return typeof value==='number'&&Number.isFinite(value)?formatBytes(Math.max(0,value)):'—';}
+    function sceneMemText(memory){
+      const p=memory.preview;
+      if(!p)return 'Point to a model to see where it would fit. Nothing starts until you use it.';
+      const status={fits:'Expected to fit',tight:'Little room to spare',blocked:'Needs more room',resident:'Already available',external:'Runs elsewhere','other-node':'Runs on another machine',unknown:'Footprint not yet known',paused:'Paused'}[p.status]||'Checking room';
+      const delta=p.additionalBytes>0?' · about '+sceneMemFormat(p.additionalBytes)+' more':'';
+      const remaining=p.additionalBytes>0&&p.remainingBytes!=null?' · '+(p.remainingBytes<0?sceneMemFormat(-p.remainingBytes)+' over capacity':sceneMemFormat(p.remainingBytes)+' available after'):'';
+      return p.label+' · '+status+delta+remaining;
+    }
+    function sceneMemSchedule(){if(sceneMemPaint)return;sceneMemPaint=true;requestAnimationFrame(()=>{sceneMemPaint=false;sceneMemRender();});}
+    function sceneMemRender(){
+      const host=$('#scene-memory-bar');if(!host||typeof buildMemoryMap!=='function')return;
+      const memory=sceneMemSnapshot(),segments=memory.segments||[],p=memory.preview;
+      const select=$('#scene-memory-machine'),nodes=sceneNodes(),base=sceneMemoryNode||nodes.find(n=>n.local)?.id||nodes[0]?.id; if(base)select.value=base;const context=$('#scene-memory-context');context.hidden=memory.nodeId===base;context.textContent=context.hidden?'':'Preview · '+sceneNodeName(nodes.find(n=>n.id===memory.nodeId)||{id:memory.nodeId});
+      const shape=JSON.stringify([memory.nodeId,memory.known,segments.map(s=>[s.id,s.kind,s.modelIds])]);
+      if(shape!==sceneMemShape){
+        const focused=host.contains(document.activeElement)?document.activeElement.dataset.memoryFocus:null;
+        sceneMemShape=shape;
+        host.innerHTML=memory.known?'<div class="scene-mem-head"><div class="scene-mem-total"><b id="scene-mem-total"></b><span>total memory</span></div><div class="scene-mem-readouts"><div class="scene-mem-readout used"><b id="scene-mem-used"></b><span>In use</span></div><div class="scene-mem-readout free"><b id="scene-mem-free"></b><span>Available</span></div></div></div><div class="scene-mem-instrument"><div id="scene-mem-ruler" class="scene-mem-ruler"></div><div class="scene-mem-bar" id="scene-mem-track" role="group" aria-label="Memory blocks">'+segments.map((s,i)=>{const model=s.modelIds?.[0],tag=model?'button':'div';return '<'+tag+(model?' type="button" data-presence-model="'+escapeHtml(model)+'"':'')+' class="scene-mem-block" data-memory-index="'+i+'" data-memory-focus="block-'+i+'" data-kind="'+escapeHtml(s.kind)+'"><span class="scene-mem-fill"></span><span class="scene-mem-rim"></span><span class="scene-mem-face"><em></em><b></b></span></'+tag+'>';}).join('')+'<span id="scene-mem-ghost" class="scene-mem-ghost" hidden><span class="scene-mem-ghost-fill"></span></span><span id="scene-mem-reserve" class="scene-mem-reserve"><b>Protected headroom</b></span></div></div><div id="scene-mem-legend" class="scene-mem-legend">'+segments.map((s,i)=>{const model=s.modelIds?.[0],tag=model?'button':'div';return '<'+tag+(model?' type="button" data-presence-model="'+escapeHtml(model)+'"':'')+' class="scene-mem-key" data-memory-key="'+i+'" data-memory-focus="key-'+i+'"><span class="scene-mem-swatch"></span><span class="scene-mem-key-name"></span><span class="scene-mem-key-size"></span></'+tag+'>';}).join('')+'</div><p id="scene-mem-note" class="scene-mem-note"></p>':'<div class="scene-mem-empty">Waiting for a memory reading from this machine.</div>';
+        if(focused)host.querySelector('[data-memory-focus="'+CSS.escape(focused)+'"]')?.focus({preventScroll:true});
+      }
+      if(memory.known){
+        $('#scene-mem-total').textContent=sceneMemFormat(memory.totalBytes);
+        $('#scene-mem-used').textContent=sceneMemFormat(memory.usedBytes);
+        $('#scene-mem-free').textContent=sceneMemFormat(memory.availableBytes);
+        const ruler=$('#scene-mem-ruler'),rulerKey=String(memory.totalBytes);
+        if(ruler.dataset.total!==rulerKey){ruler.dataset.total=rulerKey;ruler.innerHTML=Array.from({length:5},(_,i)=>'<span class="scene-mem-tick" style="left:'+i*25+'%"><span>'+escapeHtml(sceneMemFormat(memory.totalBytes*i/4))+'</span></span>').join('');}
+        for(const [i,s] of segments.entries()){
+          const block=host.querySelector('[data-memory-index="'+i+'"]'),key=host.querySelector('[data-memory-key="'+i+'"]');
+          const isPreview=Boolean(p&&s.modelIds?.includes(p.modelId)),selected=Boolean(state.selectedModelId&&s.modelIds?.includes(state.selectedModelId));
+          for(const el of [block,key]){el.style.setProperty('--seg-fill',sceneMemColor(s));el.dataset.estimated=String(Boolean(s.estimated));el.dataset.selected=String(selected);el.dataset.preview=String(isPreview);el.setAttribute('aria-label',s.label+', '+sceneMemFormat(s.bytes)+(s.estimated?', approximate':''));el.title=s.label+' · '+sceneMemFormat(s.bytes)+(s.estimated?' (estimate)':'');}
+          block.style.width=Math.max(0,Math.min(100,s.percent||0))+'%';block.classList.toggle('narrow',s.percent<10);block.classList.toggle('compact',s.percent<20);
+          block.querySelector('em').textContent=s.label;block.querySelector('b').textContent=sceneMemFormat(s.bytes);
+          key.querySelector('.scene-mem-key-name').textContent=s.label;key.querySelector('.scene-mem-key-size').textContent=sceneMemFormat(s.bytes);
+        }
+        const ghost=$('#scene-mem-ghost'),delta=p?.additionalBytes;
+        ghost.hidden=!(delta>0);ghost.style.left=memory.usedBytes/memory.totalBytes*100+'%';ghost.style.width=Math.max(0,Math.min(delta||0,memory.availableBytes))/memory.totalBytes*100+'%';ghost.dataset.overflow=String(p?.status==='blocked');
+        $('#scene-mem-track').dataset.preview=String(Boolean(p));
+        const reserve=$('#scene-mem-reserve');reserve.hidden=!(memory.reserveBytes>0);reserve.style.width=memory.reserveBytes/memory.totalBytes*100+'%';reserve.title=sceneMemFormat(memory.reserveBytes)+' protected for your machine';reserve.querySelector('b').textContent=sceneMemFormat(memory.reserveBytes)+' reserved';
+        $('#scene-mem-note').textContent=memory.attributionNote||'Live memory use. Hover previews are estimates.';
+      }
+      const forecast=$('#scene-memory-forecast');forecast.dataset.status=p?.status||'idle';const forecastText=memory.known?sceneMemText(memory):'Memory is not available yet. LLooM will check before preparing a model.';if(forecast.querySelector('p').textContent!==forecastText)forecast.querySelector('p').textContent=forecastText;
+      for(const el of document.querySelectorAll('.scene-model-row,.scene-model'))el.dataset.preview=String(el.dataset.presenceModel===p?.modelId);
+      if(state.selectedModelId){const own=sceneMemSnapshot(state.selectedModelId,true).preview;$('#scene-inspector-memory').innerHTML='<span>'+(own?.additionalBytes>0?'Expected extra memory':'Memory')+'</span><strong>'+escapeHtml(own?.additionalBytes>0?'About '+sceneMemFormat(own.additionalBytes):own?.status==='resident'?'Already available':own?.status==='external'?'Runs elsewhere':'Checked when needed')+'</strong>';}
+    }
+    const sceneMemTarget=target=>target?.closest?.('[data-presence-model]')?.dataset.presenceModel||null;
+    document.addEventListener('pointerover',event=>{if(event.pointerType==='touch')return;const id=sceneMemTarget(event.target);if(id&&id!==sceneMemPointer){sceneMemPointer=id;sceneMemSchedule();}});
+    document.addEventListener('pointerout',event=>{if(event.pointerType==='touch')return;if(sceneMemTarget(event.target)&&sceneMemTarget(event.relatedTarget)!==sceneMemPointer){sceneMemPointer=sceneMemTarget(event.relatedTarget);sceneMemSchedule();}});
+    document.addEventListener('focusin',event=>{const id=sceneMemTarget(event.target);if(id)sceneMemPointer=null;if(id!==sceneMemFocus){sceneMemFocus=id;sceneMemSchedule();}});
+    document.addEventListener('focusout',event=>{sceneMemFocus=sceneMemTarget(event.relatedTarget);sceneMemSchedule();});
+    function sceneMoveInspector(){
+      const slot=presenceView==='models'?modelDetail:$('#scene-live-detail');
+      if($('#model-inspector').parentElement!==slot)slot.append($('#model-inspector'));
+      document.querySelectorAll('.scene-placeholder').forEach(item=>item.hidden=Boolean(state.selectedModelId)&&item.parentElement===slot);
+      if(state.selectedModelId){const model=state.physicalModels.find(m=>m.id===state.selectedModelId),rt=model&&presenceRuntime(model);inspectorMemory.innerHTML='<span>Memory estimate</span><strong>'+(rt?.memoryGb!=null?escapeHtml(rt.memoryGb)+' GB':'Not reported')+'</strong>';}
+    }
+    const scenePreviousView=presenceSetView;
+    presenceSetView=function(name){sceneMemPointer=null;sceneMemFocus=null;scenePreviousView(name);sceneMoveInspector();renderScene();};
+    const scenePreviousInspector=renderModelInspector;
+    renderModelInspector=function(){scenePreviousInspector();sceneMoveInspector();sceneMemSchedule();};
+    renderPresenceModels=function(){
+      const search=$('#presence-search').value.trim().toLowerCase(),kind=$('#presence-kind').value;
+      const models=(state.physicalModels||[]).filter(m=>(!search||(m.name+' '+m.id).toLowerCase().includes(search))&&(!kind||(m.kind||'chat').startsWith(kind))).sort((a,b)=>Number(Boolean(b.runtime))-Number(Boolean(a.runtime))||Number(Boolean(presenceRuntime(b)?.healthy))-Number(Boolean(presenceRuntime(a)?.healthy)));
+      $('#presence-model-count').textContent=models.length+(models.length===1?' model':' models');
+      const key=JSON.stringify(models.map(m=>[m.id,m.name,sceneKind(m),presenceModelLabel(m),presencePolicy(presenceRuntime(m)),state.selectedModelId===m.id]));if(key===sceneModelKey)return;sceneModelKey=key;
+      const catalog=$('#presence-models'),focused=catalog.contains(document.activeElement)?document.activeElement.closest('[data-presence-model]')?.dataset.presenceModel:null;
+      $('#presence-models').innerHTML=models.map(model=>{const rt=presenceRuntime(model);return '<button type="button" class="scene-model-row" data-presence-model="'+escapeHtml(model.id)+'" data-selected="'+String(state.selectedModelId===model.id)+'" data-ready="'+String(presenceModelResident(model)||Boolean(rt?.activeRequests))+'">'+sceneIcon(sceneModelIcon(model))+'<div><h3>'+escapeHtml(model.name||model.id)+'</h3><p>'+escapeHtml(sceneKind(model))+'</p></div><span class="scene-row-status">'+escapeHtml(presenceModelLabel(model))+'</span><span class="scene-row-policy">'+escapeHtml(rt?presencePolicyNames[presencePolicy(rt)]:'Upstream')+'</span></button>';}).join('')||'<div class="empty">No models match. Choose another filter or add a model.</div>';
+      if(focused)catalog.querySelector('[data-presence-model="'+CSS.escape(focused)+'"]')?.focus({preventScroll:true});
+      if(sceneMemPointer&&!models.some(m=>m.id===sceneMemPointer))sceneMemPointer=null;
+      sceneMemSchedule();
+    };
+    renderPresenceMachines=function(){
+      const nodes=sceneNodes(),key=JSON.stringify(nodes.map(n=>[n.id,n.name,n.local,n.reachable,n.profile?.cpuBrand,sceneMemory(n)]));if(key===sceneNodeKey)return;sceneNodeKey=key;
+      network.innerHTML=nodes.map((node,index)=>(index?'<div class="scene-wire"><span>'+(node.reachable===false?'Unavailable':'Configured connection')+'</span></div>':'')+'<div class="scene-device">'+sceneIcon(sceneDeviceIcon(node))+'<strong>'+escapeHtml(sceneNodeName(node))+'</strong></div>').join('')||'<p class="muted">Waiting for machine telemetry</p>';
+      $('#presence-machines').innerHTML=nodes.map(node=>{const m=sceneMemory(node);return '<article class="presence-card">'+sceneIcon(sceneDeviceIcon(node))+'<div><h3>'+escapeHtml(sceneNodeName(node))+'</h3><p>'+escapeHtml(node.profile?.cpuBrand||node.id)+' · '+(node.reachable===false?'Unavailable':node.local?'This computer':'Configured peer')+'</p></div><div class="machine-memory"><p>'+(m.known?escapeHtml(formatBytes(m.total))+' memory':'Memory unavailable')+'</p><div class="presence-memory"><span style="width:'+(m.known?m.percent:0)+'%"></span></div><p>'+(m.known?escapeHtml(formatBytes(m.used))+' used / '+escapeHtml(formatBytes(m.total)):'No reading')+'</p></div><div class="actions"><button type="button" data-presence-node="'+escapeHtml(node.id)+'">Manage →</button></div></article>';}).join('');
+    };
+    function sceneChart(selector,values,color){
+      const svg=$(selector),width=300,height=78,max=Math.max(1,...values),points=values.map((value,index)=>[values.length>1?index/(values.length-1)*width:0,height-4-(value/max)*(height-10)]);
+      const path=points.map((p,i)=>(i?'L':'M')+p[0].toFixed(1)+','+p[1].toFixed(1)).join(' ');
+      svg.setAttribute('viewBox','0 0 300 80');svg.setAttribute('preserveAspectRatio','none');
+      svg.innerHTML='<defs><linearGradient id="'+selector.slice(1)+'-fill" x1="0" y1="0" x2="0" y2="1"><stop stop-color="'+color+'" stop-opacity=".18"/><stop offset="1" stop-color="'+color+'" stop-opacity="0"/></linearGradient></defs>'+[0,26,52,78].map(y=>'<path d="M0 '+y+'H300" stroke="#203946" stroke-width=".5"/>').join('')+(values.length>1?'<path d="'+path+' L300,80 L0,80Z" fill="url('+selector+'-fill)"/><path d="'+path+'" fill="none" stroke="'+color+'" stroke-width="1.6" vector-effect="non-scaling-stroke" style="filter:drop-shadow(0 0 3px '+color+'55)"/>':'');
+    }
+    function renderScene(){
+      if(!scene.isConnected)return;
+      const nodes=sceneNodes(),models=state.physicalModels||[],connections=(state.topologyConnections||[]).filter(c=>c.live),summary=state.topologySummary||{};
+      const clients=new Map();for(const connection of connections){const id=connection.caller||connection.requester||'API client';if(!clients.has(id))clients.set(id,{name:id,count:0});clients.get(id).count++;}
+      const clientRows=[...clients.values()].slice(0,4),clientKey=JSON.stringify(clientRows);
+      if(clientKey!==sceneClientKey||!$('#scene-clients').children.length){sceneClientKey=clientKey;$('#scene-clients').innerHTML=clientRows.map(client=>'<div class="scene-client" data-active="true">'+sceneIcon(/code|terminal|cli|agent/i.test(client.name)?'terminal':'chat')+'<div><strong>'+escapeHtml(client.name)+'</strong><small>'+client.count+' active request'+(client.count===1?'':'s')+'</small></div></div>').join('')||'<div class="scene-client" data-active="false">'+sceneIcon('terminal')+'<div><strong>Ready for a client</strong><small>Connect an app to see its requests here.</small></div></div>';sceneLinkKey='';}
+      $('#scene-client-count').textContent=connections.length?clients.size+' active client'+(clients.size===1?'':'s'):'No active requests';
+      $('#scene-active').textContent=(summary.active||0)?summary.active+' active request'+(summary.active===1?'':'s'):'Ready when you are';$('#scene-gateway').dataset.active=String(Boolean(summary.active));
+      $('#scene-machine-count').textContent=nodes.length+' machine'+(nodes.length===1?'':'s')+' · '+models.length+' models';
+      const groups=nodes.map(node=>({node,models:models.filter(model=>{const topology=(state.topologyCatalogModels||[]).find(m=>m.id===model.id);const rt=presenceRuntime(model),ids=[...new Set([...(topology?.nodes||[]),...(model.targets||[]).map(t=>t.node),...(rt?.members||[]).map(m=>m.node),rt?.node].filter(Boolean))];return ids.includes(node.id)||(!ids.length&&Boolean(model.runtime)&&node.local);})}));
+      const assigned=new Set(groups.flatMap(g=>g.models.map(m=>m.id)));const external=models.filter(m=>!assigned.has(m.id));if(external.length)groups.push({node:{id:'external',name:'External providers'},models:external,external:true});
+      const visibleGroups=groups.map(group=>({...group,models:group.models.slice().sort((a,b)=>Number(presenceRuntime(b)?.activeRequests>0)-Number(presenceRuntime(a)?.activeRequests>0)||Number(Boolean(presenceRuntime(b)?.healthy))-Number(Boolean(presenceRuntime(a)?.healthy)))}));
+      const machineKey=JSON.stringify(visibleGroups.map(g=>[g.node.id,g.node.reachable,g.models.map(m=>[m.id,m.name,presenceModelLabel(m),connections.some(c=>c.model===m.id),m.id===state.selectedModelId]),sceneFollowing]));
+      if(machineKey!==sceneMachineKey){sceneMachineKey=machineKey;$('#scene-machines').innerHTML=visibleGroups.filter(g=>!sceneFollowing||!summary.active||g.models.some(m=>presenceRuntime(m)?.activeRequests>0||connections.some(c=>c.model===m.id))).map(group=>{
+        const ranked=sceneFollowing&&summary.active?group.models.filter(m=>presenceRuntime(m)?.activeRequests>0||connections.some(c=>c.model===m.id)):group.models;
+        const shown=ranked.slice(0,group.external?2:3);
+        return '<div class="scene-machine" data-scene-node="'+escapeHtml(group.node.id)+'"><div class="scene-machine-header">'+sceneIcon(group.external?'cloud':sceneDeviceIcon(group.node))+'<div><strong>'+escapeHtml(sceneNodeName(group.node))+'</strong><small>'+escapeHtml(group.external?'Through your gateway':group.node.profile?.cpuBrand||group.node.id)+'</small></div>'+sceneMiniMemory(group.node)+'</div>'+shown.map(model=>{const label=presenceModelLabel(model),active=connections.some(c=>c.model===model.id),serving=label==='Serving'||(active&&!model.runtime&&!model.federated);return '<button type="button" class="scene-model" data-presence-model="'+escapeHtml(model.id)+'" data-serving="'+String(serving)+'" data-active="'+String(active)+'" data-selected="'+String(model.id===state.selectedModelId)+'">'+sceneIcon(sceneModelIcon(model))+'<span class="scene-model-name" title="'+escapeHtml(model.name||model.id)+'">'+escapeHtml(model.name||model.id)+'</span><span class="scene-model-state">'+escapeHtml(serving?'Serving':label==='External provider'?'External':label)+'</span></button>';}).join('')+(group.models.length>shown.length?'<button type="button" class="scene-more" data-scene-all>View all '+group.models.length+' models →</button>':'')+'</div>';
+      }).join('')||'<div class="empty">Add a model to bring your hardware to life.</div>';sceneLinkKey='';}
+      for(const group of visibleGroups){const box=[...document.querySelectorAll('[data-scene-node]')].find(el=>el.dataset.sceneNode===group.node.id),mini=box?.querySelector('.scene-mini-memory');if(mini){const m=sceneMemory(group.node);mini.querySelector('b').style.width=m.percent+'%';mini.querySelector('small').textContent=formatBytes(m.used)+' / '+formatBytes(m.total);}}
+      const minute=state.metrics?.rolling?.minute||{};$('#scene-requests-label').textContent=(summary.active||0)+' active'+(minute.requests!=null?' · '+minute.requests+'/min':'');
+      const now=Date.now();if(now-sceneSampleAt>1900){sceneSamples.push(Number(summary.active||0));if(sceneSamples.length>60)sceneSamples.shift();sceneSampleAt=now;}
+      sceneChart('#scene-request-chart',sceneSamples,'#29d9f5');
+      const durations=(state.metrics?.recent||[]).slice(0,40).reverse().map(r=>Number(r.durationMs)/1000).filter(n=>Number.isFinite(n)&&n>=0);sceneChart('#scene-latency-chart',durations,'#47e7cc');$('#scene-latency-label').textContent=durations.length?(durations.reduce((a,b)=>a+b,0)/durations.length).toFixed(1)+'s avg':'No completed requests';
+      $('#scene-memory-list').innerHTML=nodes.slice(0,4).map(node=>{const m=sceneMemory(node);return '<div class="scene-memory-row"><span>'+escapeHtml(sceneNodeName(node))+'</span><i><b style="width:'+(m.known?m.percent:0)+'%"></b></i><span>'+(m.known?escapeHtml(formatBytes(m.used))+' / '+escapeHtml(formatBytes(m.total)):'Unknown')+'</span></div>';}).join('')||'<p class="muted">Memory telemetry unavailable</p>';
+      $('#scene-health').textContent=state.status?.error?'Gateway telemetry needs attention':nodes.length?'Live gateway telemetry · '+(summary.recentErrors||0)+' errors in the last minute':'Connecting to your gateway…';
+      const machineSelect=$('#scene-memory-machine');const optionsKey=nodes.map(n=>n.id).join('|');if(machineSelect.dataset.nodes!==optionsKey){machineSelect.dataset.nodes=optionsKey;machineSelect.innerHTML=nodes.map(n=>'<option value="'+escapeHtml(n.id)+'">'+escapeHtml(sceneNodeName(n))+'</option>').join('');if(nodes.some(n=>n.id===sceneMemoryNode))machineSelect.value=sceneMemoryNode;}
+      sceneMemRender();
+      sceneMoveInspector();requestAnimationFrame(sceneDrawLinks);
+    }
+    function sceneDrawLinks(){
+      if(scene.hidden||document.hidden)return;
+      const root=$('.scene-diagram'),bounds=root.getBoundingClientRect(),gateway=$('#scene-gateway').getBoundingClientRect(),gx=gateway.left-bounds.left+gateway.width/2,gy=gateway.top-bounds.top+gateway.height/2;
+      const curves=[];for(const client of document.querySelectorAll('.scene-client')){const box=client.getBoundingClientRect();curves.push({x:box.right-bounds.left,y:box.top-bounds.top+box.height/2,toX:gx-gateway.width/2,toY:gy,active:client.dataset.active==='true'});}
+      for(const model of document.querySelectorAll('#scene-machines .scene-model')){const box=model.getBoundingClientRect(),viewport=$('#scene-machines').getBoundingClientRect();if(box.top<viewport.top||box.bottom>viewport.bottom)continue;curves.push({x:gx+gateway.width/2,y:gy,toX:box.left-bounds.left,toY:box.top-bounds.top+box.height/2,active:model.dataset.active==='true'});}
+      const key=JSON.stringify(curves.map(c=>[Math.round(c.x),Math.round(c.y),Math.round(c.toX),Math.round(c.toY),c.active]));if(key===sceneLinkKey)return;sceneLinkKey=key;
+      const reduced=window.matchMedia('(prefers-reduced-motion: reduce)').matches;
+      $('#scene-links').setAttribute('viewBox','0 0 '+bounds.width+' '+bounds.height);
+      $('#scene-links').innerHTML='<defs><filter id="scene-glow" x="-70%" y="-70%" width="240%" height="240%"><feGaussianBlur stdDeviation="3"/></filter></defs>'+curves.map((c,index)=>{const bend=(c.toX-c.x)*.5;const d='M'+c.x+','+c.y+' C'+(c.x+bend)+','+c.y+' '+(c.toX-bend)+','+c.toY+' '+c.toX+','+c.toY;return (c.active?'<path d="'+d+'" fill="none" stroke="#13d9f3" stroke-width="4" opacity=".2" filter="url(#scene-glow)"/>':'')+'<path id="scene-flow-'+index+'" d="'+d+'" fill="none" stroke="'+(c.active?'#23def5':'#294653')+'" stroke-opacity="'+(c.active?'1':'.5')+'" stroke-width="'+(c.active?'1.3':'1')+'"/>'+(c.active&&!reduced?[0,1,2].map(n=>'<circle class="scene-particle" r="'+(n===1?3:2)+'" fill="#9cf8ff" style="filter:drop-shadow(0 0 5px #37e9ff)"><animateMotion dur="3s" begin="-'+n+'s" repeatCount="indefinite" path="'+d+'"/></circle>').join(''):'')+'<circle cx="'+c.toX+'" cy="'+c.toY+'" r="'+(c.active?4:2.5)+'" fill="'+(c.active?'#45e5f3':'#34505e')+'"/>';}).join('');
+    }
+    const sceneOriginalPresence=renderPresence;renderPresence=function(){sceneOriginalPresence();renderScene();};
+    const sceneOriginalActivity=renderActivity;renderActivity=function(){sceneOriginalActivity();renderScene();};
+    $('#scene-memory-machine').addEventListener('change',()=>{sceneMemoryNode=$('#scene-memory-machine').value;renderScene();});
+    $('#scene-follow').addEventListener('change',()=>{sceneFollowing=$('#scene-follow').checked;sceneMachineKey='';renderScene();});
+    $('#scene-diagnostic').addEventListener('click',()=>{const detail=$('.topology');detail.hidden=!detail.hidden;if(!detail.hidden)detail.scrollIntoView({behavior:window.matchMedia('(prefers-reduced-motion: reduce)').matches?'auto':'smooth',block:'start'});});
+    document.addEventListener('click',event=>{const kind=event.target.closest('[data-scene-kind]');if(kind){$('#presence-kind').value=kind.dataset.sceneKind;document.querySelectorAll('[data-scene-kind]').forEach(b=>b.setAttribute('aria-pressed',String(b===kind)));renderPresenceModels();}if(event.target.closest('[data-scene-all]'))presenceSetView('models');const selected=event.target.closest('[data-presence-model]');if(selected){sceneMemoryNode=sceneMemModelNode(selected.dataset.presenceModel)||sceneMemoryNode;sceneModelKey='';sceneMachineKey='';renderPresenceModels();renderScene();}});
+    $('#scene-machines').addEventListener('scroll',()=>{sceneLinkKey='';requestAnimationFrame(sceneDrawLinks);});
+    window.addEventListener('resize',()=>{sceneLinkKey='';requestAnimationFrame(sceneDrawLinks);});
+    document.addEventListener('visibilitychange',()=>{if(!document.hidden){sceneLinkKey='';renderScene();}});
+    presenceSetView(location.hash.slice(1));
+`;
diff --git a/src/dashboard.mjs b/src/dashboard.mjs
index 7f9a34a..5f5e16d 100644
--- a/src/dashboard.mjs
+++ b/src/dashboard.mjs
@@ -1,3 +1,8 @@
+import { buildMemoryMap } from './dashboard-memory.mjs';
+import { presenceStyles, presenceNav, presenceViews } from './dashboard-presence.mjs';
+import { sceneStyles, sceneScript } from './dashboard-scene.mjs';
+import { presenceScript } from './dashboard-presence-client.mjs';
+
 const DASHBOARD_HTML = String.raw`<!doctype html>
 <html lang="en">
 <head>
@@ -458,6 +463,17 @@ const DASHBOARD_HTML = String.raw`<!doctype html>
       </div>
     </section>
 
+    <section class="band" id="fleet-band">
+      <div class="band-head">
+        <h2>Fleet Profiles</h2>
+        <span id="fleet-active" class="pill"><span class="dot"></span><span>no profile</span></span>
+      </div>
+      <div class="band-body">
+        <div id="fleet-profiles" class="grid"></div>
+        <p class="muted" style="margin-top:10px">One file describes routes and keep-warm per machine. Applying validates the whole target config first, then swaps atomically; local models load on first request.</p>
+      </div>
+    </section>
+
     <section class="band">
       <div class="band-head">
         <h2>Models</h2>
@@ -996,7 +1012,7 @@ const DASHBOARD_HTML = String.raw`<!doctype html>
           state.topologyCamera.frameKey = "";
         } else fitTopologyCameraToModels();
       }
-      if (state.selectedModelId && !state.topologyModels.some(model => model.id === state.selectedModelId)) closeModelInspector();
+      if (state.selectedModelId && !state.physicalModels.some(model => model.id === state.selectedModelId)) closeModelInspector();
       renderTopologyModelFilter();
       renderModelInspector();
     }
@@ -1004,6 +1020,7 @@ const DASHBOARD_HTML = String.raw`<!doctype html>
     function renderRuntimes() {
       const runtimes = state.status?.runtimeManager?.runtimes || {};
       const entries = Object.entries(runtimes);
+
       $("#stat-runtimes").textContent = String(entries.length);
       $("#stat-active").textContent = String(entries.reduce((sum, [, runtime]) => sum + Number(runtime.activeRequests || 0), 0));
       $("#stat-queued").textContent = String(entries.reduce((sum, [, runtime]) => sum + Number(runtime.queuedRequests || 0) + Number(runtime.admissionQueuedRequests || 0), 0));
@@ -1022,6 +1039,65 @@ const DASHBOARD_HTML = String.raw`<!doctype html>
         '</tr>'
       ).join("") : '<tr><td colspan="5"><div class="empty">No runtimes.</div></td></tr>';
     }
+    let fleetCache = null;
+    async function renderFleetProfiles() {
+      try {
+        const result = await getJson("/gateway/fleet/profiles");
+        fleetCache = result;
+      } catch {
+        return; // fleet profiles are optional; a missing profiles dir is fine
+      }
+      const active = $("#fleet-active");
+      if (active) {
+        active.querySelector("span:last-child").textContent = fleetCache.active || "no profile";
+        active.querySelector(".dot").className = "dot " + (fleetCache.active ? "ok" : "");
+      }
+      const host = $("#fleet-profiles");
+      if (!host) return;
+      const profiles = fleetCache.profiles || [];
+      const cards = profiles.map(profile => {
+        if (profile.error) {
+          return '<div class="stat"><strong>' + escapeHtml(profile.name) + '</strong><span class="muted">' + escapeHtml(profile.error) + '</span></div>';
+        }
+        const routeCount = Object.keys(profile.routes || {}).length;
+        const warmCount = Object.keys(profile.residency || {}).length;
+        const isActive = profile.active === true;
+        return '<div class="stat"><strong>' + escapeHtml(profile.name) + (isActive ? ' <span class="pill"><span class="dot ok"></span><span>active</span></span>' : '') + '</strong>' +
+          '<span class="muted">' + escapeHtml(profile.description || "") + '</span>' +
+          '<div class="muted">' + escapeHtml(String(routeCount)) + ' routes · ' + escapeHtml(String(warmCount)) + ' warm</div>' +
+          (isActive ? '' : '<button data-fleet-use="' + escapeHtml(profile.name) + '" type="button" style="margin-top:8px">Apply</button>') +
+        '</div>';
+      });
+      const current = fleetCache.active ? '' : '<button data-fleet-save type="button" style="align-self:start">Save current as profile…</button>';
+      host.innerHTML = cards.join("") + current || '<span class="muted">No profiles yet.</span>';
+    }
+
+    document.addEventListener("click", async event => {
+      const useButton = event.target.closest("[data-fleet-use]");
+      if (useButton) {
+        const name = useButton.dataset.fleetUse;
+        if (!confirm("Apply fleet profile " + name + "? Routes and keep-warm are swapped atomically.")) return;
+        try {
+          showOutput(await postJson("/gateway/fleet/profiles/" + encodeURIComponent(name) + "?apply=1", { yes: true }));
+          await refresh();
+        } catch (error) {
+          showOutput({ error: error.message });
+        }
+      }
+      if (event.target.closest("[data-fleet-save]")) {
+        const name = prompt("Profile name (letters, numbers, dots, dashes):");
+        if (!name) return;
+        const description = prompt("Description (optional):") ?? "";
+        try {
+          showOutput(await postJson("/gateway/fleet/profiles/" + encodeURIComponent(name), {
+            yes: true, description, overwrite: false
+          }));
+          await renderFleetProfiles();
+        } catch (error) {
+          showOutput({ error: error.message });
+        }
+      }
+    });
 
     function renderBackends() {
       const rows = $("#backend-rows");
@@ -1700,8 +1776,10 @@ const DASHBOARD_HTML = String.raw`<!doctype html>
       const canvas = $("#topology-canvas");
       if (!canvas || !canvas.isConnected) return;
       const viewportWidth = Math.max(1, canvas.clientWidth), viewportHeight = Math.max(1, canvas.clientHeight);
-      if (canvas.width !== Math.round(viewportWidth) || canvas.height !== Math.round(viewportHeight)) { canvas.width = Math.round(viewportWidth); canvas.height = Math.round(viewportHeight); }
+      const pixelRatio = Math.max(1, Math.min(3, Number(window.devicePixelRatio) || 1));
+      if (canvas.width !== Math.round(viewportWidth * pixelRatio) || canvas.height !== Math.round(viewportHeight * pixelRatio)) { canvas.width = Math.round(viewportWidth * pixelRatio); canvas.height = Math.round(viewportHeight * pixelRatio); }
       const ctx = canvas.getContext("2d");
+      ctx.setTransform(pixelRatio,0,0,pixelRatio,0,0);
       ctx.clearRect(0, 0, viewportWidth, viewportHeight);
       const models = state.topologyModels || [];
       const clusterNodes = state.status?.cluster?.enabled ? Object.values(state.status?.cluster?.nodes || {}) : [];
@@ -1761,7 +1839,7 @@ const DASHBOARD_HTML = String.raw`<!doctype html>
         return hashUnit(aSeed * 29) - hashUnit(bSeed * 29);
       });
       const threadField = { left: 24, right: gate.left - 24, top: 112, bottom: height - 45 };
-      ctx.font = '11px "SFMono-Regular",monospace';
+      ctx.font = '12px system-ui,sans-serif';
       ctx.textAlign = "left";
       const connectionLabels = new Map(orderedConnections.map(connection => {
         const outputRate = smoothRate("connection:" + connection.id + ":out", connection.outputRate, now);
@@ -1921,7 +1999,7 @@ const DASHBOARD_HTML = String.raw`<!doctype html>
         if (active) { ctx.shadowColor = "rgba(47,230,200,.55)"; ctx.shadowBlur = 12; }
         ctx.beginPath(); ctx.roundRect(cardLeft, cardTop, nodeCardWidth, point.cardHeight, 6); ctx.fill(); ctx.stroke();
         ctx.shadowBlur = 0;
-        ctx.textAlign = "left"; ctx.font = '700 10px "SFMono-Regular",monospace';
+        ctx.textAlign = "left"; ctx.font = '600 11px system-ui,sans-serif';
         ctx.fillStyle = node.reachable === false ? "#ff6f7d" : "#e9fffb";
         ctx.fillText(fitCanvasText(ctx, node.name || node.id, clusterEnabled ? 116 : 94), cardLeft + 10, cardTop + 19);
         ctx.textAlign = "right"; ctx.fillStyle = node.local ? "#8fb4ff" : "rgba(153,163,176,.9)";
@@ -1930,10 +2008,10 @@ const DASHBOARD_HTML = String.raw`<!doctype html>
         if (clusterEnabled) {
           const platform = node.profile?.platformId || [node.system?.platform, node.system?.arch].filter(Boolean).join("-") || "unknown architecture";
           const accelerator = node.profile?.accelerators?.[0] || node.labels?.hardware || "cpu";
-          ctx.textAlign = "left"; ctx.font = '8px "SFMono-Regular",monospace'; ctx.fillStyle = "rgba(143,180,255,.72)";
+          ctx.textAlign = "left"; ctx.font = '8px system-ui,sans-serif'; ctx.fillStyle = "rgba(143,180,255,.72)";
           ctx.fillText(fitCanvasText(ctx, platform + " · " + accelerator, nodeCardWidth - 20), cardLeft + 10, cardTop + 35);
         }
-        ctx.font = '9px "SFMono-Regular",monospace';
+        ctx.font = '9px system-ui,sans-serif';
         point.resources.forEach((row, rowIndex) => {
           const y = cardTop + 58 + rowIndex * 18;
           const barLeft = cardLeft + 39;
@@ -1992,16 +2070,16 @@ const DASHBOARD_HTML = String.raw`<!doctype html>
           ctx.beginPath(); ctx.roundRect(cardLeft - 5, cardTop - 5, cardWidth + 10, 78, 8); ctx.stroke();
           ctx.restore();
         }
-        ctx.beginPath(); ctx.roundRect(cardLeft, cardTop, cardWidth, 68, 5); ctx.fill(); ctx.stroke();
+        ctx.beginPath(); ctx.roundRect(cardLeft, cardTop, cardWidth, 68, 12); ctx.fill(); ctx.stroke();
         if (processing) {
           const scanX = cardLeft + ((now * .08) % (cardWidth + 36)) - 18;
           const scan = ctx.createLinearGradient(scanX - 16, 0, scanX + 16, 0);
           const scanRgb = externalProcessing ? "192,153,255" : "243,189,79";
           scan.addColorStop(0, "rgba(" + scanRgb + ",0)"); scan.addColorStop(.5, "rgba(" + scanRgb + ",.13)"); scan.addColorStop(1, "rgba(" + scanRgb + ",0)");
-          ctx.save(); ctx.beginPath(); ctx.roundRect(cardLeft, cardTop, cardWidth, 68, 5); ctx.clip(); ctx.fillStyle = scan; ctx.fillRect(scanX - 16, cardTop, 32, 68); ctx.restore();
+          ctx.save(); ctx.beginPath(); ctx.roundRect(cardLeft, cardTop, cardWidth, 68, 12); ctx.clip(); ctx.fillStyle = scan; ctx.fillRect(scanX - 16, cardTop, 32, 68); ctx.restore();
         }
         ctx.fillStyle = serving ? "#42d77d" : externalProcessing ? "#f3bd4f" : external ? "#c099ff" : hot ? "#2fe6c8" : warming ? "#f3bd4f" : evicting ? "#ff7e66" : unavailable ? "#ff6f7d" : "#8fb4ff"; ctx.fillRect(cardLeft, cardTop, 4, 68);
-        ctx.textAlign = "left"; ctx.font = '11px "SFMono-Regular",monospace';
+        ctx.textAlign = "left"; ctx.font = '12px system-ui,sans-serif';
         const vendor = modelFamily(point.model.id).toUpperCase();
         const vendorText = fitCanvasText(ctx, vendor, 54);
         const titleWidth = Math.max(24, cardWidth - 12 - 10 - ctx.measureText(vendorText).width - 12);
@@ -2059,8 +2137,16 @@ const DASHBOARD_HTML = String.raw`<!doctype html>
       const instantaneousOutputRate = Math.max(0, Number(summary.outputRate || 0));
       ctx.textAlign = "center";
       if (clusterEnabled) {
-        ctx.fillStyle = "#e9fffb"; ctx.font = '700 17px "SFMono-Regular",monospace'; ctx.fillText("LLooM", center.x, gate.top + 20);
-        ctx.fillStyle = "rgba(153,163,176,.9)"; ctx.font = '9px "SFMono-Regular",monospace';
+        ctx.save();
+        ctx.shadowColor = "rgba(56,223,245,.45)";
+        ctx.shadowBlur = summary.active ? 26 : 12;
+        ctx.fillStyle = "rgba(18,43,54,.92)";
+        ctx.strokeStyle = "rgba(56,223,245,.45)";
+        ctx.lineWidth = 1;
+        ctx.beginPath(); ctx.roundRect(center.x - nodeCardWidth / 2, gate.top - 6, nodeCardWidth, 80, 20); ctx.fill(); ctx.stroke();
+        ctx.restore();
+        ctx.fillStyle = "#e9fffb"; ctx.font = '700 17px system-ui,sans-serif'; ctx.fillText("LLooM", center.x, gate.top + 20);
+        ctx.fillStyle = "rgba(153,163,176,.9)"; ctx.font = '9px system-ui,sans-serif';
         ctx.fillText(fitCanvasText(ctx, (state.status?.cluster?.id || "CLUSTER") + " · ROUTING", nodeCardWidth), center.x, gate.top + 35);
         const clusterTraffic = [
           (summary.active || 0) + " ACTIVE",
@@ -2077,13 +2163,13 @@ const DASHBOARD_HTML = String.raw`<!doctype html>
         ctx.beginPath(); ctx.roundRect(gate.left, gate.top, gate.right - gate.left, gate.bottom - gate.top, 9); ctx.fill(); ctx.stroke();
         ctx.strokeStyle = "rgba(143,180,255,.35)"; ctx.lineWidth = 1;
         for (let i = 0; i < 9; i++) { const y = gate.top + 30 + i * 25; ctx.beginPath(); ctx.moveTo(gate.left + 13, y); ctx.lineTo(gate.right - 13, y); ctx.stroke(); }
-        ctx.fillStyle = "#e9fffb"; ctx.font = '700 18px "SFMono-Regular",monospace'; ctx.fillText("LLooM", center.x, gate.top + 39);
-        ctx.fillStyle = "rgba(153,163,176,.9)"; ctx.font = '10px "SFMono-Regular",monospace'; ctx.fillText("ROUTING LOOM", center.x, gate.top + 57);
+        ctx.fillStyle = "#e9fffb"; ctx.font = '700 18px system-ui,sans-serif'; ctx.fillText("LLooM", center.x, gate.top + 39);
+        ctx.fillStyle = "rgba(153,163,176,.9)"; ctx.font = '10px system-ui,sans-serif'; ctx.fillText("ROUTING LOOM", center.x, gate.top + 57);
         const statRows = [["ACTIVE", summary.active || 0], ["PROMPT", promptTokens > 0 ? (summary.promptEstimated ? "~" : "") + formatCompact(Math.round(promptTokens)) + " tok" : "—"], ["OUTPUT", formatRate(instantaneousOutputRate) + " ~t/s"], ["ERRORS/1M", summary.recentErrors || 0]];
-        ctx.font = '11px "SFMono-Regular",monospace';
+        ctx.font = '12px system-ui,sans-serif';
         statRows.forEach((row, index) => { const y = gate.top + 94 + index * 27; ctx.textAlign = "left"; ctx.fillStyle = "rgba(153,163,176,.9)"; ctx.fillText(row[0], gate.left + 18, y); ctx.textAlign = "right"; ctx.fillStyle = "rgba(242,245,247,.95)"; ctx.fillText(String(row[1]), gate.right - 18, y); });
         const resourceRows = hostResourceRows(summary.host);
-        ctx.font = '9px "SFMono-Regular",monospace';
+        ctx.font = '9px system-ui,sans-serif';
         resourceRows.forEach((row, index) => {
           const y = gate.top + 218 + index * 24, value = row[1];
           ctx.textAlign = "left"; ctx.fillStyle = "rgba(153,163,176,.9)"; ctx.fillText(row[0], gate.left + 17, y);
@@ -2097,7 +2183,7 @@ const DASHBOARD_HTML = String.raw`<!doctype html>
     }
 
     function animateTopology() {
-      if (!document.hidden) {
+      if (!document.hidden && !$(".topology").hidden) {
         const reducedMotion = window.matchMedia("(prefers-reduced-motion: reduce)").matches;
         // Reduced motion still resolves the layout, it just settles it and
         // snaps the camera instead of easing both across frames.
@@ -2287,6 +2373,7 @@ const DASHBOARD_HTML = String.raw`<!doctype html>
     }
 
     async function refreshActivity() {
+      if (document.hidden) return;
       try {
         state.metrics = await getJson("/gateway/metrics?period=" + encodeURIComponent(state.metricsPeriod));
         renderActivity();
@@ -2306,6 +2393,7 @@ const DASHBOARD_HTML = String.raw`<!doctype html>
     }
 
     async function refresh() {
+      if (document.hidden) return;
       const healthPill = $("#health");
       healthPill?.classList.add("refreshing");
       try {
@@ -2317,7 +2405,8 @@ const DASHBOARD_HTML = String.raw`<!doctype html>
         const [health, models, status, library, backends] = await Promise.all([
           getJson("/health"),
           getJson("/gateway/models").catch(error => ({ models: [], error: error.message })),
-          getJson("/gateway/status").catch(error => ({ error: error.message })),
+          getJson("/gateway/status" + (typeof presenceView === "string" && presenceView === "live" ? "?memoryUsage=1" : ""))
+            .catch(error => ({ error: error.message })),
           getJson("/gateway/library").catch(error => ({ error: error.message })),
           getJson("/gateway/backends").catch(error => ({ backends: [], error: error.message })),
         ]);
@@ -2332,6 +2421,7 @@ const DASHBOARD_HTML = String.raw`<!doctype html>
         renderRuntimes();
         renderBackends();
         renderLibrary();
+        void renderFleetProfiles();
         renderNodeInspector();
         const authHint = security.adminAuthRequired
           ? "admin auth on"
@@ -2715,5 +2805,25 @@ const DASHBOARD_HTML = String.raw`<!doctype html>
 </html>`;
 
 export function renderDashboardPage() {
-  return DASHBOARD_HTML;
+  return DASHBOARD_HTML.replace('</style>', () => presenceStyles + sceneStyles + '\n  </style>')
+    .replace('<body>', () => '<body>' + presenceNav)
+    .replace(
+      '<main>',
+      '<main><aside id="presence-memory-safety" role="alert" hidden><strong>YOLO mode</strong> · Memory safety is disabled. Loading a model can exhaust this machine’s memory.</aside><div class="presence-heading" data-presence-panel="live"><div><h2>Your AI, in motion.</h2><p>Your hardware. Your models. One gateway.</p></div><button type="button" class="primary" data-add-model>Add model</button></div>'
+    )
+    .replace('<section class="topology"', '<section data-presence-panel="live" class="topology"')
+    .replace(
+      '<details class="operations-dock">',
+      () => presenceViews + '<details class="operations-dock" data-presence-panel="settings" hidden>'
+    )
+    .replace(
+      '    refresh();\n    refreshActivity();',
+      () =>
+        'const buildMemoryMap = (' +
+        buildMemoryMap.toString() +
+        ');\n' +
+        presenceScript +
+        sceneScript +
+        '\n    refresh();\n    refreshActivity();'
+    );
 }
diff --git a/src/first-run-page.mjs b/src/first-run-page.mjs
new file mode 100644
index 0000000..dfb6e75
--- /dev/null
+++ b/src/first-run-page.mjs
@@ -0,0 +1,775 @@
+/**
+ * Self-contained first-run setup page.
+ *
+ * No CDN, no external fonts, no network UI dependencies: everything is inline
+ * so the loopback setup server works offline. All dynamic content is inserted
+ * with textContent (never innerHTML) and every stage/reason string is escaped
+ * as data before it reaches the DOM.
+ *
+ * Visual design only: the markup carries no fake evidence. Decorative graphics
+ * (chip, logo, lock, workload/check icons) are aria-hidden. Hardware name,
+ * facts, model, sizes, and readiness all come from real dynamic data.
+ */
+
+export const PRESENCE_THEME = {
+  background: '#050c10',
+  surface: '#08141a',
+  accent: '#21d8f3',
+  danger: '#ff6b6b'
+};
+
+export function firstRunPage() {
+  return `<!doctype html>
+<html lang="en">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<meta name="color-scheme" content="dark">
+<meta name="referrer" content="no-referrer">
+<title>LLooM first-run setup</title>
+<style>
+:root{
+  --bg:#050c10; --panel:#08141a; --panel-2:#0a1a22; --border:#14313d; --border-soft:#0f2731;
+  --accent:#21d8f3; --accent-soft:rgba(33,216,243,.12);
+  --text:#e7f6fb; --muted:#7f98a6; --danger:#ff6b6b; --ok:#2ea043;
+  --radius:16px;
+}
+*{box-sizing:border-box}
+html,body{margin:0;padding:0}
+body{
+  background:var(--bg); color:var(--text);
+  font:15px/1.55 ui-sans-serif,-apple-system,BlinkMacSystemFont,"Segoe UI",Roboto,Helvetica,Arial,sans-serif;
+  min-height:100vh; padding:22px 20px 72px; -webkit-font-smoothing:antialiased;
+}
+.shell{max-width:1440px;margin:0 auto}
+p{margin:0 0 10px}
+a{color:var(--accent)}
+code{background:#050a0e;padding:2px 6px;border-radius:5px;overflow-wrap:anywhere}
+
+/* ---------- top bar ---------- */
+.topbar{display:flex;align-items:center;justify-content:space-between;gap:18px;padding:4px 4px 24px}
+.brand{display:flex;align-items:center;gap:12px;min-width:0}
+.logo{width:42px;height:42px;flex:none;filter:drop-shadow(0 0 10px rgba(33,216,243,.45))}
+.logo svg{width:100%;height:100%;display:block}
+.brand-text{display:flex;flex-direction:column;line-height:1.12;min-width:0}
+.brand-text strong{font-size:17px;font-weight:600;letter-spacing:.01em}
+.brand-text small{color:var(--muted);font-size:12px;letter-spacing:.05em}
+.progress{display:flex;align-items:center;gap:10px;font-size:13px;letter-spacing:.02em}
+.pstep{color:var(--muted);display:inline-flex;align-items:center;gap:8px;white-space:nowrap}
+.pstep::before{content:'';width:7px;height:7px;border-radius:50%;background:#1d3b46;transition:background .25s,box-shadow .25s}
+.pstep[data-active="true"]{color:var(--accent)}
+.pstep[data-active="true"]::before{background:var(--accent);box-shadow:0 0 10px rgba(33,216,243,.85)}
+.psep{width:20px;height:1px;background:linear-gradient(90deg,transparent,#1d3b46,transparent)}
+.local{display:inline-flex;align-items:center;gap:8px;font-size:13px;color:var(--muted);
+  border:1px solid var(--border);border-radius:999px;padding:7px 14px;background:rgba(9,20,26,.55);white-space:nowrap}
+.local svg{width:14px;height:14px;flex:none}
+
+/* ---------- steps ---------- */
+.step{display:none}
+.step[data-active="true"]{display:block;animation:fade .35s ease}
+@keyframes fade{from{opacity:0;transform:translateY(6px)}to{opacity:1;transform:none}}
+
+/* ---------- shared panels ---------- */
+.card,.panel{background:var(--panel);border:1px solid var(--border);border-radius:var(--radius);
+  padding:22px;margin-bottom:18px;min-width:0}
+.panel-title{font-size:20px;font-weight:600;letter-spacing:-0.01em;margin:0 0 10px}
+
+/* ---------- choose layout ---------- */
+.choose-grid{display:grid;grid-template-columns:minmax(260px,300px) minmax(0,1fr);gap:30px;align-items:start}
+.choose-main{min-width:0}
+.hero{font-size:clamp(34px,4.4vw,54px);font-weight:600;letter-spacing:-0.035em;line-height:1.04;margin:2px 0 14px;max-width:16ch}
+.hero.sm{font-size:clamp(27px,3.4vw,42px);max-width:26ch}
+.hero-sub{color:var(--muted);font-size:16px;margin:0 0 26px;max-width:62ch}
+.visually-hidden{position:absolute;width:1px;height:1px;padding:0;margin:-1px;overflow:hidden;clip:rect(0 0 0 0);white-space:nowrap;border:0}
+
+/* ---------- hardware panel ---------- */
+.machine{min-height:620px;display:flex;flex-direction:column;padding:24px 22px;
+  background:linear-gradient(180deg,rgba(10,26,34,.92),rgba(7,17,22,.92))}
+.eyebrow{font-size:11px;letter-spacing:.18em;text-transform:uppercase;color:var(--muted);margin:0 0 7px}
+.machine-name{font-size:19px;font-weight:600;letter-spacing:-0.01em;margin:0 0 4px;overflow-wrap:anywhere}
+.chip-wrap{display:flex;align-items:center;justify-content:center;margin:16px 0 22px;flex:1 1 auto;min-height:200px}
+.chip{width:100%;max-width:200px;height:auto;overflow:visible}
+.machine dl{display:grid;grid-template-columns:auto 1fr;gap:9px 16px;margin:0;font-size:13.5px}
+.machine dt{color:var(--muted);white-space:nowrap}
+.machine dd{margin:0;overflow-wrap:anywhere}
+
+/* chip visualization (decorative) */
+.ring{fill:none;stroke:rgba(33,216,243,.22);stroke-width:1;stroke-dasharray:2 7}
+.ring-2{stroke:rgba(33,216,243,.15);stroke-dasharray:1 10}
+.ring-3{stroke:rgba(33,216,243,.11)}
+.pin{stroke:#15505f;stroke-width:2;stroke-linecap:round}
+.body{fill:#08222b;stroke:#21d8f3;stroke-width:1.6;filter:drop-shadow(0 0 6px rgba(33,216,243,.55))}
+.glyph{fill:none;stroke:#7fe9fb;stroke-width:1.4;opacity:.92}
+.orbit{transform-box:fill-box;transform-origin:center;animation:chipSpin 26s linear infinite}
+.orbit-rev{animation-duration:38s;animation-direction:reverse}
+.orbit-dot{fill:var(--accent);filter:drop-shadow(0 0 6px rgba(33,216,243,.9))}
+.scan{stroke:var(--accent);stroke-width:1.4;stroke-linecap:round;opacity:0;animation:chipScan 5s ease-in-out infinite}
+@keyframes chipSpin{to{transform:rotate(360deg)}}
+@keyframes chipScan{
+  0%{transform:translateY(-30px);opacity:0}
+  12%{opacity:.85}
+  50%{opacity:.7}
+  88%{opacity:.15}
+  100%{transform:translateY(30px);opacity:0}
+}
+
+/* ---------- workload choices ---------- */
+.choices{display:grid;grid-template-columns:repeat(4,minmax(0,1fr));gap:14px;margin-bottom:24px}
+.choice{position:relative;display:flex;flex-direction:column;align-items:flex-start;gap:16px;
+  min-height:170px;text-align:left;background:linear-gradient(180deg,rgba(12,26,34,.78),rgba(8,18,24,.78));
+  color:inherit;border:1px solid var(--border);border-radius:14px;padding:20px 16px;cursor:pointer;font:inherit;
+  transition:border-color .2s,transform .2s,background .2s,box-shadow .2s}
+.choice:hover{border-color:#22505f;transform:translateY(-1px)}
+.choice[aria-pressed="true"]{border-color:var(--accent);
+  background:linear-gradient(180deg,rgba(16,42,52,.92),rgba(9,22,28,.92));
+  box-shadow:0 0 0 1px rgba(33,216,243,.35),0 14px 34px -20px rgba(33,216,243,.75)}
+.wk-icon{width:52px;height:52px;border-radius:13px;border:1px solid #1b3d4a;display:flex;align-items:center;
+  justify-content:center;color:var(--accent);background:rgba(33,216,243,.05);transition:.2s}
+.wk-icon svg{width:26px;height:26px;display:block}
+.choice[aria-pressed="true"] .wk-icon{border-color:rgba(33,216,243,.6);background:rgba(33,216,243,.12);
+  box-shadow:0 0 18px -6px rgba(33,216,243,.85)}
+.reco h3{color:var(--text)}
+.wk-body{display:flex;flex-direction:column;gap:5px;min-width:0}
+.wk-body strong{font-size:17px;font-weight:600;letter-spacing:-0.01em;overflow-wrap:anywhere}
+.wk-desc{color:var(--muted);font-size:13.5px;display:block;overflow-wrap:anywhere}
+.wk-check{position:absolute;right:12px;top:12px;width:22px;height:22px;border-radius:50%;border:1px solid #21404c;display:flex;align-items:center;
+  justify-content:center;justify-self:end;color:transparent;transition:.2s}
+.wk-check::after{content:'';width:11px;height:6px;border-left:1.8px solid currentColor;border-bottom:1.8px solid currentColor;
+  transform:rotate(-45deg) translate(.5px,-1px)}
+.choice[aria-pressed="true"] .wk-check{border-color:var(--accent);color:#04141a;background:var(--accent)}
+.choice:focus-visible,button:focus-visible,summary:focus-visible,a:focus-visible{outline:2px solid var(--accent);outline-offset:2px}
+
+/* ---------- recommendation ---------- */
+#recommendation{min-width:0}
+.card.reco{border-color:rgba(33,216,243,.42);padding:26px 26px 22px;
+  background:linear-gradient(180deg,rgba(9,26,34,.94),rgba(6,16,21,.94));
+  box-shadow:0 26px 64px -44px rgba(33,216,243,.85)}
+.reco-icon{display:inline-flex;align-items:center;justify-content:center;width:38px;height:38px;border-radius:11px;
+  border:1px solid rgba(33,216,243,.45);background:rgba(33,216,243,.09);color:var(--accent);margin-bottom:14px}
+.reco-icon svg{width:21px;height:21px;display:block}
+.card.reco h3{font-size:23px;font-weight:600;letter-spacing:-0.015em;text-transform:none;margin:12px 0 8px}
+.reco .mono{color:var(--muted)}
+.badge{display:inline-flex;align-items:center;font-size:11.5px;border-radius:999px;padding:4px 11px;
+  border:1px solid var(--border);color:var(--muted);margin:4px 6px 0 0;background:rgba(255,255,255,.02)}
+.badge.ok{color:var(--accent);border-color:rgba(33,216,243,.5)}
+.badge.warn{color:#ffd166;border-color:rgba(255,209,102,.35)}
+.badge.bad{color:var(--danger);border-color:rgba(255,107,107,.45)}
+.mono{font-family:ui-monospace,SFMono-Regular,Menlo,monospace;font-size:12.5px;overflow-wrap:anywhere}
+h3{font-size:14px;margin:0 0 6px;letter-spacing:.02em;color:var(--muted)}
+ul.plain{margin:0;padding-left:18px}
+ul.plain li{margin-bottom:6px}
+
+/* ---------- details / disclosures ---------- */
+details{margin-top:16px}
+summary{cursor:pointer;color:var(--accent);font-size:14px;list-style:none}
+summary::-webkit-details-marker{display:none}
+summary::before{content:'+ ';color:var(--accent);font-weight:600}
+details[open] summary::before{content:'\\2013  '}
+
+/* ---------- buttons ---------- */
+.actions{display:flex;gap:12px;flex-wrap:wrap;align-items:center;margin-top:6px}
+.actions.center{justify-content:center}
+button{font:inherit;border-radius:11px;padding:13px 22px;cursor:pointer;border:1px solid var(--border);
+  background:var(--panel-2);color:var(--text)}
+button:hover:not(:disabled){border-color:#25515f}
+button.primary{background:var(--accent);color:#031318;border-color:var(--accent);font-weight:600}
+button.cta{font-size:16.5px;padding:16px 40px;border-radius:14px;box-shadow:0 14px 36px -16px rgba(33,216,243,.9)}
+button:disabled{opacity:.45;cursor:not-allowed;box-shadow:none}
+
+/* ---------- status ---------- */
+.status{color:var(--muted);min-height:20px;margin:12px 0 0}
+.status[data-kind="error"]{color:var(--danger)}
+.status[data-kind="ok"]{color:var(--accent)}
+.center-text{text-align:center;max-width:70ch;margin-left:auto;margin-right:auto}
+
+/* ---------- stages / readiness ---------- */
+.stages{list-style:none;margin:0;padding:0}
+.stages li{display:flex;gap:14px;align-items:flex-start;padding:12px 0;border-bottom:1px solid var(--border-soft);font-size:14.5px}
+.stages li:last-child{border-bottom:0}
+.stages .dot{width:10px;height:10px;border-radius:50%;background:#24404b;margin-top:6px;flex:none}
+.stages li[data-state="active"] .dot{background:var(--accent);box-shadow:0 0 0 4px rgba(33,216,243,.15)}
+.stages li[data-state="complete"] .dot{background:var(--ok)}
+.stages li[data-state="failed"] .dot{background:var(--danger)}
+.verify{font-weight:600;margin:0 0 6px}
+.verify[data-state="verified"]{color:var(--accent)}
+.verify[data-state="installed"]{color:#ffd166}
+.verify[data-state="failed"]{color:var(--danger)}
+footer{color:var(--muted);font-size:12px;margin-top:26px;text-align:center}
+
+/* ---------- responsive ---------- */
+@media (max-width:1000px){
+  .choose-grid{grid-template-columns:minmax(210px,240px) minmax(0,1fr);gap:22px}
+  .machine{min-height:0;padding:20px}
+  .chip{max-width:168px}
+}
+@media (max-width:720px){
+  body{padding:18px 14px 56px}
+  .topbar{flex-wrap:wrap;gap:12px 16px}
+  .progress{order:3;width:100%;font-size:12.5px}
+  .choose-grid{grid-template-columns:1fr;gap:20px}
+  .machine{min-height:0;flex-direction:row;flex-wrap:wrap;align-items:center;gap:14px 18px;padding:18px}
+  .machine .eyebrow,.machine-name{width:100%;margin-bottom:0}
+  .chip-wrap{flex:0 0 auto;width:132px;margin:0;min-height:0}
+  .chip{max-width:132px}
+  .machine dl{flex:1 1 200px;gap:6px 12px}
+  .choices{grid-template-columns:repeat(2,minmax(0,1fr));gap:12px}
+  .choice{min-height:160px;padding:16px;gap:12px}
+  .wk-icon{width:44px;height:44px}
+  .wk-icon svg{width:22px;height:22px}
+  .wk-check{display:none}
+  .hero,.hero.sm{max-width:100%}
+}
+@media (max-width:520px){
+  .hero{font-size:30px}
+  .actions button{width:100%}
+  .actions.center{flex-direction:column-reverse}
+}
+@media (prefers-reduced-motion:reduce){
+  *{transition:none!important;animation:none!important}
+  .orbit,.scan{animation:none!important}
+}
+</style>
+</head>
+<body>
+<div class="shell">
+  <header class="topbar">
+    <div class="brand">
+      <span class="logo" aria-hidden="true">
+        <svg viewBox="0 0 32 32" focusable="false">
+          <defs>
+            <linearGradient id="logoGrad" x1="0" y1="0" x2="1" y2="1">
+              <stop offset="0%" stop-color="#8ff4ff"/>
+              <stop offset="60%" stop-color="#21d8f3"/>
+              <stop offset="100%" stop-color="#0f7f96"/>
+            </linearGradient>
+          </defs>
+          <rect x="3" y="3" width="26" height="26" rx="8" fill="none" stroke="url(#logoGrad)" stroke-width="1.5"/>
+          <path d="M12.4 9.4v11.4h7.6" fill="none" stroke="url(#logoGrad)" stroke-width="2.4" stroke-linecap="round" stroke-linejoin="round"/>
+        </svg>
+      </span>
+      <span class="brand-text"><strong>LLooM</strong><small>by Enntity</small></span>
+    </div>
+    <nav class="progress" aria-label="Setup progress">
+      <span class="pstep" data-stage="discover">Discover</span>
+      <span class="psep" aria-hidden="true"></span>
+      <span class="pstep" data-stage="choose">Choose</span>
+      <span class="psep" aria-hidden="true"></span>
+      <span class="pstep" data-stage="review">Review</span>
+      <span class="psep" aria-hidden="true"></span>
+      <span class="pstep" data-stage="ready">Ready</span>
+    </nav>
+    <div class="local">
+      <svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.7" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true" focusable="false">
+        <rect x="4.5" y="10.5" width="15" height="9.5" rx="2.2"/>
+        <path d="M8.2 10.5V8a3.8 3.8 0 0 1 7.6 0v2.5"/>
+      </svg>
+      Local setup
+    </div>
+  </header>
+
+  <main>
+    <section class="step" id="step-discover" data-active="false" aria-labelledby="discover-title">
+      <div class="panel">
+        <h2 class="panel-title">Discover local weight sources</h2>
+        <p id="discover-note" class="status">Looking for already-downloaded models and compatible recipes locally. This stays on this computer.</p>
+        <div class="actions">
+          <button class="primary" id="to-choose" type="button">Continue</button>
+        </div>
+      </div>
+    </section>
+
+    <section class="step" id="step-choose" data-active="true" aria-labelledby="choose-title">
+      <div class="choose-grid">
+        <div class="machine-slot" id="machine-slot">
+          <div class="card machine">
+            <p class="eyebrow">Local hardware</p>
+            <h2 id="discover-title" class="machine-name">Detecting hardware…</h2>
+            <div class="chip-wrap">
+              <svg class="chip" viewBox="0 0 200 200" aria-hidden="true" focusable="false">
+                <defs>
+                  <radialGradient id="chipCore" cx="50%" cy="50%" r="62%">
+                    <stop offset="0%" stop-color="#0d2b36"/>
+                    <stop offset="100%" stop-color="#071820"/>
+                  </radialGradient>
+                  <linearGradient id="chipEdge" x1="0" y1="0" x2="1" y2="1">
+                    <stop offset="0%" stop-color="#7cf0ff"/>
+                    <stop offset="55%" stop-color="#21d8f3"/>
+                    <stop offset="100%" stop-color="#0f7f96"/>
+                  </linearGradient>
+                </defs>
+                <g class="orbit">
+                  <circle class="ring" cx="100" cy="100" r="93"/>
+                  <circle class="orbit-dot" cx="100" cy="7" r="2.7"/>
+                </g>
+                <g class="orbit orbit-rev">
+                  <circle class="ring ring-2" cx="100" cy="100" r="82"/>
+                  <circle class="orbit-dot" cx="100" cy="182" r="2.2"/>
+                </g>
+                <circle class="ring ring-3" cx="100" cy="100" r="70"/>
+                <g class="pins">
+                  <line class="pin" x1="52" y1="76" x2="64" y2="76"/>
+                  <line class="pin" x1="52" y1="88" x2="64" y2="88"/>
+                  <line class="pin" x1="52" y1="100" x2="64" y2="100"/>
+                  <line class="pin" x1="52" y1="112" x2="64" y2="112"/>
+                  <line class="pin" x1="52" y1="124" x2="64" y2="124"/>
+                  <line class="pin" x1="136" y1="76" x2="148" y2="76"/>
+                  <line class="pin" x1="136" y1="88" x2="148" y2="88"/>
+                  <line class="pin" x1="136" y1="100" x2="148" y2="100"/>
+                  <line class="pin" x1="136" y1="112" x2="148" y2="112"/>
+                  <line class="pin" x1="136" y1="124" x2="148" y2="124"/>
+                  <line class="pin" x1="76" y1="52" x2="76" y2="64"/>
+                  <line class="pin" x1="88" y1="52" x2="88" y2="64"/>
+                  <line class="pin" x1="100" y1="52" x2="100" y2="64"/>
+                  <line class="pin" x1="112" y1="52" x2="112" y2="64"/>
+                  <line class="pin" x1="124" y1="52" x2="124" y2="64"/>
+                  <line class="pin" x1="76" y1="136" x2="76" y2="148"/>
+                  <line class="pin" x1="88" y1="136" x2="88" y2="148"/>
+                  <line class="pin" x1="100" y1="136" x2="100" y2="148"/>
+                  <line class="pin" x1="112" y1="136" x2="112" y2="148"/>
+                  <line class="pin" x1="124" y1="136" x2="124" y2="148"/>
+                </g>
+                <rect class="body" x="64" y="64" width="72" height="72" rx="14"/>
+                <rect class="glyph" x="86" y="86" width="28" height="28" rx="7"/>
+                <circle class="orbit-dot" cx="100" cy="100" r="4"/>
+                <line class="scan" x1="70" y1="100" x2="130" y2="100"/>
+              </svg>
+            </div>
+            <dl id="machine-facts"><dt>Scanning</dt><dd>Detecting hardware…</dd></dl>
+          </div>
+        </div>
+
+        <div class="choose-main">
+          <h1 class="hero">What will you create?</h1>
+          <p class="hero-sub">AI chosen for your hardware. Review the recommendation, then let LLooM set it up.</p>
+          <h2 id="choose-title" class="visually-hidden">What will you use it for?</h2>
+          <div class="choices" id="workloads" role="group" aria-label="Workload"></div>
+          <div id="recommendation" aria-live="polite">
+            <p class="status">Choose a workload to see a recommendation.</p>
+          </div>
+          <div class="actions center">
+            <button id="back-discover" type="button">Back</button>
+            <button class="primary cta" id="to-review" type="button" disabled>Review plan</button>
+          </div>
+        </div>
+      </div>
+    </section>
+
+    <section class="step" id="step-review" aria-labelledby="review-title">
+      <h1 class="hero sm" id="review-title">Review the plan</h1>
+      <p class="hero-sub">Nothing installs until you approve it. This is the last step before any writes.</p>
+      <div id="plan-review" aria-live="polite"></div>
+      <div class="actions center">
+        <button id="back-choose" type="button">Back</button>
+        <button class="primary cta" id="authorize" type="button">Set up my AI</button>
+      </div>
+      <p class="status center-text" id="authorize-note">This installs the reviewed vendor recipe, downloads its models, and configures LLooM.</p>
+    </section>
+
+    <section class="step" id="step-ready" aria-labelledby="ready-title">
+      <h1 class="hero sm" id="ready-title">Setup progress</h1>
+      <p class="hero-sub">Real stages, reported as they happen. Nothing here is estimated in advance.</p>
+      <div class="panel">
+        <ul class="stages" id="stages"></ul>
+        <p class="status" id="install-status" role="status" aria-live="polite"></p>
+      </div>
+      <div class="panel" id="result-card" hidden>
+        <h3>Result</h3>
+        <p class="verify" id="verify-state" data-state="pending"></p>
+        <p id="result-detail" class="status"></p>
+        <p id="endpoint-line" class="mono" hidden></p>
+        <div class="actions">
+          <button class="primary" id="open-lloom" type="button" disabled>Open LLooM</button>
+          <button id="retry" type="button" hidden>Retry setup</button>
+        </div>
+      </div>
+      <footer>Local setup server. It stops when this process exits.</footer>
+    </section>
+  </main>
+</div>
+<script>
+(function(){
+  'use strict';
+  var STORAGE_KEY='lloom.firstRunToken';
+  var stage={workloadId:'chat',plan:null,job:null};
+  var planSequence=0;
+  var SVGNS='http://www.w3.org/2000/svg';
+  var el=function(id){return document.getElementById(id)};
+
+  function readToken(){
+    var hash=window.location.hash||'';
+    var match=hash.match(/setup=([^&]+)/);
+    var token=match?decodeURIComponent(match[1]):null;
+    if(token){
+      try{ window.sessionStorage.setItem(STORAGE_KEY,token); }catch(e){}
+      // Strip the secret from the address bar and history immediately.
+      try{ window.history.replaceState(null,'',window.location.pathname+window.location.search); }catch(e){}
+      return token;
+    }
+    try{ return window.sessionStorage.getItem(STORAGE_KEY); }catch(e){ return null; }
+  }
+
+  var token=readToken();
+
+  function api(path,options){
+    var opts=options||{};
+    return fetch(path,{
+      method:opts.method||'GET',
+      headers:Object.assign({'authorization':'Bearer '+(token||''),'accept':'application/json'},
+        opts.body?{'content-type':'application/json'}:{}),
+      body:opts.body?JSON.stringify(opts.body):undefined,
+      credentials:'omit',
+      cache:'no-store'
+    }).then(function(res){
+      return res.json().catch(function(){return {};}).then(function(data){
+        if(!res.ok||data.ok===false){throw new Error(data.error||('Request failed ('+res.status+')'));}
+        return data;
+      });
+    });
+  }
+
+  function text(node,value){node.textContent=value==null?'':String(value);return node}
+  function clear(node){while(node.firstChild)node.removeChild(node.firstChild);return node}
+  function humanBytes(bytes){
+    var n=Number(bytes);
+    if(!isFinite(n)||n<=0)return null;
+    var units=['B','KB','MB','GB','TB'],i=0;
+    while(n>=1024&&i<units.length-1){n=n/1024;i+=1}
+    return (Math.round(n*10)/10)+' '+units[i];
+  }
+  function svgEl(name,attrs){
+    var node=document.createElementNS(SVGNS,name);
+    if(attrs){for(var k in attrs){if(Object.prototype.hasOwnProperty.call(attrs,k))node.setAttribute(k,attrs[k]);}}
+    return node;
+  }
+  function svgIcon(kind){
+    var svg=svgEl('svg',{viewBox:'0 0 24 24',fill:'none',stroke:'currentColor',
+      'stroke-width':'1.6','stroke-linecap':'round','stroke-linejoin':'round',focusable:'false'});
+    if(kind==='chat'){
+      svg.appendChild(svgEl('path',{d:'M4.5 5.5h15v9.4h-8L7.6 18.7V14.9H4.5z'}));
+      svg.appendChild(svgEl('path',{d:'M8.2 9.1h7.6'}));
+    }else if(kind==='code'){
+      svg.appendChild(svgEl('path',{d:'M9.2 8.2 5.6 12l3.6 3.8'}));
+      svg.appendChild(svgEl('path',{d:'M14.8 8.2 18.4 12l-3.6 3.8'}));
+      svg.appendChild(svgEl('path',{d:'M13.2 6.6 10.8 17.4'}));
+    }else if(kind==='images'){
+      svg.appendChild(svgEl('rect',{x:'4',y:'5',width:'16',height:'14',rx:'2.2'}));
+      svg.appendChild(svgEl('circle',{cx:'9',cy:'9.5',r:'1.5'}));
+      svg.appendChild(svgEl('path',{d:'M4.8 17.2 9.4 12.6l3.2 3.1 2.5-2.4 4 3.9'}));
+    }else{
+      svg.appendChild(svgEl('path',{d:'M5 10.5v3'}));
+      svg.appendChild(svgEl('path',{d:'M8.4 7.6v8.8'}));
+      svg.appendChild(svgEl('path',{d:'M11.8 5.6v12.8'}));
+      svg.appendChild(svgEl('path',{d:'M15.2 7.6v8.8'}));
+      svg.appendChild(svgEl('path',{d:'M18.6 10.5v3'}));
+    }
+    return svg;
+  }
+  function sparkIcon(){
+    var svg=svgEl('svg',{viewBox:'0 0 24 24',fill:'none',stroke:'currentColor',
+      'stroke-width':'1.5','stroke-linejoin':'round',focusable:'false'});
+    svg.appendChild(svgEl('path',{d:'M12 3.4 13.6 9 19.4 10.5 13.6 12 12 17.6 10.4 12 4.6 10.5 10.4 9z'}));
+    return svg;
+  }
+  function updateProgress(stepId){
+    var current=String(stepId||'').replace('step-','');
+    if(current==='install')current='review';
+    var items=document.querySelectorAll('.progress .pstep');
+    for(var i=0;i<items.length;i+=1){
+      items[i].setAttribute('data-active',items[i].getAttribute('data-stage')===current?'true':'false');
+    }
+  }
+  function show(id){
+    var steps=document.querySelectorAll('.step');
+    for(var i=0;i<steps.length;i+=1){steps[i].setAttribute('data-active',steps[i].id===id?'true':'false');}
+    updateProgress(id);
+  }
+
+  function setStatus(node,message,kind){
+    text(node,message);
+    if(kind){node.setAttribute('data-kind',kind)}else{node.removeAttribute('data-kind')}
+  }
+
+  function renderMachine(plan){
+    var list=clear(el('machine-facts'));
+    var machine=plan.machine||{};
+    var title=el('discover-title');
+    if(title){text(title,machine.name||'This machine');}
+    var facts=[
+      ['Name',machine.name||'Unknown'],
+      ['Platform',machine.platformId||'Unknown'],
+      ['Memory',machine.totalMemoryGb?machine.totalMemoryGb+' GB':'Unknown'],
+      ['Accelerator',(machine.accelerators&&machine.accelerators.length)?machine.accelerators.join(', '):'None detected']
+    ];
+    facts.forEach(function(entry){
+      var dt=document.createElement('dt');text(dt,entry[0]);
+      var dd=document.createElement('dd');text(dd,entry[1]);
+      list.appendChild(dt);list.appendChild(dd);
+    });
+  }
+
+  function optionCard(option,selected){
+    var card=document.createElement('div');
+    card.className='card'+(selected?' reco':'');
+    if(selected){
+      var icon=document.createElement('span');
+      icon.className='reco-icon';icon.setAttribute('aria-hidden','true');
+      icon.appendChild(sparkIcon());
+      card.appendChild(icon);
+    }
+    var title=document.createElement('h3');text(title,option.name||option.id);card.appendChild(title);
+    if(option.id){var id=document.createElement('p');id.className='mono';text(id,option.id);card.appendChild(id)}
+    if(option.reason){var reason=document.createElement('p');text(reason,option.reason);card.appendChild(reason)}
+    var badges=document.createElement('p');
+    var footprint=humanBytes(option.downloadSizeBytes);
+    var mem=option.memoryRequiredGb?option.memoryRequiredGb+' GB memory':'Memory requirement unknown';
+    [
+      {label:footprint?('Download '+footprint):'Download size unknown',cls:footprint?'ok':'warn'},
+      {label:mem,cls:option.memoryRequiredGb?'':'warn'},
+      {label:option.credentials?(option.credentials.required?'Credentials: '+option.credentials.required:'Credentials: none declared'):'Credentials: check vendor requirements',cls:option.credentials&&option.credentials.required?'warn':''},
+      {label:option.license?('Recipe license: '+option.license):'Licenses: see vendor recipe',cls:''}
+    ].forEach(function(badge){
+      var span=document.createElement('span');
+      span.className='badge'+(badge.cls?' '+badge.cls:'');
+      text(span,badge.label);
+      badges.appendChild(span);
+    });
+    card.appendChild(badges);
+    return card;
+  }
+
+  function renderRecommendation(plan){
+    var host=clear(el('recommendation'));
+    var selected=plan.selected;
+    if(!selected){setStatus(host,'No compatible recipe was found for this machine.','error');return}
+    host.appendChild(optionCard(selected,true));
+    if(plan.alternatives&&plan.alternatives.length){
+      var details=document.createElement('details');
+      var summary=document.createElement('summary');text(summary,'Other compatible options ('+plan.alternatives.length+')');
+      details.appendChild(summary);
+      plan.alternatives.slice(0,6).forEach(function(option){
+        var line=document.createElement('p');
+        text(line,option.name+' — '+option.id+(option.reason?' — '+option.reason:''));
+        var choose=document.createElement('button');choose.type='button';text(choose,'Choose this recipe');
+        choose.addEventListener('click',function(){loadPlan(stage.workloadId,option.id);});
+        line.appendChild(choose);
+        details.appendChild(line);
+      });
+      host.appendChild(details);
+    }
+    el('to-review').disabled=false;
+  }
+
+  function renderReview(plan){
+    var host=clear(el('plan-review'));
+    host.appendChild(optionCard(plan.selected,true));
+    var adv=document.createElement('details');
+    var summary=document.createElement('summary');text(summary,'Advanced details: packages, config, and paths');
+    adv.appendChild(summary);
+    var stages=(plan.review&&plan.review.stages)||[];
+    if(stages.length){
+      var title=document.createElement('h3');text(title,'Planned stages');adv.appendChild(title);
+      var list=document.createElement('ul');list.className='plain';
+      stages.forEach(function(stage){
+        var item=document.createElement('li');
+        text(item,(stage.title||stage.id)+(stage.summary?' — '+stage.summary:''));
+        list.appendChild(item);
+      });
+      adv.appendChild(list);
+    }
+    var paths=plan.paths||{};
+    [['Config path',paths.configPath],['Model root',paths.modelRoot]].forEach(function(entry){
+      if(!entry[1])return;
+      var p=document.createElement('p');
+      text(p,entry[0]+': ');
+      var code=document.createElement('code');text(code,entry[1]);
+      p.appendChild(code);adv.appendChild(p);
+    });
+    if(plan.ports){
+      var ports=document.createElement('p');
+      text(ports,'Ports: gateway '+plan.ports.gateway+' / backend range '+plan.ports.backendStart+'-'+plan.ports.backendEnd);
+      adv.appendChild(ports);
+    }
+    if(plan.review&&plan.review.doctorCommands){
+      var cmd=document.createElement('p');
+      text(cmd,'Verify later: ');
+      var code2=document.createElement('code');text(code2,plan.review.doctorCommands);
+      cmd.appendChild(code2);adv.appendChild(cmd);
+    }
+    host.appendChild(adv);
+    var note=document.createElement('p');
+    note.className='status';
+    text(note,'Installs and downloads are authorized only by the button below. Nothing has run yet.');
+    host.appendChild(note);
+  }
+
+  function renderStages(stages){
+    var list=clear(el('stages'));
+    (stages||[]).forEach(function(stage){
+      var item=document.createElement('li');
+      item.setAttribute('data-state',stage.status||'pending');
+      var dot=document.createElement('span');dot.className='dot';
+      var label=document.createElement('span');
+      text(label,(stage.title||stage.id||'Stage')+(stage.status?' ('+stage.status+')':''));
+      item.appendChild(dot);item.appendChild(label);
+      list.appendChild(item);
+    });
+  }
+
+  function applyResult(job){
+    el('result-card').hidden=false;
+    var verify=el('verify-state');
+    var detail=el('result-detail');
+    var endpoint=el('endpoint-line');
+    var open=el('open-lloom');
+    var retry=el('retry');
+    endpoint.hidden=true;
+    open.disabled=true;
+    retry.hidden=true;
+    if(job.status==='failed'){
+      verify.setAttribute('data-state','failed');
+      text(verify,'Setup failed');
+      setStatus(detail,job.error||'Setup did not complete. Retry, or run lloom doctor for details.','error');
+      retry.hidden=false;
+      return;
+    }
+    if(job.inferenceVerified){
+      updateProgress('step-ready');
+      verify.setAttribute('data-state','verified');
+      text(verify,'Ready — inference verified through the LLooM gateway.');
+      setStatus(detail,job.detail||'The gateway answered a real model request.',null);
+    }else{
+      verify.setAttribute('data-state','installed');
+      text(verify,job.healthy?'Installed — gateway running, inference not yet verified.':'Installed — verification needs attention.');
+      setStatus(detail,job.detail||'Files and configuration are in place. Verify the model through the gateway.','error');
+    }
+    if(job.endpoint){
+      text(endpoint,'Endpoint: '+job.endpoint);
+      endpoint.hidden=false;
+      open.disabled=false;
+    }
+    retry.hidden=true;
+  }
+
+  function pollJob(){
+    api('/gateway/first-run/job').then(function(data){
+      var job=data.job;
+      if(!job){setStatus(el('install-status'),'No active setup job.',null);return}
+      stage.job=job;
+      renderStages(job.stages);
+      if(job.stage&&job.stage.detail){setStatus(el('install-status'),job.stage.detail,null)}
+      else if(job.status==='running'){setStatus(el('install-status'),'Working…',null)}
+      if(job.status==='running'||job.status==='pending'){
+        window.setTimeout(pollJob,1200);
+        return;
+      }
+      setStatus(el('install-status'),job.status==='succeeded'?'Setup finished.':'Setup stopped.','');
+      applyResult(job);
+    }).catch(function(error){
+      setStatus(el('install-status'),'Lost contact with the setup server: '+error.message,'error');
+    });
+  }
+
+  function loadPlan(workloadId,recipeId){
+    var sequence=++planSequence;
+    stage.plan=null;
+    el('authorize').disabled=true;
+    el('to-review').disabled=true;
+    setStatus(el('recommendation'),'Reviewing compatible recipes for this machine…',null);
+    return api('/gateway/first-run/plan?workload='+encodeURIComponent(workloadId)+(recipeId?'&recipe='+encodeURIComponent(recipeId):'')).then(function(data){
+      if(sequence!==planSequence)return;
+      stage.plan=data.plan;
+      el('authorize').disabled=false;
+      renderMachine(stage.plan);
+      renderRecommendation(stage.plan);
+      renderReview(stage.plan);
+    }).catch(function(error){
+      if(sequence!==planSequence)return;
+      setStatus(el('recommendation'),'Could not build a plan: '+error.message,'error');
+    });
+  }
+
+  function initWorkloads(){
+    var host=clear(el('workloads'));
+    var workloads=[
+      {id:'chat',label:'Chat & write',description:'Everyday conversation, drafting, and reasoning.'},
+      {id:'code',label:'Code',description:'Editor and agent assistants that read and write code.'},
+      {id:'images',label:'Images',description:'Image generation and editing models.'},
+      {id:'voice',label:'Voice',description:'Speech synthesis and transcription.'}
+    ];
+    workloads.forEach(function(workload){
+      var button=document.createElement('button');
+      button.type='button';
+      button.className='choice';
+      button.setAttribute('aria-pressed',workload.id===stage.workloadId?'true':'false');
+      var icon=document.createElement('span');
+      icon.className='wk-icon';icon.setAttribute('aria-hidden','true');
+      icon.appendChild(svgIcon(workload.id));
+      var body=document.createElement('span');body.className='wk-body';
+      var label=document.createElement('strong');text(label,workload.label);
+      var description=document.createElement('span');description.className='wk-desc';text(description,workload.description);
+      body.appendChild(label);body.appendChild(description);
+      var check=document.createElement('span');check.className='wk-check';check.setAttribute('aria-hidden','true');
+      button.appendChild(icon);button.appendChild(body);button.appendChild(check);
+      button.addEventListener('click',function(){
+        stage.workloadId=workload.id;
+        var all=host.querySelectorAll('.choice');
+        for(var i=0;i<all.length;i+=1){all[i].setAttribute('aria-pressed',all[i]===button?'true':'false')}
+        loadPlan(workload.id);
+      });
+      host.appendChild(button);
+    });
+  }
+
+  el('to-choose').addEventListener('click',function(){show('step-choose')});
+  el('back-discover').addEventListener('click',function(){show('step-discover')});
+  el('to-review').addEventListener('click',function(){show('step-review')});
+  el('back-choose').addEventListener('click',function(){show('step-choose')});
+  el('authorize').addEventListener('click',function(){
+    if(!stage.plan||!stage.plan.planId){setStatus(el('authorize-note'),'No reviewed plan is loaded. Go back and plan again.','error');return}
+    var button=el('authorize');
+    button.disabled=true;
+    setStatus(el('authorize-note'),'Authorizing the reviewed plan. This is the only step that installs anything.',null);
+    api('/gateway/first-run/apply',{method:'POST',body:{planId:stage.plan.planId,yes:true,workloadId:stage.workloadId}})
+      .then(function(data){
+        show('step-ready');
+        stage.job=data.job;
+        renderStages(data.job&&data.job.stages);
+        setStatus(el('install-status'),'Setup started. Progress is reported by real stages below.',null);
+        pollJob();
+      })
+      .catch(function(error){
+        button.disabled=false;
+        setStatus(el('authorize-note'),'Could not start setup: '+error.message,'error');
+      });
+  });
+  el('retry').addEventListener('click',function(){
+    el('result-card').hidden=true;
+    setStatus(el('install-status'),'Reloading the plan to retry.',null);
+    show('step-choose');
+    loadPlan(stage.workloadId).then(function(){show('step-review')});
+  });
+  el('open-lloom').addEventListener('click',function(){
+    if(stage.job&&stage.job.endpoint){
+      window.open(stage.job.endpoint,'_blank','noopener');
+    }
+  });
+
+  initWorkloads();
+  var machineCard=document.querySelector('.machine');
+  var machineSlot=el('machine-slot');
+  if(machineCard&&machineSlot){machineSlot.appendChild(machineCard);}
+  el('back-discover').hidden=true;
+  show('step-choose');
+  if(!token){
+    setStatus(el('recommendation'),'This setup session is missing its local token. Re-run lloom from a terminal to reopen setup.','error');
+  }else{
+    api('/gateway/first-run/job').then(function(data){
+      if(data.job){stage.job=data.job;show('step-ready');pollJob();}
+      else loadPlan(stage.workloadId);
+    }).catch(function(error){setStatus(el('recommendation'),error.message,'error');});
+  }
+})();
+</script>
+</body>
+</html>`;
+}
diff --git a/src/first-run.mjs b/src/first-run.mjs
new file mode 100644
index 0000000..be10040
--- /dev/null
+++ b/src/first-run.mjs
@@ -0,0 +1,257 @@
+import { randomBytes, timingSafeEqual, createHash } from 'node:crypto';
+import http from 'node:http';
+import { spawn } from 'node:child_process';
+import { firstRunPage } from './first-run-page.mjs';
+
+const MAX_BODY = 8192;
+const TTL = 30 * 60 * 1000;
+const WORKLOADS = new Set(['chat', 'code', 'images', 'voice']);
+const digest = (value) => createHash('sha256').update(String(value)).digest();
+const failure = (message, statusCode = 400) => Object.assign(new Error(message), { statusCode });
+
+export function shouldOpenSetup(args, { installed = false, interactive = false } = {}) {
+  if (
+    installed ||
+    args.some((arg) => ['--json', '--offline', '--no-browser', '--go', '--apply', '--recipe', '--format'].includes(arg))
+  )
+    return false;
+  return args.includes('--browser') || interactive;
+}
+
+export function openLocalBrowser(url) {
+  const command = process.platform === 'darwin' ? 'open' : process.platform === 'win32' ? 'rundll32' : 'xdg-open';
+  const args = process.platform === 'win32' ? ['url.dll,FileProtocolHandler', url] : [url];
+  return new Promise((resolve) => {
+    const child = spawn(command, args, { stdio: 'ignore' });
+    const timer = setTimeout(() => {
+      child.kill();
+      resolve(false);
+    }, 10000);
+    child.once('error', () => {
+      clearTimeout(timer);
+      resolve(false);
+    });
+    child.once('exit', (code) => {
+      clearTimeout(timer);
+      resolve(code === 0);
+    });
+  });
+}
+
+async function readBody(req) {
+  if (!/^application\/json(?:;|$)/i.test(req.headers['content-type'] || ''))
+    throw failure('Expected application/json.', 415);
+  let bytes = 0,
+    text = '';
+  for await (const chunk of req) {
+    bytes += chunk.length;
+    if (bytes > MAX_BODY) throw failure('Request body too large.', 413);
+    text += chunk;
+  }
+  let body;
+  try {
+    body = JSON.parse(text);
+  } catch {
+    throw failure('Invalid JSON.');
+  }
+  if (!body || typeof body !== 'object' || Array.isArray(body)) throw failure('Expected a JSON object.');
+  return body;
+}
+
+export function createFirstRunServer({
+  planBuilder,
+  applyRunner,
+  gatewayStarter,
+  gatewayProbe,
+  now = Date.now,
+  port = 0,
+  host = '127.0.0.1',
+  planTtlMs = TTL,
+  logger = console
+}) {
+  if (host !== '127.0.0.1' && host !== '::1') throw failure('Setup must bind to loopback.');
+  if (typeof planBuilder !== 'function' || typeof applyRunner !== 'function')
+    throw failure('Setup requires a planner and installer.');
+  const secret = randomBytes(32).toString('base64url');
+  const plans = new Map();
+  let job = null;
+  let activeRun = null;
+  let activeEntry = null;
+  const hostname = host === '::1' ? '[::1]' : host;
+  const origin = () => 'http://' + hostname + ':' + server.address().port;
+  const reply = (res, status, body) => {
+    if (res.destroyed) return;
+    res.writeHead(status, {
+      'content-type': 'application/json; charset=utf-8',
+      'cache-control': 'no-store',
+      'x-content-type-options': 'nosniff',
+      'referrer-policy': 'no-referrer'
+    });
+    res.end(JSON.stringify(body));
+  };
+  const snapshot = () => (job ? structuredClone(job) : null);
+  function progress(event) {
+    if (!job || job.status !== 'running') return;
+    const detail = String(event?.message ?? event?.detail ?? event?.label ?? event?.type ?? 'Installing').slice(
+      0,
+      1500
+    );
+    const id = String(event?.phase ?? event?.stage ?? event?.id ?? 'install').slice(0, 120);
+    job.stage = { id, status: 'active', detail };
+    if (job.stages.at(-1)?.id !== id) job.stages.push({ id, title: detail, status: 'active' });
+    else job.stages.at(-1).title = detail;
+    if (job.stages.length > 50) job.stages.shift();
+  }
+  async function run(entry) {
+    try {
+      const report = await applyRunner(entry.plan, { onProgress: progress });
+      if (report?.ok === false)
+        throw new Error(
+          report.error ||
+            report.stages?.find((s) => ['failed', 'blocked'].includes(s.status))?.summary ||
+            'Installation did not complete. Run lloom doctor for details.'
+        );
+      job.stage = { id: 'gateway', status: 'active', detail: 'Starting the gateway' };
+      const gateway = gatewayStarter ? await gatewayStarter(report.configPath ?? entry.plan.configPath) : null;
+      if (gateway?.ok === false || gateway?.status === 'failed')
+        throw new Error(gateway.error || 'The gateway did not start.');
+      job.stage = { id: 'verify', status: 'active', detail: 'Checking a request through the gateway' };
+      const probe = gatewayProbe ? await gatewayProbe(report, entry.plan, gateway) : {};
+      job.inferenceVerified = probe?.inferenceVerified === true;
+      job.ready = job.inferenceVerified;
+      job.healthy = probe?.healthy === true;
+      job.endpoint = probe?.endpoint ?? gateway?.url ?? report.dashboardUrl ?? null;
+      job.detail = probe?.detail ?? 'Installed. Model inference has not been verified yet.';
+      job.status = 'succeeded';
+      job.stages.forEach((stage) => (stage.status = 'complete'));
+      job.stage = { id: 'verify', status: job.ready ? 'complete' : 'pending', detail: job.detail };
+      const verifyStage = job.stages.find((stage) => stage.id === 'verify');
+      const finalVerify = {
+        id: 'verify',
+        title: job.ready ? 'Inference verified' : 'Inference not verified',
+        status: job.ready ? 'complete' : 'pending'
+      };
+      if (verifyStage) Object.assign(verifyStage, finalVerify);
+      else job.stages.push(finalVerify);
+    } catch (error) {
+      job.status = 'failed';
+      job.error = String(error.message).slice(0, 1500);
+      job.stage = { ...job.stage, status: 'failed', detail: job.error };
+      if (job.stages.length) job.stages.at(-1).status = 'failed';
+    }
+  }
+  const server = http.createServer(async (req, res) => {
+    try {
+      // Exact authority stops DNS rebinding and alternate-port attacks.
+      if (req.headers.host !== new URL(origin()).host)
+        throw failure('Use the local setup address from your terminal.', 403);
+      const url = new URL(req.url, origin());
+      if (url.origin !== origin()) throw failure('Invalid setup address.', 403);
+      if (req.headers.origin !== undefined && req.headers.origin !== origin())
+        throw failure('Cross-site setup requests are denied.', 403);
+      if (req.headers['sec-fetch-site'] === 'cross-site') throw failure('Cross-site setup requests are denied.', 403);
+      if (req.method === 'GET' && url.pathname === '/') {
+        res.writeHead(200, {
+          'content-type': 'text/html; charset=utf-8',
+          'cache-control': 'no-store',
+          'referrer-policy': 'no-referrer',
+          'x-content-type-options': 'nosniff',
+          'content-security-policy':
+            "default-src 'none'; script-src 'unsafe-inline'; style-src 'unsafe-inline'; connect-src 'self'; base-uri 'none'; frame-ancestors 'none'; form-action 'none'"
+        });
+        res.end(firstRunPage());
+        return;
+      }
+      const token = String(req.headers.authorization || '').match(/^Bearer (.+)$/)?.[1] || '';
+      if (!timingSafeEqual(digest(token), digest(secret)))
+        throw failure('Setup session expired. Reopen setup from your terminal.', 401);
+      if (req.method === 'GET' && url.pathname === '/gateway/first-run/job') {
+        reply(res, 200, { ok: true, job: snapshot() });
+        return;
+      }
+      if (req.method === 'GET' && url.pathname === '/gateway/first-run/plan') {
+        if (job?.status === 'running')
+          throw failure('Installation is already running. Resume its progress instead.', 409);
+        const workloadId = url.searchParams.get('workload') || 'chat';
+        const recipeId = url.searchParams.get('recipe') || undefined;
+        if (!WORKLOADS.has(workloadId) || (recipeId && recipeId.length > 200))
+          throw failure('Unknown workload or recipe.');
+        const retry =
+          job?.status === 'failed' &&
+          activeEntry?.view.workloadId === workloadId &&
+          (!recipeId || activeEntry.view.selected.id === recipeId);
+        if (retry && now() - activeEntry.createdAt > planTtlMs)
+          throw failure('The reviewed recovery plan expired. Continue with lloom bootstrap from your terminal.', 410);
+        const built = retry ? activeEntry : await planBuilder({ workloadId, recipeId });
+        if (!built?.plan || !built?.view?.selected?.id)
+          throw failure('No compatible recipe was found for this machine.', 422);
+        for (const [id, value] of plans) if (now() - value.createdAt > planTtlMs) plans.delete(id);
+        while (plans.size >= 12) plans.delete(plans.keys().next().value);
+        const planId = randomBytes(24).toString('base64url');
+        plans.set(planId, { plan: built.plan, view: built.view, createdAt: retry ? activeEntry.createdAt : now() });
+        reply(res, 200, { ok: true, plan: { ...built.view, workloadId, planId } });
+        return;
+      }
+      if (req.method === 'POST' && url.pathname === '/gateway/first-run/apply') {
+        const body = await readBody(req);
+        if (Object.keys(body).some((key) => !['planId', 'yes', 'workloadId'].includes(key)))
+          throw failure('Apply accepts only a reviewed plan, workload, and confirmation.');
+        if (body.yes !== true) throw failure('Review and confirm the plan before installing.');
+        if (job?.planId === body.planId) {
+          reply(res, 200, { ok: true, job: snapshot(), resumed: true });
+          return;
+        }
+        if (job?.status === 'running') throw failure('Another installation is already running.', 409);
+        const entry = plans.get(body.planId);
+        if (!entry) throw failure('Plan not found. Review a fresh plan.', 404);
+        if (now() - entry.createdAt > planTtlMs) throw failure('This plan expired. Review a fresh plan.', 410);
+        if (body.workloadId && body.workloadId !== entry.view.workloadId)
+          throw failure('The workload differs from the reviewed plan.', 409);
+        plans.delete(body.planId);
+        activeEntry = entry;
+        job = {
+          id: randomBytes(16).toString('hex'),
+          planId: body.planId,
+          status: 'running',
+          stages: [],
+          stage: { id: 'install', status: 'active', detail: 'Installing the reviewed recipe' },
+          ready: false,
+          inferenceVerified: false,
+          healthy: false,
+          error: null,
+          endpoint: null,
+          detail: null
+        };
+        activeRun = run(entry);
+        reply(res, 202, { ok: true, job: snapshot() });
+        return;
+      }
+      throw failure('Setup endpoint not found.', 404);
+    } catch (error) {
+      if (!error.statusCode) logger.error?.('Setup request failed: ' + error.message);
+      reply(res, error.statusCode || 500, { ok: false, error: error.message });
+    }
+  });
+  server.requestTimeout = 30000;
+  return {
+    server,
+    bootstrapUrl: () => origin() + '/#setup=' + secret,
+    async listen() {
+      await new Promise((resolve, reject) => {
+        server.once('error', reject);
+        server.listen(port, host, () => {
+          server.off('error', reject);
+          resolve();
+        });
+      });
+      return this;
+    },
+    async close() {
+      await activeRun;
+      await new Promise((resolve, reject) => {
+        server.close((error) => (error ? reject(error) : resolve()));
+        server.closeIdleConnections();
+      });
+    }
+  };
+}
diff --git a/src/host-memory.mjs b/src/host-memory.mjs
index b78d8af..a5486fd 100644
--- a/src/host-memory.mjs
+++ b/src/host-memory.mjs
@@ -34,16 +34,30 @@ export function parseLinuxMeminfo(text) {
 }
 
 export function parseMacMemoryPressure(text, totalBytes) {
-  const match = String(text).match(/System-wide memory free percentage:\s*([\d.]+)%/i);
+  const text_ = String(text);
+  const page = (name) => Number(text_.match(new RegExp('^Pages ' + name + ':\\s*(\\d+)', 'm'))?.[1]);
+  const pageSize = Number(text_.match(/page size of (\d+)/)?.[1]) || 16384;
+  const free = page('free');
+  const inactive = page('inactive');
+  const speculative = page('speculative');
+  const purgeable = page('purgeable');
+  // XNU considers free + inactive + speculative + purgeable pages available;
+  // the "System-wide memory free percentage" is an opaque kernel estimate that
+  // can understate true availability while the file cache holds pages.
+  if ([free, inactive, speculative, purgeable].every(Number.isFinite)) {
+    return memorySnapshot(totalBytes, Math.min(totalBytes, (free + inactive + speculative + purgeable) * pageSize), 'macos-memory-pages');
+  }
+  const match = text_.match(/System-wide memory free percentage:\s*([\d.]+)%/i);
   const percentage = Number(match?.[1]);
   if (!Number.isFinite(percentage)) return null;
-  return memorySnapshot(totalBytes, (Number(totalBytes) * clamp(percentage, 0, 100)) / 100, 'macos-memory-pressure');
+  return memorySnapshot(totalBytes, (totalBytes * clamp(percentage, 0, 100)) / 100, 'macos-memory-pressure');
 }
 
 export async function readHostMemory({
   platform = process.platform,
   totalBytes = os.totalmem(),
   freeBytes = os.freemem(),
+  strict = false,
   readFile = fs.readFile,
   execFileImpl = execFileAsync
 } = {}) {
@@ -51,18 +65,22 @@ export async function readHostMemory({
     try {
       const snapshot = parseLinuxMeminfo(await readFile('/proc/meminfo', 'utf8'));
       if (snapshot) return snapshot;
-    } catch {
+    } catch (error) {
+      if (strict) throw error;
       // Fall through to the portable free-memory estimate.
     }
+    if (strict) throw new Error('Linux available-memory telemetry is unavailable');
   }
   if (platform === 'darwin') {
     try {
-      const { stdout } = await execFileImpl('/usr/bin/memory_pressure', ['-Q'], { timeout: 1500 });
+      const { stdout } = await execFileImpl('/usr/bin/memory_pressure', [], { timeout: strict ? 750 : 1500 });
       const snapshot = parseMacMemoryPressure(stdout, totalBytes);
       if (snapshot) return snapshot;
-    } catch {
+    } catch (error) {
+      if (strict) throw error;
       // Fall through when memory_pressure is unavailable.
     }
+    if (strict) throw new Error('macOS memory-pressure telemetry is unavailable');
   }
   return memorySnapshot(totalBytes, freeBytes, 'os-freemem');
 }
diff --git a/src/init.mjs b/src/init.mjs
index 66451c2..eade39b 100644
--- a/src/init.mjs
+++ b/src/init.mjs
@@ -1548,7 +1548,7 @@ export async function createInitPlan(
   };
 }
 
-async function writeJson(filePath, value) {
+async function writeJson(filePath, value, { exclusive = false } = {}) {
   await fs.mkdir(path.dirname(filePath), { recursive: true });
   const temporaryPath = path.join(
     path.dirname(filePath),
@@ -1556,22 +1556,37 @@ async function writeJson(filePath, value) {
   );
   try {
     await fs.writeFile(temporaryPath, `${JSON.stringify(value, null, 2)}\n`, { mode: 0o600 });
-    await fs.rename(temporaryPath, filePath);
+    if (exclusive) await fs.link(temporaryPath, filePath);
+    else await fs.rename(temporaryPath, filePath);
   } finally {
     await fs.unlink(temporaryPath).catch(() => {});
   }
 }
 
-export async function applyInit(config, { dryRun = true, yes = false, integrate = false, ...options } = {}) {
+export async function applyInit(
+  config,
+  {
+    dryRun = true,
+    yes = false,
+    integrate = false,
+    reviewedPlan,
+    exclusiveConfig = false,
+    onConfigWritten,
+    writeConfig,
+    ...options
+  } = {}
+) {
   if (!dryRun && !yes) {
     throw new Error(
       'Refusing to initialize LLooM without yes=true. Re-run with --yes after reviewing the dry-run plan.'
     );
   }
-  const plan = await createInitPlan(config, options);
+  const plan = reviewedPlan ?? (await createInitPlan(config, options));
   if (dryRun) return plan;
 
-  await writeJson(plan.configPath, plan.config);
+  if (writeConfig) await writeConfig(plan.configPath, plan.config);
+  else await writeJson(plan.configPath, plan.config, { exclusive: exclusiveConfig });
+  await onConfigWritten?.(plan.configPath, plan.config);
   const registry = createRegistry(plan.config);
   const generatedClients = await writeGeneratedIntegrationArtifacts(plan.config, registry, {
     clientId: options.clientId ?? 'all',
diff --git a/src/installation-jobs.mjs b/src/installation-jobs.mjs
new file mode 100644
index 0000000..5c0fb8e
--- /dev/null
+++ b/src/installation-jobs.mjs
@@ -0,0 +1,63 @@
+import { randomBytes } from 'node:crypto';
+
+const failure = (message, statusCode = 409) => Object.assign(new Error(message), { statusCode });
+
+// Installation survives a browser refresh; bootstrap's persisted stage state
+// remains responsible for resuming downloads after a gateway restart.
+export function createInstallationJobs({ plan, apply, now = Date.now, ttlMs = 1800000 }) {
+  const plans = new Map();
+  let job = null;
+  let running = null;
+  const snapshot = () => (job ? structuredClone(job) : null);
+  return {
+    snapshot,
+    async review(input) {
+      if (job?.status === 'running') throw failure('An installation is already running.');
+      const prepared = await plan(input);
+      for (const [id, value] of plans) if (now() - value.at > ttlMs) plans.delete(id);
+      while (plans.size >= 12) plans.delete(plans.keys().next().value);
+      const planId = randomBytes(24).toString('hex');
+      plans.set(planId, { prepared, at: now() });
+      return { planId, ...prepared.view };
+    },
+    start({ planId, yes, ...extra } = {}) {
+      if (yes !== true || Object.keys(extra).length) throw failure('Apply only the reviewed plan with yes: true.', 400);
+      if (job?.planId === planId) return snapshot();
+      if (job?.status === 'running') throw failure('An installation is already running.');
+      const entry = plans.get(planId);
+      if (!entry || now() - entry.at > ttlMs) throw failure('Review a fresh installation plan before continuing.');
+      plans.delete(planId);
+      job = {
+        id: randomBytes(16).toString('hex'),
+        planId,
+        status: 'running',
+        detail: 'Preparing the reviewed installation',
+        error: null
+      };
+      running = Promise.resolve()
+        .then(() =>
+          apply(entry.prepared, (event) => {
+            job.detail = String(event?.message ?? event?.detail ?? event?.label ?? event?.phase ?? 'Installing').slice(
+              0,
+              1500
+            );
+          })
+        )
+        .then((result) => {
+          if (result?.ok === false)
+            throw new Error(result.error || 'Installation needs attention. Run lloom doctor for details.');
+          job.status = 'succeeded';
+          job.detail = result?.message || 'Installation complete. Load the model to verify it through the gateway.';
+        })
+        .catch((error) => {
+          job.status = 'failed';
+          job.error = String(error.message).slice(0, 1500);
+        });
+      return snapshot();
+    },
+    async wait() {
+      await running;
+      return snapshot();
+    }
+  };
+}
diff --git a/src/installer.mjs b/src/installer.mjs
index 5e3ffce..b35a0b5 100644
--- a/src/installer.mjs
+++ b/src/installer.mjs
@@ -73,8 +73,8 @@ function backendState(state, backendId) {
   return state.backends[backendId];
 }
 
-function huggingFaceCommandCandidates(step, destination = step.destination) {
-  const configured = process.env.LLOOM_HF_BIN || process.env.HF_HUB_CLI;
+function huggingFaceCommandCandidates(step, destination = step.destination, env = process.env) {
+  const configured = env.LLOOM_HF_BIN || env.HF_HUB_CLI;
   const recipeOwned = destination ? path.join(path.dirname(destination), '.hf-cli', 'bin', 'hf') : null;
   return [configured, recipeOwned, 'hf', 'huggingface-cli'].filter(Boolean).map((command) => [
     command,
@@ -105,12 +105,43 @@ function pythonExecutableForVenv(venvPath) {
 }
 
 async function resolveHuggingFaceDownloadCommand(step, { env = process.env } = {}) {
-  for (const candidate of huggingFaceCommandCandidates(step)) {
+  for (const candidate of huggingFaceCommandCandidates(step, step.destination, env)) {
     if (await commandAvailable(candidate[0], { env })) return candidate;
   }
   return null;
 }
 
+// Bind browser-reviewed downloads to the executable selected at review. If a
+// recipe installs its own CLI later, its deterministic path remains reviewed.
+export async function pinDownloadCommands(plan, { env = process.env, requireAvailable = false } = {}) {
+  for (const step of plan.steps ?? []) {
+    const legacyDownload =
+      step.action === 'command' &&
+      Array.isArray(step.command) &&
+      ['hf', 'huggingface-cli'].includes(path.basename(step.command[0])) &&
+      step.command[1] === 'download';
+    if (step.action !== 'download-model' && !legacyDownload) continue;
+    const destinationIndex = step.command?.indexOf('--local-dir') ?? -1;
+    const destination = step.destination ?? (destinationIndex >= 0 ? step.command[destinationIndex + 1] : null);
+    const resolved = await resolveHuggingFaceDownloadCommand({ ...step, destination }, { env });
+    let executable = legacyDownload && path.isAbsolute(step.command[0]) ? step.command[0] : resolved?.[0];
+    if (executable && !path.isAbsolute(executable)) {
+      const which = await runCommand('/usr/bin/which', [executable], { env, allowFailure: true });
+      executable = which.stdout.trim();
+    }
+    if (!executable && requireAvailable)
+      throw new Error(
+        'The Hugging Face downloader is not installed. Choose a vendor recipe or install the Hugging Face CLI, then review again.'
+      );
+    if (!executable && !destination)
+      throw new Error('A reviewed Hugging Face download needs an installed CLI or an explicit --local-dir.');
+    step.downloadExecutable = executable || path.join(path.dirname(destination), '.hf-cli', 'bin', 'hf');
+    step.commands = (step.commands ?? [step.command]).map((command) => [step.downloadExecutable, ...command.slice(1)]);
+    step.command = step.commands[0];
+  }
+  return plan;
+}
+
 function stepCommand(step, { preferConfigured = true } = {}) {
   if (step.action === 'download-model') {
     if (step.provider !== 'huggingface') {
@@ -167,7 +198,7 @@ async function executeDownloadModel(step, { env = process.env, stdio } = {}) {
     };
   }
 
-  const command = await resolveHuggingFaceDownloadCommand(step, { env });
+  const command = step.downloadExecutable ? step.command : await resolveHuggingFaceDownloadCommand(step, { env });
   if (!command) {
     return {
       ok: false,
@@ -187,9 +218,12 @@ async function executeDownloadModel(step, { env = process.env, stdio } = {}) {
   const commands = [];
   let execution;
   for (const include of selections) {
-    const workCommand = huggingFaceCommandCandidates({ ...step, include }, prepared.workPath).find(
-      (candidate) => candidate[0] === command[0]
-    );
+    const template = step.downloadExecutable
+      ? (step.commands ?? [step.command])[commands.length]
+      : huggingFaceCommandCandidates({ ...step, include }, prepared.workPath, env).find(
+          (candidate) => candidate[0] === command[0]
+        );
+    const workCommand = template.map((arg, index) => (template[index - 1] === '--local-dir' ? prepared.workPath : arg));
     commands.push(workCommand);
     execution = await executeCommand(workCommand, { env, stdio });
     if (!execution.ok) return { ...execution, commands, status: 'failed', partialDestination: prepared.workPath };
@@ -345,14 +379,18 @@ export async function applyRecipe(
     onlyStep,
     env = process.env,
     onProgress,
-    stdio
+    stdio,
+    reviewedPlan
   } = {}
 ) {
   if (!dryRun && !yes) {
     throw new Error('Refusing to execute recipe without yes=true. Re-run with --yes after reviewing the dry-run plan.');
   }
 
-  const plan = planRecipe(recipe, config, { modelRoot, checkLocalReferences: false });
+  // A reviewed plan pins the exact executable steps approved by the user. When
+  // present, execute those steps verbatim without replanning, so a later
+  // profile/catalog change cannot alter the commands.
+  const plan = reviewedPlan ?? planRecipe(recipe, config, { modelRoot, checkLocalReferences: false });
   if (plan.validationErrors.length) {
     throw new Error(
       `Recipe ${recipe.id} is invalid:\n${plan.validationErrors.map((error) => `- ${error}`).join('\n')}`
@@ -447,7 +485,15 @@ export async function applyRecipe(
 
 export async function applyBackend(
   backend,
-  { dryRun = true, yes = false, statePath = defaultInstallStatePath, onlyStep, variables, env = process.env } = {}
+  {
+    dryRun = true,
+    yes = false,
+    statePath = defaultInstallStatePath,
+    onlyStep,
+    variables,
+    env = process.env,
+    reviewedPlan
+  } = {}
 ) {
   if (!dryRun && !yes) {
     throw new Error(
@@ -455,10 +501,13 @@ export async function applyBackend(
     );
   }
 
-  const plan = await planBackend(backend, {
-    variables,
-    checkCommands: true
-  });
+  // Reviewed backend plan from the approved dry run; execute its steps exactly.
+  const plan =
+    reviewedPlan ??
+    (await planBackend(backend, {
+      variables,
+      checkCommands: true
+    }));
   if (!plan.platformSupported) {
     throw new Error(`Backend ${backend.id} is not supported on ${plan.platform}`);
   }
diff --git a/src/model-installation.mjs b/src/model-installation.mjs
new file mode 100644
index 0000000..39108fa
--- /dev/null
+++ b/src/model-installation.mjs
@@ -0,0 +1,41 @@
+import { defaultBackendVariables, getBackend, loadBackendCatalog } from './backend-catalog.mjs';
+import { applyBackend } from './installer.mjs';
+import { runCommand } from './process-control.mjs';
+import { modelAcquisitionStatus, prepareModelAcquisition, finalizeModelAcquisition } from './model-acquisition.mjs';
+
+export async function installImportedModelAssets(plan, { onProgress, backendCatalog, env = process.env } = {}) {
+  const variables = defaultBackendVariables(env);
+  const installEnv = { ...env, PATH: [variables.shimDir, env.PATH].filter(Boolean).join(':') };
+  if (plan.additions.runtimeId) {
+    onProgress?.({ message: 'Installing the vendor backend' });
+    const catalog = backendCatalog ?? (await loadBackendCatalog());
+    const backend = getBackend(catalog, plan.inference.backend);
+    if (!backend) throw new Error('Unknown backend ' + plan.inference.backend);
+    const result = await applyBackend(backend, {
+      dryRun: false,
+      yes: true,
+      variables,
+      env: installEnv,
+      onProgress,
+      reviewedPlan: plan.backendPlan
+    });
+    const failed = result.results.filter((step) => ['failed', 'manual-required'].includes(step.status));
+    if (failed.length)
+      throw new Error(failed.map((step) => step.stderr || step.message || step.id + ' failed').join('\n'));
+  }
+  if (!plan.download?.command) return;
+  onProgress?.({ message: 'Downloading model files. Existing files are reused.' });
+  const acquisition = plan.download.acquisition;
+  if (plan.reference?.type === 'huggingface' && !acquisition)
+    throw new Error('Hugging Face imports require a reviewed acquisition plan.');
+  if (acquisition && (await modelAcquisitionStatus(acquisition)).complete) return;
+  const prepared = acquisition ? await prepareModelAcquisition(acquisition) : null;
+  const planned = plan.download.command;
+  const [command, ...args] = prepared
+    ? planned.map((arg, index) => (planned[index - 1] === '--local-dir' ? prepared.workPath : arg))
+    : planned;
+  const result = await runCommand(command, args, { allowFailure: true, env: installEnv, stdio: 'inherit' });
+  if (result.code !== 0)
+    throw new Error(result.stderr || 'Model download failed. Check the vendor requirements and credentials.');
+  if (prepared) await finalizeModelAcquisition(acquisition, prepared);
+}
diff --git a/src/protocol/openrouter-provider.mjs b/src/protocol/openrouter-provider.mjs
new file mode 100644
index 0000000..c42f3ad
--- /dev/null
+++ b/src/protocol/openrouter-provider.mjs
@@ -0,0 +1,103 @@
+/**
+ * OpenRouter provider-preference policy.
+ *
+ * OpenRouter routes each chat completion across multiple upstream providers and
+ * accepts a `provider` object on the request body. That object lets a caller
+ * constrain routing. LLooM exposes an operator-owned policy at
+ * `backends.<id>.openrouterProvider` so a configured lane always routes the way
+ * the operator intends.
+ *
+ * The policy is deliberately narrow:
+ *
+ *   { "openrouterProvider": { "only": ["z-ai"], "allow_fallbacks": false } }
+ *
+ * - `only` is required, non-empty, and every entry must be a non-empty string.
+ * - `allow_fallbacks` is optional and must be a boolean.
+ * - Any other key is rejected so the operator cannot believe a setting took
+ *   effect when the gateway did not enforce it.
+ *
+ * Enforcement is fail-closed. A malformed policy on an OpenRouter backend
+ * throws instead of silently sending an unconstrained request. Non-OpenRouter
+ * backends and backends without a policy leave the body untouched.
+ */
+
+const POLICY_KEY = 'openrouterProvider';
+const ALLOWED_KEYS = new Set(['only', 'allow_fallbacks']);
+
+function isPlainObject(value) {
+  return typeof value === 'object' && value !== null && !Array.isArray(value);
+}
+
+/** True only for the real OpenRouter API host, never a lookalike name. */
+export function isOpenRouterBackend(backend = {}) {
+  const baseUrl = backend?.baseUrl;
+  if (typeof baseUrl !== 'string' || !baseUrl.trim()) return false;
+  try {
+    return new URL(baseUrl).hostname === 'openrouter.ai';
+  } catch {
+    return false;
+  }
+}
+
+/**
+ * Validate the configured policy. Returns the normalized policy, `null` when no
+ * policy is configured, and throws for any malformed configuration.
+ */
+export function normalizeOpenRouterProviderPolicy(configured, { backendId } = {}) {
+  if (configured === undefined || configured === null) return null;
+  const where = backendId ? `backends.${backendId}.${POLICY_KEY}` : POLICY_KEY;
+  if (!isPlainObject(configured)) {
+    throw new Error(`${where} must be an object`);
+  }
+  for (const key of Object.keys(configured)) {
+    if (!ALLOWED_KEYS.has(key)) {
+      throw new Error(`${where} has unsupported key "${key}"; allowed keys are only, allow_fallbacks`);
+    }
+  }
+  if (!Object.hasOwn(configured, 'only')) {
+    throw new Error(`${where}.only is required`);
+  }
+  if (!Array.isArray(configured.only)) {
+    throw new Error(`${where}.only must be an array of provider slugs`);
+  }
+  if (configured.only.length === 0) {
+    throw new Error(`${where}.only must not be empty`);
+  }
+  const only = configured.only.map((entry) => {
+    if (typeof entry !== 'string' || !entry.trim()) {
+      throw new Error(`${where}.only entries must be non-empty strings`);
+    }
+    return entry.trim();
+  });
+  if (Object.hasOwn(configured, 'allow_fallbacks') && typeof configured.allow_fallbacks !== 'boolean') {
+    throw new Error(`${where}.allow_fallbacks must be a boolean`);
+  }
+  return {
+    only,
+    // Fail closed: an operator who constrains `only` almost never wants
+    // OpenRouter to silently substitute a different provider.
+    allow_fallbacks: configured.allow_fallbacks ?? false
+  };
+}
+
+/**
+ * Apply the OpenRouter provider policy to an outbound chat-completions body.
+ *
+ * Configured values win over caller-supplied `provider` fields while every
+ * other caller field is preserved. Bodies that are not plain objects, backends
+ * on another host, and backends without a policy are returned unchanged.
+ */
+export function applyOpenRouterProviderPolicy(body, backend = {}) {
+  if (!isPlainObject(body)) return body;
+  // Only the real OpenRouter host is ever treated as OpenRouter. Every other
+  // backend, including lookalike hostnames, is left exactly as the caller sent
+  // it and never has a policy applied to it.
+  if (!isOpenRouterBackend(backend)) return body;
+  const policy = normalizeOpenRouterProviderPolicy(backend?.[POLICY_KEY], { backendId: backend?.id });
+  if (!policy) return body;
+  const callerProvider = isPlainObject(body.provider) ? body.provider : {};
+  return {
+    ...body,
+    provider: { ...callerProvider, only: [...policy.only], allow_fallbacks: policy.allow_fallbacks }
+  };
+}
diff --git a/src/runtime-manager.mjs b/src/runtime-manager.mjs
index 41a9389..a2681aa 100644
--- a/src/runtime-manager.mjs
+++ b/src/runtime-manager.mjs
@@ -15,8 +15,15 @@ import {
   runtimeResourcesByNode
 } from './cluster.mjs';
 import { cleanupPortListener, terminateProcessTree } from './process-control.mjs';
+import { memorySafetyPolicy, createMemorySafetyGuard, RuntimeMemorySafetyError } from './runtime-memory-safety.mjs';
+import { createRuntimeMemoryUsageSampler } from './runtime-memory-usage.mjs';
 
-import { maintenanceBlocksRouting, assertMaintenanceStartAllowed, maintenanceError } from './model-maintenance.mjs';
+import {
+  runtimeMaintenance,
+  maintenanceBlocksRouting,
+  assertMaintenanceStartAllowed,
+  maintenanceError
+} from './model-maintenance.mjs';
 
 const execFileAsync = promisify(execFile);
 const packageRoot = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '..');
@@ -90,6 +97,8 @@ function compactRuntime(runtimeId, runtime, config) {
   return {
     enabled: runtime.enabled === true,
     keepWarm: runtime.keepWarm === true,
+    maintenance: runtimeMaintenance(config, runtimeId),
+    memoryGb: runtime.memoryGb ?? runtime.memory?.requiredGb ?? null,
     maxConcurrency: runtimeMaxConcurrency(runtime),
     maxQueuedRequests: runtimeMaxQueuedRequests(runtime),
     interactiveReservedSlots: interactiveReservedSlots(runtime),
@@ -155,6 +164,7 @@ const LIVE_ADMISSION_FIELDS = new Set([
   'requestStartupWaitMs',
   'startupRetryAfterSeconds',
   'healthTimeoutMs',
+  'keepWarm',
   'preferredWarm'
 ]);
 
@@ -335,6 +345,7 @@ async function dockerContainerState(runtime) {
     const state = inspected.State ?? {};
     return {
       exists: true,
+      id: inspected.Id ?? null,
       running: state.Running === true,
       status: state.Status ?? (state.Running ? 'running' : 'stopped'),
       pid: state.Pid ?? null,
@@ -663,10 +674,18 @@ export function effectiveRuntimeArgs(runtimeId, runtime) {
 }
 
 export class RuntimeManager {
-  constructor(config, { logger = console, captureOutput = true, clusterCoordinator = null } = {}) {
+  constructor(
+    config,
+    { logger = console, captureOutput = true, clusterCoordinator = null, memorySampler, memoryUsageSampler } = {}
+  ) {
     this.config = config;
     this.logger = logger;
     this.captureOutput = captureOutput;
+    this.memorySampler = memorySampler;
+    this.memoryUsageSampler =
+      memoryUsageSampler ??
+      createRuntimeMemoryUsageSampler({ nodeId: currentNodeId(config), platform: process.platform });
+    this.memorySafetyFailures = new Map();
     this.processes = new Map();
     this.state = new Map();
     this.queues = new Map();
@@ -677,6 +696,7 @@ export class RuntimeManager {
     this.watchdogOperations = new Map();
     this.admissionQueue = Promise.resolve();
     this.activeAdmission = null;
+    this.desiredResidency = new Map();
     // Set when the owning server begins shutting down. Queued background
     // admissions must not start a runtime after shutdown has been requested;
     // the guard is re-evaluated inside the admission mutex.
@@ -690,6 +710,24 @@ export class RuntimeManager {
   // before any side effects. It sees the freshest runtimeManager.config, so a
   // queued admission whose runtime was disabled, suspended, unowned, or whose
   // supervision was torn down is skipped safely rather than started.
+  noteDesiredResidency(runtimeId, policy, generation) {
+    this.desiredResidency.set(runtimeId, { policy, generation });
+  }
+  settleDesiredResidency(runtimeId, policy, generation) {
+    if (
+      this.desiredResidency.get(runtimeId)?.generation === generation &&
+      (policy === 'always'
+        ? this.config.runtimes?.[runtimeId]?.keepWarm === true
+        : policy === 'preferred'
+          ? this.config.runtimes?.[runtimeId]?.preferredWarm === true
+          : !this.config.runtimes?.[runtimeId]?.keepWarm && !this.config.runtimes?.[runtimeId]?.preferredWarm)
+    )
+      this.desiredResidency.delete(runtimeId);
+  }
+  desiredResidencyPolicy(runtimeId) {
+    return this.desiredResidency.get(runtimeId)?.policy;
+  }
+
   resolveAdmissionGuard({ runtimeId, reason } = {}) {
     if (!runtimeId) return { ok: true };
     if (this.shuttingDown === true) {
@@ -699,7 +737,20 @@ export class RuntimeManager {
         message: `runtime manager is shutting down; ${runtimeId} admission skipped`
       };
     }
+    if (reason === 'keep-warm') {
+      const desired = this.desiredResidencyPolicy(runtimeId);
+      const runtime = this.config.runtimes?.[runtimeId];
+      if ((desired && desired !== 'always') || !runtime?.enabled || !runtime.keepWarm)
+        return { ok: false, code: 'keep_warm_revoked', message: 'Always ready was removed or disabled while queued.' };
+    }
     if (reason === 'preferred-warm-reconcile' || reason === 'preferred-warm') {
+      const desired = this.desiredResidencyPolicy(runtimeId);
+      if (desired && desired !== 'preferred')
+        return {
+          ok: false,
+          code: 'preferred_warm_revoked',
+          message: 'The saved readiness policy no longer requests a preferred restore.'
+        };
       const runtime = this.config.runtimes?.[runtimeId];
       if (!runtime || runtime.enabled !== true || runtime.preferredWarm !== true || runtime.keepWarm === true) {
         return {
@@ -916,9 +967,12 @@ export class RuntimeManager {
     return this.queues.get(runtimeId);
   }
 
-  withAdmissionLock(fn, { runtimeId = null, reason = 'runtime-admission', preemptible = false } = {}) {
+  withAdmissionLock(
+    fn,
+    { runtimeId = null, reason = 'runtime-admission', preemptible = false, allowPreemption = true } = {}
+  ) {
     const active = this.activeAdmission;
-    if (active?.runtimeId && active.runtimeId !== runtimeId) {
+    if (allowPreemption && active?.runtimeId && active.runtimeId !== runtimeId) {
       if (active.preemptible && !active.preemptionRequested) {
         active.preemptionRequested = true;
         const error = new Error(
@@ -1008,11 +1062,12 @@ export class RuntimeManager {
     return aborted;
   }
 
-  async status({ localOnly = false } = {}) {
+  async status({ localOnly = false, includeMemoryUsage = false } = {}) {
     const runtimes = {};
     const keepWarm = new Set(this.keepWarmRuntimeIds());
     const preferredWarm = new Set(this.preferredWarmRuntimeIds());
     const remoteNodes = new Map();
+    const localRuntimeIds = [];
     for (const [runtimeId, runtime] of Object.entries(this.config.runtimes ?? {})) {
       const placement = runtimePlacement(runtime, this.config);
       if (placement.mode === 'distributed') continue;
@@ -1021,6 +1076,7 @@ export class RuntimeManager {
         ? !runtime.node || nodeId === currentNodeId(this.config)
         : this.clusterCoordinator.isLocalNode(nodeId);
       if (localOnly && !isLocal) continue;
+      if (isLocal) localRuntimeIds.push(runtimeId);
       if (!isLocal && this.clusterCoordinator) {
         if (!remoteNodes.has(nodeId)) remoteNodes.set(nodeId, this.clusterCoordinator.nodeStatus(nodeId));
         const node = await remoteNodes.get(nodeId);
@@ -1146,8 +1202,19 @@ export class RuntimeManager {
         };
       }
     }
+    if (includeMemoryUsage && this.memoryUsageSampler && localRuntimeIds.length) {
+      try {
+        const usage = await this.memoryUsageSampler.sample({ runtimes, runtimeIds: localRuntimeIds });
+        for (const runtimeId of localRuntimeIds) {
+          if (usage[runtimeId]) runtimes[runtimeId].memoryUsage = usage[runtimeId];
+        }
+      } catch {
+        // Memory telemetry is observational only and must not change runtime health or lifecycle.
+      }
+    }
     return {
       runtimes,
+      memorySafety: memorySafetyPolicy(this.config),
       events: this.events
     };
   }
@@ -1469,6 +1536,19 @@ export class RuntimeManager {
     const leaderNode = nextConfig.cluster?.leaderNode ?? previousConfig.cluster?.leaderNode ?? nodeId;
     const changed = reconfigureRuntimeIds(previousConfig, nextConfig, { nodeId, leaderNode });
     const liveAdmissionChanged = liveAdmissionRuntimeIds(previousConfig, nextConfig);
+    if (!changed.length && liveAdmissionChanged.length) {
+      // Wait for a running admission decision to finish before changing its
+      // residency inputs. A readiness click must never interrupt a cold start
+      // or resume a runtime paused for another reason.
+      return this.withAdmissionLock(
+        () => {
+          this.config = nextConfig;
+          for (const runtimeId of liveAdmissionChanged) this.admitQueuedRequests(runtimeId);
+          return { changed, liveAdmissionChanged, results: [] };
+        },
+        { reason: 'readiness-change', allowPreemption: false }
+      );
+    }
     const distributed = new Set(
       changed.filter((runtimeId) => {
         const runtime = nextConfig.runtimes?.[runtimeId] ?? previousConfig.runtimes?.[runtimeId];
@@ -1530,7 +1610,7 @@ export class RuntimeManager {
         }
       }
       this.config = nextConfig;
-      for (const runtimeId of liveAdmissionChanged) this.resumeRuntime(runtimeId);
+      for (const runtimeId of liveAdmissionChanged) this.admitQueuedRequests(runtimeId);
       const startOrder = [...changed].sort(
         (left, right) => Number(distributed.has(right)) - Number(distributed.has(left))
       );
@@ -1777,6 +1857,84 @@ export class RuntimeManager {
       };
     }
 
+    return this.startLocalWithMemorySafety(runtimeId, runtime, { force, warmup, reason, signal });
+  }
+
+  async startLocalWithMemorySafety(runtimeId, runtime, { force, warmup, reason, signal }) {
+    const policy = memorySafetyPolicy(this.config);
+    const previousFailure = this.memorySafetyFailures.get(runtimeId);
+    const manualRetry = ['manual-start', 'admin-start', 'admin-admit', 'cli-start', 'cli-admit'].includes(reason);
+    if (previousFailure && policy.mode !== 'yolo' && !manualRetry) throw previousFailure;
+    if (manualRetry || policy.mode === 'yolo') this.memorySafetyFailures.delete(runtimeId);
+    const owned = { child: null, containerId: null };
+    let cleanupPromise;
+    const cleanup = () =>
+      (cleanupPromise ??= (async () => {
+        if (owned.child?.pid && owned.child.exitCode == null && owned.child.signalCode == null) {
+          if (owned.child.connected) owned.child.send({ type: 'memory-safety-abort' }, () => {});
+          const result = await terminateProcessTree([owned.child.pid], { termTimeoutMs: 150, killTimeoutMs: 1000 });
+          if (result.survivors.length || result.failed.length)
+            throw new Error('Memory safety cleanup could not stop the owned process tree');
+        }
+        if (owned.containerId) {
+          await execFileAsync('docker', ['kill', owned.containerId], { timeout: 5000 }).catch(async (error) => {
+            const state = await dockerContainerState(runtime);
+            if (state.id === owned.containerId && state.running) throw error;
+          });
+        }
+      })());
+    const guard = createMemorySafetyGuard({
+      policy,
+      runtimeId,
+      sample: this.memorySampler,
+      onAbort: (error) => {
+        this.memorySafetyFailures.set(runtimeId, error);
+        this.record({ runtimeId, event: 'memory-safety-abort', message: error.message, snapshot: error.snapshot });
+        this.logger.error?.(error.message);
+        void cleanup().catch(() => {});
+      }
+    });
+    const combinedSignal = signal ? AbortSignal.any([signal, guard.signal]) : guard.signal;
+    try {
+      await guard.check();
+      combinedSignal.throwIfAborted();
+      guard.start();
+      const result = await this.startLocalUnlocked(runtimeId, runtime, {
+        force,
+        warmup,
+        reason,
+        signal: combinedSignal,
+        owned,
+        guard,
+        policy
+      });
+      await guard.check();
+      combinedSignal.throwIfAborted();
+      if (owned.child?.connected) owned.child.send({ type: 'memory-safety-complete' }, () => {});
+      return result;
+    } catch (cause) {
+      const error = guard.signal.aborted ? guard.signal.reason : cause;
+      if (owned.child || owned.containerId || guard.signal.aborted || combinedSignal.aborted) {
+        // A container start may have completed after the first abort cleanup.
+        await cleanupPromise?.catch(() => {});
+        cleanupPromise = null;
+        await cleanup();
+        const state = this.stateFor(runtimeId);
+        state.lastError = error.message;
+        this.setStatus(
+          runtimeId,
+          'failed',
+          error.code === 'runtime_memory_safety_abort' ? 'memory-safety-abort' : 'start-aborted'
+        );
+      }
+      throw error;
+    } finally {
+      await guard.stop();
+    }
+  }
+
+  async startLocalUnlocked(runtimeId, runtime, { force, warmup, reason, signal, owned, guard, policy }) {
+    const state = this.stateFor(runtimeId);
     if (runtimeAdapter(runtime) === 'docker') {
       if (runtimeManagement(runtime) !== 'managed') {
         return { runtimeId, started: false, healthy: false, reason: 'externally-managed' };
@@ -1803,7 +1961,11 @@ export class RuntimeManager {
         this.record({ runtimeId, event: 'docker-create', bootstrapResult, reason });
       }
       this.setStatus(runtimeId, 'starting', reason);
+      container = await dockerContainerState(runtime);
+      if (!container.running) owned.containerId = container.id;
+      signal.throwIfAborted();
       const processResult = await dockerLifecycle('start', runtime);
+      signal.throwIfAborted();
       state.starts += 1;
       state.startedAt = nowIso();
       this.record({ runtimeId, event: 'docker-start', processResult, reason });
@@ -1839,16 +2001,28 @@ export class RuntimeManager {
     // CLI managers disable capture so their detached runtime does not keep the
     // short-lived command process open through an inherited pipe.
     const supervised = process.platform !== 'win32';
+    signal.throwIfAborted();
     const child = spawn(
       supervised ? process.execPath : runtime.command,
       supervised ? [path.join(packageRoot, 'src', 'runtime-supervisor.mjs'), runtime.command, ...args] : args,
       {
         cwd: runtime.cwd,
-        env: runtimeEnvironment(this.config, runtime),
-        stdio: ['ignore', this.captureOutput ? 'pipe' : 'ignore', this.captureOutput ? 'pipe' : 'ignore'],
+        env: { ...runtimeEnvironment(this.config, runtime), LLOOM_MEMORY_SAFETY_POLICY: JSON.stringify(policy) },
+        stdio: [
+          'ignore',
+          this.captureOutput ? 'pipe' : 'ignore',
+          this.captureOutput ? 'pipe' : 'ignore',
+          ...(supervised ? ['ipc'] : [])
+        ],
         detached: true
       }
     );
+    owned.child = child;
+    child.on('message', (message) => {
+      if (message?.type === 'memory-safety-abort')
+        guard.trip(new RuntimeMemorySafetyError(message.message, { runtimeId, snapshot: message.snapshot, policy }));
+    });
+    child.channel?.unref();
     child.unref();
     this.processes.set(runtimeId, child);
     state.starts += 1;
@@ -1874,6 +2048,13 @@ export class RuntimeManager {
       this.record({ runtimeId, event: 'error', message: state.lastError });
     });
     child.on('exit', (code, signal) => {
+      if (code === 78 && policy.mode === 'enforce')
+        guard.trip(
+          new RuntimeMemorySafetyError('Backend supervisor rejected the load to protect host memory.', {
+            runtimeId,
+            policy
+          })
+        );
       const expectedStop = state.status === 'stopping' || ['SIGTERM', 'SIGKILL'].includes(signal);
       state.status = code === 0 || expectedStop ? 'stopped' : 'failed';
       state.stoppedAt = nowIso();
@@ -1886,6 +2067,34 @@ export class RuntimeManager {
       }
     });
 
+    if (supervised) {
+      // Health cannot authorize a load until the independent supervisor has
+      // checked memory and actually spawned this backend.
+      await new Promise((resolve, reject) => {
+        const finish = (error) => {
+          clearTimeout(timer);
+          child.off('message', onMessage);
+          child.off('exit', onExit);
+          child.off('error', onError);
+          signal.removeEventListener('abort', onAbort);
+          if (error) reject(error);
+          else resolve();
+        };
+        const onMessage = (message) => {
+          if (message?.type === 'memory-safety-ready') finish();
+        };
+        const onExit = () =>
+          finish(signal.reason ?? new Error(`runtime ${runtimeId} supervisor exited before startup`));
+        const onError = (error) => finish(error);
+        const onAbort = () => finish(signal.reason);
+        const timer = setTimeout(() => finish(new Error(`runtime ${runtimeId} supervisor startup timed out`)), 5000);
+        child.on('message', onMessage);
+        child.once('exit', onExit);
+        child.once('error', onError);
+        signal.addEventListener('abort', onAbort, { once: true });
+        if (signal.aborted) onAbort();
+      });
+    }
     const result = await this.waitForHealth(runtimeId, runtime, child, { signal });
     let warmupResult = null;
     if (result.healthy && warmup && runtime.warmup) {
@@ -1905,6 +2114,7 @@ export class RuntimeManager {
     while (Date.now() < deadline) {
       if (signal?.aborted) throw signal.reason ?? new Error(`runtime ${runtimeId} start aborted`);
       if (await runtimeHealthOk(runtime)) {
+        signal?.throwIfAborted?.();
         this.setStatus(runtimeId, 'running');
         this.record({ runtimeId, event: 'healthy' });
         return { runtimeId, healthy: true };
@@ -1975,6 +2185,7 @@ export class RuntimeManager {
         signal
       });
       const text = await response.text().catch(() => '');
+      signal?.throwIfAborted?.();
       const result = {
         runtimeId,
         warmed: response.ok,
diff --git a/src/runtime-memory-safety.mjs b/src/runtime-memory-safety.mjs
new file mode 100644
index 0000000..20afda4
--- /dev/null
+++ b/src/runtime-memory-safety.mjs
@@ -0,0 +1,150 @@
+import os from 'node:os';
+import { readHostMemory } from './host-memory.mjs';
+
+const GiB = 1024 ** 3;
+
+export class RuntimeMemorySafetyError extends Error {
+  constructor(message, { runtimeId, snapshot, policy } = {}) {
+    super(message);
+    this.name = 'RuntimeMemorySafetyError';
+    this.code = 'runtime_memory_safety_abort';
+    this.type = 'runtime_memory_safety_error';
+    this.statusCode = 503;
+    this.temporary = false;
+    this.runtimeId = runtimeId;
+    this.snapshot = snapshot;
+    this.policy = policy;
+  }
+}
+
+// An operator configuration problem, not a transient load abort: wrong
+// statusCode/type so clients and the CLI report it as a config error.
+export class MemorySafetyConfigError extends Error {
+  constructor(message) {
+    super(message);
+    this.name = 'MemorySafetyConfigError';
+    this.code = 'memory_safety_config_invalid';
+    this.type = 'memory_safety_config_error';
+    this.statusCode = 500;
+    this.temporary = false;
+  }
+}
+
+export function memorySafetyPolicy(config, totalMemoryGb = os.totalmem() / GiB) {
+  const input = config.runtimePolicy?.memorySafety ?? {};
+  const configuredReserve = config.runtimePolicy?.reserveMemoryGb;
+  const minAvailableMemoryGb = input.minAvailableMemoryGb ?? Math.min(configuredReserve ?? 8, totalMemoryGb * 0.15);
+  const maxMemoryUtilization = Math.min(
+    input.maxMemoryUtilization ?? 0.9,
+    config.runtimePolicy?.maxMemoryUtilization ?? 0.9
+  );
+  const pollIntervalMs = input.pollIntervalMs ?? 250;
+  const mode = input.mode ?? 'enforce';
+  if (
+    !['enforce', 'yolo'].includes(mode) ||
+    !Number.isFinite(minAvailableMemoryGb) ||
+    minAvailableMemoryGb <= 0 ||
+    !Number.isFinite(maxMemoryUtilization) ||
+    maxMemoryUtilization <= 0 ||
+    maxMemoryUtilization >= 1 ||
+    !Number.isFinite(pollIntervalMs) ||
+    pollIntervalMs < 50 ||
+    pollIntervalMs > 1000
+  ) {
+    throw new MemorySafetyConfigError('Invalid memory safety limits; refusing to load a model.');
+  }
+  return { mode, minAvailableMemoryGb, maxMemoryUtilization, pollIntervalMs };
+}
+
+export function assertMemorySafety(policy, snapshot, runtimeId) {
+  if (policy.mode === 'yolo') return;
+  const total = snapshot?.totalBytes;
+  const available = snapshot?.availableBytes;
+  let detail;
+  if (!Number.isFinite(total) || total <= 0 || !Number.isFinite(available) || available < 0 || available > total) {
+    detail = 'host memory could not be measured';
+  } else if (available <= policy.minAvailableMemoryGb * GiB) {
+    detail = `${(available / GiB).toFixed(1)} GB available reached the ${policy.minAvailableMemoryGb.toFixed(1)} GB hard reserve`;
+  } else if (1 - available / total >= policy.maxMemoryUtilization) {
+    detail = `host memory use reached the ${(policy.maxMemoryUtilization * 100).toFixed(1)}% hard ceiling`;
+  }
+  if (detail)
+    throw new RuntimeMemorySafetyError(`Load aborted to protect this machine: ${detail}.`, {
+      runtimeId,
+      snapshot,
+      policy
+    });
+}
+
+// One operation owns one guard. Samples never overlap, and stop() waits for a
+// pending sample so a stale callback cannot kill a later operation.
+export function createMemorySafetyGuard({
+  policy,
+  runtimeId,
+  sample = () => readHostMemory({ strict: true }),
+  onAbort = () => {}
+}) {
+  const controller = new AbortController();
+  let stopped = false;
+  let timer;
+  let inFlight;
+  const trip = (cause) => {
+    if (stopped || controller.signal.aborted) return;
+    const error =
+      cause instanceof RuntimeMemorySafetyError
+        ? cause
+        : new RuntimeMemorySafetyError('Load aborted: host memory could not be measured.', { runtimeId, policy });
+    controller.abort(error);
+    onAbort(error);
+  };
+  const check = () => {
+    if (policy.mode === 'yolo' || stopped) return Promise.resolve();
+    controller.signal.throwIfAborted();
+    if (!inFlight) {
+      inFlight = (async () => {
+        // A stuck sampler is also unsafe; do not leave the load unguarded.
+        let deadline;
+        try {
+          const snapshot = await Promise.race([
+            Promise.resolve().then(sample),
+            new Promise((_, reject) => {
+              deadline = setTimeout(() => reject(new Error('Memory sampler timed out')), 1000);
+            })
+          ]);
+          if (!stopped) assertMemorySafety(policy, snapshot, runtimeId);
+        } catch (error) {
+          trip(error);
+        } finally {
+          clearTimeout(deadline);
+        }
+      })().finally(() => {
+        inFlight = null;
+      });
+    }
+    return inFlight.then(() => controller.signal.throwIfAborted());
+  };
+  const tick = async () => {
+    try {
+      await check();
+    } catch {
+      return;
+    }
+    if (!stopped) {
+      timer = setTimeout(tick, policy.pollIntervalMs);
+      timer.unref?.();
+    }
+  };
+  return {
+    signal: controller.signal,
+    check,
+    trip,
+    start() {
+      if (!stopped && policy.mode !== 'yolo') void tick();
+    },
+    async stop() {
+      stopped = true;
+      clearTimeout(timer);
+      await inFlight;
+    }
+  };
+}
diff --git a/src/runtime-memory-usage.mjs b/src/runtime-memory-usage.mjs
new file mode 100644
index 0000000..9379d86
--- /dev/null
+++ b/src/runtime-memory-usage.mjs
@@ -0,0 +1,415 @@
+import { execFile } from 'node:child_process';
+import { promisify } from 'node:util';
+import { fileURLToPath } from 'node:url';
+
+const execFileAsync = promisify(execFile);
+const LSOF = process.platform === 'darwin' ? '/usr/sbin/lsof' : 'lsof';
+const MAX_MODEL_BODY_BYTES = 64 * 1024;
+
+async function withDeadline(work, milliseconds) {
+  let timer;
+  try {
+    return await Promise.race([
+      Promise.resolve().then(work),
+      new Promise((_, reject) => {
+        timer = setTimeout(() => reject(new Error('Memory observation timed out')), milliseconds);
+      })
+    ]);
+  } finally {
+    clearTimeout(timer);
+  }
+}
+
+function positiveInteger(value) {
+  const number = Number(value);
+  return Number.isInteger(number) && number > 0 ? number : null;
+}
+
+function commandBasename(value) {
+  const first = Array.isArray(value) ? value[0] : value;
+  return (
+    String(first ?? '')
+      .trim()
+      .split(/\s+/)[0]
+      .split(/[\\/]/)
+      .pop() ?? ''
+  );
+}
+
+function loopbackHostname(hostname) {
+  return ['127.0.0.1', 'localhost', '[::1]', '::1'].includes(String(hostname ?? '').toLowerCase());
+}
+
+function loopbackUrl(value) {
+  try {
+    const url = new URL(String(value));
+    return url.protocol === 'http:' || url.protocol === 'https:' ? (loopbackHostname(url.hostname) ? url : null) : null;
+  } catch {
+    return null;
+  }
+}
+
+function endpointUrl(runtime, path = '/') {
+  const configured = loopbackUrl(runtime?.healthUrl);
+  if (runtime?.healthUrl && !configured) return null;
+  if (configured) {
+    if (path === '/') return configured;
+    const url = new URL(configured);
+    url.pathname = path;
+    url.search = '';
+    url.hash = '';
+    return url;
+  }
+  const port = positiveInteger(runtime?.port);
+  return port ? new URL(`http://127.0.0.1:${port}${path.startsWith('/') ? path : `/${path}`}`) : null;
+}
+
+function parseProcessRows(text) {
+  const rows = new Map();
+  for (const match of String(text ?? '').matchAll(/^\s*(\d+)\s+(\d+)\s+(\d+)\s*$/gm)) {
+    const pid = Number(match[1]);
+    const ppid = Number(match[2]);
+    if (pid > 0 && ppid >= 0) rows.set(pid, { pid, ppid, rss: Number(match[3]) * 1024 });
+  }
+  return [...rows.values()];
+}
+
+function parseLoopbackListeners(text) {
+  const listeners = new Map();
+  for (const line of String(text ?? '').split(/\r?\n/)) {
+    const match = line.match(/^\s*\S+\s+(\d+)\s+.*?\sTCP\s+(\S+)\s*(?:\(\s*LISTEN\s*\))?$/i);
+    if (!match) continue;
+    const pid = Number(match[1]);
+    const address = match[2];
+    const separator = address.lastIndexOf(':');
+    if (!Number.isInteger(pid) || pid <= 0 || separator < 1) continue;
+    const host = address.slice(0, separator);
+    const port = Number(address.slice(separator + 1));
+    if (!loopbackHostname(host) || !positiveInteger(port)) continue;
+    if (!listeners.has(port)) listeners.set(port, new Set());
+    listeners.get(port).add(pid);
+  }
+  return listeners;
+}
+
+function expandProcessTree(rootPids, rowsByPid) {
+  const selected = new Set(rootPids);
+  let changed = true;
+  while (changed) {
+    changed = false;
+    for (const row of rowsByPid ?? []) {
+      if (selected.has(row.ppid) && !selected.has(row.pid)) {
+        selected.add(row.pid);
+        changed = true;
+      }
+    }
+  }
+  return [...selected];
+}
+
+function boundedText(response, maxBytes) {
+  const stream = response?.body;
+  if (stream && typeof stream.getReader === 'function') {
+    return (async () => {
+      const reader = stream.getReader();
+      const chunks = [];
+      let bytes = 0;
+      try {
+        for (;;) {
+          const { done, value } = await reader.read();
+          if (done) break;
+          bytes += value?.byteLength ?? value?.length ?? 0;
+          if (bytes > maxBytes) throw new Error('memory usage response too large');
+          chunks.push(value);
+        }
+      } finally {
+        await reader.cancel().catch(() => {});
+        reader.releaseLock();
+      }
+      return Buffer.concat(chunks.map((chunk) => Buffer.from(chunk))).toString('utf8');
+    })();
+  }
+  return response.text().then((text) => {
+    if (Buffer.byteLength(text) > maxBytes) throw new Error('memory usage response too large');
+    return text;
+  });
+}
+
+function extractLoadedModelIds(payload, kind) {
+  if (payload == null || typeof payload !== 'object') throw new Error('unexpected model response');
+  if (kind === 'ollama') {
+    if (!Array.isArray(payload.models)) throw new Error('missing model residency list');
+    return (Array.isArray(payload.models) ? payload.models : [])
+      .map((item) => (typeof item === 'string' ? item : (item?.model ?? item?.name ?? item?.id)))
+      .filter((value) => typeof value === 'string' && value.trim())
+      .map((value) => value.trim());
+  }
+  if (!Array.isArray(payload.tts_loaded) || !Array.isArray(payload.stt_loaded))
+    throw new Error('missing audio residency lists');
+  const values = [
+    ...(Array.isArray(payload.tts_loaded) ? payload.tts_loaded : []),
+    ...(Array.isArray(payload.stt_loaded) ? payload.stt_loaded : [])
+  ];
+  return values
+    .map((value) => (typeof value === 'string' ? value : (value?.id ?? value?.name ?? value?.model)))
+    .filter((value) => typeof value === 'string' && value.trim())
+    .map((value) => value.trim());
+}
+
+export function createRuntimeMemoryUsageSampler({
+  nodeId = null,
+  platform = process.platform,
+  psReader = async () => {
+    const result = await execFileAsync('/bin/ps', ['-axo', 'pid=,ppid=,rss='], {
+      timeout: 1000,
+      maxBuffer: 2 * 1024 * 1024
+    });
+    return result.stdout;
+  },
+  listenerReader = async () => {
+    const result = await execFileAsync(LSOF, ['-nP', '-iTCP', '-sTCP:LISTEN'], {
+      timeout: 1000,
+      maxBuffer: 2 * 1024 * 1024
+    });
+    return result.stdout;
+  },
+  footprintReader = async (pids) => {
+    const result = await execFileAsync(
+      'python3',
+      [fileURLToPath(new URL('./darwin-memory-usage.py', import.meta.url)), ...pids.map(String)],
+      { timeout: 1000, maxBuffer: 2 * 1024 * 1024 }
+    );
+    return JSON.parse(result.stdout);
+  },
+  fetchImpl = globalThis.fetch,
+  clock = Date.now,
+  nowIso = () => new Date().toISOString(),
+  scheduleTimeout = (callback, milliseconds) => {
+    const timer = setTimeout(callback, milliseconds);
+    timer.unref?.();
+    return timer;
+  },
+  cacheTtlMs = 5000,
+  modelTimeoutMs = 500
+} = {}) {
+  const now = typeof clock?.now === 'function' ? clock.now.bind(clock) : (clock ?? Date.now);
+  let processInFlight = null;
+  let processCache = null;
+  const modelInFlight = new Map();
+  const modelCache = new Map();
+
+  async function readProcessSnapshot() {
+    const at = now();
+    if (processCache && at < processCache.expiresAt) return processCache.value;
+    if (processInFlight) return processInFlight;
+    processInFlight = Promise.allSettled([withDeadline(psReader, 1000), withDeadline(listenerReader, 1000)])
+      .then(async ([ps, lsof]) => {
+        const value = {
+          rows: ps.status === 'fulfilled' ? parseProcessRows(ps.value) : [],
+          listeners: lsof.status === 'fulfilled' ? parseLoopbackListeners(lsof.value) : new Map(),
+          processOk: ps.status === 'fulfilled',
+          listenerOk: lsof.status === 'fulfilled'
+        };
+        if (platform === 'darwin' && value.rows.length) {
+          try {
+            const footprints = await withDeadline(() => footprintReader(value.rows.map((row) => row.pid)), 1000);
+            for (const row of value.rows) {
+              const bytes = footprints?.[row.pid];
+              if (typeof bytes === 'number' && Number.isFinite(bytes) && bytes >= 0) row.footprint = bytes;
+            }
+          } catch {
+            /* Python or per-process accounting may be unavailable; retain RSS. */
+          }
+        }
+        processCache = { value, expiresAt: at + cacheTtlMs };
+        return value;
+      })
+      .finally(() => {
+        processInFlight = null;
+      });
+    return processInFlight;
+  }
+
+  async function readModelIds(url, kind) {
+    const key = url.toString();
+    const at = now();
+    const cached = modelCache.get(key);
+    if (cached && at < cached.expiresAt) return cached.value;
+    const inFlight = modelInFlight.get(key);
+    if (inFlight) return inFlight;
+    const request = withDeadline(
+      async () => {
+        let timer = null;
+        let signal = null;
+        if (typeof AbortController === 'function' && modelTimeoutMs > 0) {
+          const controller = new AbortController();
+          signal = controller.signal;
+          timer = scheduleTimeout(
+            () => controller.abort(new Error('model residency request timed out')),
+            modelTimeoutMs
+          );
+        }
+        try {
+          const response = await fetchImpl(url, { signal, redirect: 'error' });
+          if (!response?.ok) throw new Error(`model residency request failed (${response?.status ?? 'unknown'})`);
+          const body = await boundedText(response, MAX_MODEL_BODY_BYTES);
+          return extractLoadedModelIds(JSON.parse(body), kind);
+        } finally {
+          if (timer) clearTimeout(timer);
+        }
+      },
+      Math.max(1, modelTimeoutMs) + 50
+    )
+      .then(
+        (value) => {
+          const result = { value, expiresAt: now() + cacheTtlMs };
+          modelCache.set(key, result);
+          return value;
+        },
+        () => {
+          const result = { value: null, expiresAt: now() + cacheTtlMs };
+          modelCache.set(key, result);
+          return null;
+        }
+      )
+      .finally(() => {
+        modelInFlight.delete(key);
+      });
+    modelInFlight.set(key, request);
+    return request;
+  }
+
+  function runtimeRoots(runtime, snapshot) {
+    const roots = [];
+    const adapter = String(runtime?.adapter ?? '').toLowerCase();
+    if (adapter === 'docker') {
+      if (platform !== 'linux') return [];
+      const pid = positiveInteger(runtime?.containerPid ?? runtime?.container?.pid);
+      if (pid) roots.push(pid);
+    } else {
+      const pid = positiveInteger(runtime?.pid);
+      if (pid) roots.push(pid);
+    }
+    const url = endpointUrl(runtime);
+    const port = url ? positiveInteger(url.port || (url.protocol === 'https:' ? 443 : 80)) : null;
+    if (port) for (const pid of snapshot.listeners.get(port) ?? []) roots.push(pid);
+    return [...new Set(roots)];
+  }
+
+  return {
+    nodeId,
+    async sample({ runtimes = {}, runtimeIds = [] } = {}) {
+      if (!runtimeIds.length) return {};
+      const ids = [...new Set(runtimeIds)].filter((runtimeId) => {
+        const runtime = runtimes[runtimeId];
+        return (
+          runtime &&
+          runtime.remote !== true &&
+          runtime.distributed !== true &&
+          runtime.placement?.mode !== 'distributed' &&
+          (!nodeId ||
+            !(runtime.node ?? runtime.placement?.node) ||
+            (runtime.node ?? runtime.placement?.node) === nodeId) &&
+          ['running', 'external', 'starting', 'warming', 'stopping', 'draining'].includes(runtime.status)
+        );
+      });
+      if (!ids.length) return {};
+      const snapshot = await readProcessSnapshot();
+      const expansions = new Map();
+      const roots = new Map();
+      for (const runtimeId of ids) {
+        const runtime = runtimes[runtimeId];
+        const selectedRoots = runtimeRoots(runtime, snapshot);
+        roots.set(runtimeId, selectedRoots);
+        expansions.set(runtimeId, expandProcessTree(selectedRoots, snapshot.rows));
+      }
+      const parent = new Map();
+      const find = (pid) => {
+        if (!parent.has(pid)) return pid;
+        const root = find(parent.get(pid));
+        parent.set(pid, root);
+        return root;
+      };
+      const union = (left, right) => {
+        const a = find(left);
+        const b = find(right);
+        if (a === b) return;
+        parent.set(b, a);
+      };
+      for (const pids of expansions.values()) {
+        if (pids.length > 1) for (const pid of pids.slice(1)) union(pids[0], pid);
+      }
+      const groups = new Map();
+      for (const runtimeId of ids) {
+        const pids = expansions.get(runtimeId) ?? [];
+        const groupKey = pids.length ? `proc:${find(pids[0])}` : `runtime:${runtimeId}`;
+        if (!groups.has(groupKey)) {
+          groups.set(groupKey, { runtimeIds: [], pids: new Set(), urls: [] });
+        }
+        const group = groups.get(groupKey);
+        group.runtimeIds.push(runtimeId);
+        for (const pid of pids) group.pids.add(pid);
+      }
+      const result = {};
+      await Promise.all(
+        [...groups].map(async ([groupKey, group]) => {
+          group.runtimeIds.sort();
+          let residentBytes = null;
+          const usesFootprint =
+            platform === 'darwin' &&
+            group.pids.size > 0 &&
+            [...group.pids].every((pid) => snapshot.rows.find((row) => row.pid === pid)?.footprint != null);
+          let rowsAvailable = snapshot.rows.length > 0;
+          if (rowsAvailable && group.pids.size) {
+            let sum = 0;
+            for (const pid of group.pids) {
+              const row = snapshot.rows.find((row) => row.pid === pid);
+              if (!row) {
+                rowsAvailable = false;
+                break;
+              }
+              sum += usesFootprint ? row.footprint : row.rss;
+            }
+            if (rowsAvailable) residentBytes = sum;
+          }
+          const loadedByRuntime = new Map();
+          const modelCalls = [];
+          for (const runtimeId of group.runtimeIds) {
+            const runtime = runtimes[runtimeId];
+            loadedByRuntime.set(runtimeId, null);
+            const base = commandBasename(runtime.command);
+            const audio = [base, ...(runtime.args ?? []).map(commandBasename)].some((part) =>
+              ['lloom-audio-server', 'lloom_audio_server.py', 'lloom_audio_server'].includes(part)
+            );
+            const kind = base === 'ollama' ? 'ollama' : audio ? 'audio' : null;
+            if (!kind) continue;
+            const url = kind === 'ollama' ? endpointUrl(runtime, '/api/ps') : endpointUrl(runtime, '/health');
+            if (!url) continue;
+            modelCalls.push(
+              readModelIds(url, kind).then((value) => {
+                loadedByRuntime.set(runtimeId, value);
+              })
+            );
+          }
+          await Promise.all(modelCalls.splice(0, modelCalls.length));
+          const knownLists = [...loadedByRuntime.values()].filter((value) => Array.isArray(value));
+          const loadedModelIds =
+            knownLists.length === group.runtimeIds.length
+              ? [...new Set(knownLists.flat())].sort((left, right) => left.localeCompare(right))
+              : null;
+          const memoryUsage = {
+            residentBytes,
+            source: usesFootprint ? 'process-footprint' : 'process-rss',
+            sampledAt: nowIso(),
+            groupId: groupKey,
+            sharedRuntimeIds: group.runtimeIds,
+            loadedModelIds,
+            residencyKnown: loadedModelIds !== null
+          };
+          for (const runtimeId of group.runtimeIds) result[runtimeId] = { ...memoryUsage };
+        })
+      );
+      return result;
+    }
+  };
+}
diff --git a/src/runtime-policy.mjs b/src/runtime-policy.mjs
index 92e8cd7..c695194 100644
--- a/src/runtime-policy.mjs
+++ b/src/runtime-policy.mjs
@@ -169,7 +169,7 @@ function policyConfig(config, profile = {}) {
       ? Math.max(0, totalMemoryGb * maxMemoryUtilization)
       : Math.max(0, totalMemoryGb - reserveMemoryGb));
   return {
-    enabled: policy.enabled !== false,
+    enabled: policy.enabled !== false && policy.memorySafety?.mode !== 'yolo',
     autoEvict: policy.autoEvict === true,
     totalMemoryGb,
     reserveMemoryGb,
@@ -338,8 +338,10 @@ function clusterRuntimePolicyPlan(
       0,
       totalMemoryGb - (numberOrNull(nodeProfile.availableMemoryGb) ?? totalMemoryGb)
     );
-    const predictive = maxMemoryUtilization != null && nodeProfile.availableMemoryGb != null;
-    const projectedMemoryGb = (predictive ? actualUsedMemoryGb : loadedMemoryGb) + requestedAddsMemoryGb;
+    // The host reserve protects every process, even when the operator only
+    // configured a reserve or an absolute budget (rather than a percentage).
+    const predictive = nodeProfile.availableMemoryGb != null;
+    const projectedMemoryGb = Math.max(actualUsedMemoryGb, loadedMemoryGb) + requestedAddsMemoryGb;
     const overBudgetGb = Math.max(0, projectedMemoryGb - memoryBudgetGb);
     nodes[nodeId] = {
       ...nodeProfile,
@@ -370,7 +372,7 @@ function clusterRuntimePolicyPlan(
   }
 
   const policy = {
-    enabled: policyTemplate.enabled !== false,
+    enabled: policyTemplate.enabled !== false && policyTemplate.memorySafety?.mode !== 'yolo',
     autoEvict: policyTemplate.autoEvict === true,
     protectActiveRequests: policyTemplate.protectActiveRequests !== false,
     clustered: true
@@ -505,9 +507,25 @@ export async function createRuntimePolicyPlan(
     0,
     policy.totalMemoryGb - (numberOrNull(memoryProfile.availableMemoryGb) ?? policy.totalMemoryGb)
   );
-  const predictive = policy.maxMemoryUtilization != null;
-  const projectedMemoryGb = (predictive ? actualUsedMemoryGb : loadedMemoryGb) + requestedAddsMemory;
+  const predictive = memoryProfile.availableMemoryGb != null;
+
+  // Two separate limits:
+  //  - The modeled budget governs LLooM-attributable runtimes only. Host-wide
+  //    usage (OS, browser, other apps) is not evictable by admission, so it
+  //    must never count against this budget — folding it in makes the plan
+  //    impossible on any busy machine regardless of evictions.
+  //  - The live host signal feeds the hard reserve check: after this admission
+  //    (and any planned evictions), the host must keep reserve headroom.
+  const hostProjectedMemoryGb = Math.max(actualUsedMemoryGb, loadedMemoryGb) + requestedAddsMemory;
+  const projectedMemoryGb = loadedMemoryGb + requestedAddsMemory;
   let overBudgetGb = requested?.loaded ? 0 : Math.max(0, projectedMemoryGb - policy.memoryBudgetGb);
+  const evictionFreesGb = rows
+    .filter((row) => row.loaded && protectedReasons(row, policy).length === 0)
+    .reduce((sum, row) => sum + row.memoryGb, 0);
+  const hostShortfallGb = predictive
+    ? Math.max(0, policy.reserveMemoryGb - memoryProfile.availableMemoryGb + requestedAddsMemory - evictionFreesGb)
+    : 0;
+  overBudgetGb = Math.max(overBudgetGb, hostShortfallGb);
 
   const actions = [];
   const evictions = [];
@@ -555,7 +573,7 @@ export async function createRuntimePolicyPlan(
     overBudgetGb = Math.max(0, overBudgetGb - row.memoryGb);
   }
 
-  if (projectedMemoryGb > policy.memoryBudgetGb) {
+  if (projectedMemoryGb > policy.memoryBudgetGb || hostShortfallGb > 0) {
     actions.unshift(...evictions);
   }
 
@@ -594,7 +612,10 @@ export async function createRuntimePolicyPlan(
       availableMemoryGb: numberOrNull(memoryProfile.availableMemoryGb),
       predictive,
       requestedAddsMemoryGb: requestedAddsMemory,
-      projectedMemoryGb,
+      // Host-wide projection (loaded vs actual usage, plus the add): kept for
+      // observability; the allow decision uses the modeled budget + reserve.
+      projectedMemoryGb: Math.max(hostProjectedMemoryGb, projectedMemoryGb),
+      hostShortfallGb,
       memoryBudgetGb: policy.memoryBudgetGb
     },
     runtimes: rows,
@@ -631,7 +652,7 @@ export async function applyRuntimePolicyPlan(
     // that config toggles/reloads queued behind an earlier admission cannot
     // start a runtime the operator already disabled or suspended.
     const isPreferred = preferredRestore || reason === 'preferred-warm';
-    const readLiveConfig = () => (isPreferred ? (runtimeManager.config ?? config) : config);
+    const readLiveConfig = () => runtimeManager.config ?? config;
     let planSource = null;
     const assertFreshGuard = () => {
       admissionSignal?.throwIfAborted?.();
@@ -661,9 +682,20 @@ export async function applyRuntimePolicyPlan(
     liveConfig = readLiveConfig();
     assertFreshGuard();
     planSource = liveConfig;
+    // Overlay only pending hard pins for planning. Other policy changes adopt
+    // the live configuration under the admission mutex.
+    const desiredPins = Object.fromEntries(
+      Object.entries(liveConfig.runtimes ?? {}).map(([id, runtime]) => [
+        id,
+        runtimeManager.desiredResidencyPolicy?.(id) === 'always' ? { ...runtime, keepWarm: true } : runtime
+      ])
+    );
+    liveConfig = { ...liveConfig, runtimes: desiredPins };
     const planConfig = isPreferred
       ? { ...liveConfig, runtimePolicy: { ...liveConfig.runtimePolicy, enabled: true, protectActiveRequests: true } }
-      : liveConfig;
+      : reason === 'keep-warm'
+        ? { ...liveConfig, runtimePolicy: { ...liveConfig.runtimePolicy, enabled: true } }
+        : liveConfig;
     const idleGrace = preferredRestore
       ? Number(liveConfig.runtimePolicy?.preferredWarmIdleMs ?? preferredWarmIdleMs)
       : preferredWarmIdleMs;
@@ -676,7 +708,7 @@ export async function applyRuntimePolicyPlan(
       profile
     });
     assertFreshGuard();
-    if (isPreferred && readLiveConfig() !== liveConfig)
+    if (isPreferred && readLiveConfig() !== planSource)
       throw new RuntimeAdmissionError('Residency configuration changed; retry reconciliation', {
         code: 'runtime_residency_changed'
       });
@@ -775,6 +807,10 @@ export async function applyRuntimePolicyPlan(
         const statusBeforeDrain = (runtimeManager.stateFor?.(action.runtimeId) ?? status.runtimes?.[action.runtimeId])
           ?.statusSince;
         const assertIdleVictim = (afterDrain = false) => {
+          if (runtimeManager.desiredResidencyPolicy?.(action.runtimeId) === 'always')
+            throw new RuntimeAdmissionError('A saved Always ready preference protects this runtime.', {
+              code: 'runtime_keep_warm_conflict'
+            });
           if (!preferredRestore) return;
           const current =
             runtimeManager.config?.runtimes?.[action.runtimeId] ?? liveConfig.runtimes?.[action.runtimeId];
diff --git a/src/runtime-preferences.mjs b/src/runtime-preferences.mjs
new file mode 100644
index 0000000..6e89c76
--- /dev/null
+++ b/src/runtime-preferences.mjs
@@ -0,0 +1,109 @@
+import { keepWarmOwnership } from './runtime-manager.mjs';
+
+export const RESIDENCY_POLICIES = new Set(['auto', 'preferred', 'always']);
+
+function invalid(message, statusCode = 400) {
+  return Object.assign(new Error(message), { statusCode, code: 'invalid_residency_policy' });
+}
+
+export function residencyPolicy(runtime) {
+  if (runtime?.keepWarm === true || runtime?.evictable === false || runtime?.policy?.evictable === false) {
+    return 'always';
+  }
+  return runtime?.preferredWarm === true ? 'preferred' : 'auto';
+}
+
+export function updateRuntimePreference(raw, runtimeId, policy) {
+  if (!RESIDENCY_POLICIES.has(policy)) throw invalid('Choose auto, preferred, or always.');
+  if (!Object.hasOwn(raw.runtimes ?? {}, runtimeId)) throw invalid('Unknown managed runtime.', 404);
+  const runtime = raw.runtimes[runtimeId];
+  if (!runtime || runtime.enabled !== true || runtime.managed === false || runtime.management === 'external') {
+    throw invalid('Readiness is available only for enabled, managed runtimes.');
+  }
+  const ownership = keepWarmOwnership(raw, runtimeId);
+  if (!ownership.owned) throw invalid('Change readiness on the machine that owns this runtime.', 409);
+  if (runtime.maintenance) throw invalid('Resume this model before changing its readiness.', 409);
+  const previousPolicy = residencyPolicy(runtime);
+  // Remove the legacy pin aliases, otherwise normalization silently restores
+  // the pin even after a user chooses Auto.
+  delete runtime.evictable;
+  if (runtime.policy && typeof runtime.policy === 'object') {
+    delete runtime.policy.evictable;
+  }
+  runtime.keepWarm = policy === 'always';
+  runtime.preferredWarm = policy === 'preferred';
+  return { runtimeId, policy, previousPolicy };
+}
+
+export function createRuntimePreferenceController({
+  getConfig,
+  mutateSource,
+  reload,
+  assertControl,
+  onPersisted,
+  onApplied
+}) {
+  let writes = Promise.resolve(),
+    reconciles = Promise.resolve(),
+    sequence = 0;
+  const jobs = new Map();
+  const snapshot = (id) => {
+    const job = jobs.get(id);
+    return job ? { ...job, completion: undefined } : null;
+  };
+  async function submit(runtimeId, { policy, yes = false, requestedBy } = {}) {
+    const write = writes.then(async () => {
+      if (yes !== true) throw invalid('Review the readiness policy and confirm with yes: true.');
+      if (!RESIDENCY_POLICIES.has(policy)) throw invalid('Choose auto, preferred, or always.');
+      if (!getConfig().sourcePath) throw invalid('This gateway has no writable installed configuration.', 409);
+      assertControl?.(runtimeId, requestedBy);
+      let result;
+      const mutation = await mutateSource((raw) => {
+        result = updateRuntimePreference(raw, runtimeId, policy);
+      });
+      for (const id of jobs.keys()) if (!Object.hasOwn(getConfig().runtimes ?? {}, id)) jobs.delete(id);
+      const job = { ...result, id: ++sequence, changed: mutation.changed, status: 'pending', error: null };
+      jobs.set(runtimeId, job);
+      onPersisted?.(runtimeId, policy, job.id);
+      const completion = reconciles.then(async () => {
+        try {
+          await reload();
+          if (jobs.get(runtimeId) !== job || residencyPolicy(getConfig().runtimes?.[runtimeId]) !== policy) {
+            job.status = 'superseded';
+            return { ok: true, ...result, status: 'superseded', message: 'Replaced by a newer readiness change.' };
+          }
+          onApplied?.(runtimeId, policy, job.id);
+          job.status = 'succeeded';
+          return {
+            ok: true,
+            ...result,
+            changed: mutation.changed,
+            message: 'Readiness saved. Loading still uses normal memory admission.'
+          };
+        } catch (error) {
+          job.status = 'failed';
+          job.error = error.message;
+          throw error;
+        }
+      });
+      job.completion = completion;
+      reconciles = completion.catch(() => {});
+      return job;
+    });
+    writes = write.then(
+      () => {},
+      () => {}
+    );
+    return write;
+  }
+  return {
+    snapshot,
+    async request(id, options) {
+      const job = await submit(id, options);
+      return { ok: true, ...job, completion: undefined };
+    },
+    async set(id, options) {
+      return (await submit(id, options)).completion;
+    }
+  };
+}
diff --git a/src/runtime-supervisor.mjs b/src/runtime-supervisor.mjs
index bc3ac4f..495b836 100644
--- a/src/runtime-supervisor.mjs
+++ b/src/runtime-supervisor.mjs
@@ -2,6 +2,8 @@
 // alive too, so an exited backend cannot leave model workers behind.
 import { spawn } from 'node:child_process';
 import { setTimeout as delay } from 'node:timers/promises';
+import { memorySafetyPolicy, assertMemorySafety, createMemorySafetyGuard } from './runtime-memory-safety.mjs';
+import { readHostMemory } from './host-memory.mjs';
 
 const [command, ...args] = process.argv.slice(2);
 if (!command || process.platform === 'win32') {
@@ -9,12 +11,29 @@ if (!command || process.platform === 'win32') {
   process.exit(2);
 }
 
+let memoryPolicy = null;
+try {
+  if (process.env.LLOOM_MEMORY_SAFETY_POLICY) {
+    memoryPolicy = memorySafetyPolicy({
+      runtimePolicy: { memorySafety: JSON.parse(process.env.LLOOM_MEMORY_SAFETY_POLICY) }
+    });
+    if (memoryPolicy.mode !== 'yolo') assertMemorySafety(memoryPolicy, await readHostMemory({ strict: true }));
+  }
+} catch (error) {
+  process.send?.({ type: 'memory-safety-abort', message: error.message, snapshot: error.snapshot });
+  process.stderr.write(`${error.message}\n`);
+  process.exit(78);
+}
+
 // Separate the backend group from the supervisor, allowing escalation even
 // after the backend exits or while a worker ignores SIGTERM.
 const child = spawn(command, args, { detached: true, stdio: ['ignore', 'pipe', 'pipe'] });
 let exitCode = 1;
 let stopping = false;
 let cleanupPromise;
+let memoryGuard;
+let memoryComplete = false;
+let memoryAborted = false;
 
 function forward(source, destination) {
   source.pipe(destination, { end: false });
@@ -40,6 +59,14 @@ function signalGroup(signal) {
 
 function cleanup() {
   cleanupPromise ??= (async () => {
+    await memoryGuard?.stop();
+    if (memoryAborted) {
+      // SIGKILL was already delivered to the owned group. Reap the child;
+      // probing a dying process group can return EPERM on macOS.
+      const deadline = Date.now() + 1000;
+      while (Date.now() < deadline && child.exitCode == null && child.signalCode == null) await delay(10);
+      process.exit(78);
+    }
     signalGroup('SIGTERM');
     const deadline = Date.now() + 2000;
     while (Date.now() < deadline && signalGroup(0)) await delay(50);
@@ -54,12 +81,50 @@ function cleanup() {
   return cleanupPromise;
 }
 
+function abortForMemory(error) {
+  if (memoryAborted || memoryComplete) return;
+  memoryAborted = true;
+  exitCode = 78;
+  if (process.connected)
+    process.send({ type: 'memory-safety-abort', message: error.message, snapshot: error.snapshot });
+  if (!process.stderr.destroyed) process.stderr.write(`${error.message}\n`);
+  // This group belongs to this supervisor. Do not wait for the gateway or a
+  // graceful backend shutdown while the host is running out of memory.
+  signalGroup('SIGKILL');
+  void cleanup();
+}
+
+if (memoryPolicy?.mode === 'enforce') {
+  memoryGuard = createMemorySafetyGuard({ policy: memoryPolicy, onAbort: abortForMemory });
+  memoryGuard.start();
+}
+process.on('message', (message) => {
+  if (message?.type === 'memory-safety-complete') {
+    memoryComplete = true;
+    void memoryGuard?.stop();
+    // Disconnecting synchronously while Node drains queued IPC messages can
+    // crash its message dispatcher, leaving the backend without a supervisor.
+    setImmediate(() => {
+      if (process.connected) process.disconnect();
+    });
+  } else if (message?.type === 'memory-safety-abort') {
+    abortForMemory(new Error('Gateway aborted this load to protect host memory.'));
+  }
+});
+process.on('disconnect', () => {
+  if (memoryGuard && !memoryComplete && !memoryAborted)
+    abortForMemory(new Error('Gateway disconnected during a guarded load.'));
+});
+
 child.on('error', (error) => {
   if (!process.stderr.destroyed) process.stderr.write(`runtime launch failed: ${error.message}\n`);
   void cleanup();
 });
+child.on('spawn', () => {
+  if (process.connected) process.send({ type: 'memory-safety-ready' }, () => {});
+});
 child.on('exit', (code) => {
-  exitCode = code ?? 1;
+  if (!memoryAborted) exitCode = code ?? 1;
   void cleanup();
 });
 for (const signal of ['SIGTERM', 'SIGINT', 'SIGHUP']) {
diff --git a/src/security.mjs b/src/security.mjs
index 1a40597..4bbbceb 100644
--- a/src/security.mjs
+++ b/src/security.mjs
@@ -1,3 +1,5 @@
+import { isIP } from 'node:net';
+
 /**
  * Local gateway auth and admin perimeter.
  *
@@ -20,10 +22,10 @@ export function isLoopbackAddress(host) {
   value = value.split('%')[0];
   if (value.includes(':') && !value.includes('::') && value.split(':').length === 2) {
     const [maybeHost] = value.split(':');
-    if (LOOPBACK_HOSTS.has(maybeHost) || maybeHost === '127.0.0.1') return true;
+    if (LOOPBACK_HOSTS.has(maybeHost) || (isIP(maybeHost) === 4 && maybeHost.startsWith('127.'))) return true;
   }
   if (LOOPBACK_HOSTS.has(value)) return true;
-  if (value.startsWith('127.')) return true;
+  if (isIP(value) === 4 && value.startsWith('127.')) return true;
   return false;
 }
 
@@ -73,7 +75,7 @@ export function classifyRoute(method, pathname) {
   ) {
     return 'public';
   }
-  if (path.startsWith('/gateway/')) {
+  if (path.startsWith('/gateway/') || path === '/v1/integrations') {
     return verb === 'POST' ? 'admin-write' : 'admin-read';
   }
   return 'inference';
@@ -87,6 +89,11 @@ export function authorizeRequest(req, config, { method, pathname } = {}) {
     effectiveMethod === 'GET' &&
     ['/gateway/status', '/gateway/metrics', '/gateway/models'].includes(effectivePath);
   const routeKind = publicTelemetry ? 'public' : classifyRoute(effectiveMethod, effectivePath);
+  const browserBoundary = authorizeAdminBrowserRequest(req, config, {
+    method: effectiveMethod,
+    pathname: effectivePath
+  });
+  if (!browserBoundary.ok) return { ...browserBoundary, routeKind };
   const bindHost = config.server?.host ?? '127.0.0.1';
   const loopbackBind = isLoopbackAddress(bindHost);
   const allowMissing = config.security?.allowMissingAuth === true;
@@ -112,6 +119,15 @@ export function authorizeRequest(req, config, { method, pathname } = {}) {
           'Admin write endpoints are disabled on non-loopback binds. Bind to 127.0.0.1 or set security.allowRemoteAdmin=true with API keys.'
       };
     }
+    if (routeKind === 'admin-write' && (!adminRequired || !validAdminKey)) {
+      return {
+        ok: false,
+        status: 403,
+        code: 'admin_key_required',
+        routeKind,
+        message: 'Remote admin writes require a separate security.adminApiKeys credential.'
+      };
+    }
     if (isAdminRoute) {
       if (!validAdminKey) {
         return {
@@ -163,6 +179,50 @@ export function authorizeRequest(req, config, { method, pathname } = {}) {
   };
 }
 
+// Local command-line clients may omit Origin. Browsers may not use another
+// website's origin (or a rebound DNS hostname) to reach the local control plane.
+// This gate is independent of bearer auth: a stolen browser origin must never
+// gain admin access merely because loopback allows missing inference auth.
+export function authorizeAdminBrowserRequest(req, config, { method = req.method, pathname = '/' } = {}) {
+  if (!pathname.startsWith('/gateway/') && pathname !== '/v1/integrations') return { ok: true };
+  if (pathname === '/gateway/dashboard' || pathname === '/gateway/security') return { ok: true };
+  const denied = {
+    ok: false,
+    status: 403,
+    code: 'admin_origin_denied',
+    message: 'Open the dashboard on this gateway to use its management API.'
+  };
+  const host = req.headers?.host;
+  const origin = req.headers?.origin;
+  const site = req.headers?.['sec-fetch-site'];
+  const loopback = isLoopbackAddress(config.server?.host ?? '127.0.0.1');
+  let target;
+  try {
+    if (host) target = new URL('http://' + host);
+    if (target && (target.username || target.password || target.pathname !== '/' || target.search || target.hash))
+      return denied;
+    if (loopback && target) {
+      const hostname = target.hostname.replace(/^\[|\]$/g, '');
+      if (hostname !== 'localhost' && (!isIP(hostname) || !isLoopbackAddress(hostname))) return denied;
+    }
+    if (site === 'cross-site') return denied;
+    if (origin !== undefined) {
+      if (!target || origin === 'null') return denied;
+      const source = new URL(origin);
+      if (!['http:', 'https:'].includes(source.protocol) || source.origin !== origin) return denied;
+      // An HTTPS reverse proxy may supply the browser-facing host. Do not trust
+      // forwarded headers or broaden this to arbitrary sibling origins.
+      if (source.host !== target.host) return denied;
+      if (req.socket?.encrypted && source.protocol !== 'https:') return denied;
+      if (loopback && source.protocol !== 'http:') return denied;
+    }
+    if (method === 'POST' && site && !['same-origin', 'none'].includes(site)) return denied;
+  } catch {
+    return denied;
+  }
+  return { ok: true };
+}
+
 /** @deprecated Prefer authorizeRequest — kept for simple key checks. */
 export function hasAuth(req, config) {
   return authorizeRequest(req, config, {
diff --git a/src/server.mjs b/src/server.mjs
index cc2508b..5cac260 100644
--- a/src/server.mjs
+++ b/src/server.mjs
@@ -6,6 +6,7 @@ import { generateProviderVideo } from './video-providers.mjs';
 import http from 'node:http';
 import { readErrorDiagnostic, streamProviderError } from './protocol/upstream-error.mjs';
 import { fetchWithStreamProgress } from './protocol/stream-progress.mjs';
+import { applyOpenRouterProviderPolicy } from './protocol/openrouter-provider.mjs';
 import {
   appendFileSync,
   existsSync,
@@ -20,7 +21,9 @@ import path from 'node:path';
 import os from 'node:os';
 import { execFile } from 'node:child_process';
 import { promisify } from 'node:util';
-import { Agent as UndiciAgent } from 'undici';
+// Dispatcher and fetch must come from the same undici copy: the package's
+// v8 Agent speaks a dispatcher contract Node 22's internal undici v6 rejects.
+import { Agent as UndiciAgent, fetch as undiciFetch } from 'undici';
 import {
   backendIds,
   defaultBackendVariables,
@@ -44,6 +47,7 @@ import {
   selectedRecipeIdFromCommunityPlan
 } from './community-client.mjs';
 import { defaultLloomHome, loadConfig } from './config.mjs';
+import { createFleetProfileController } from './config-profiles.mjs';
 import { createDoctorReport } from './doctor.mjs';
 import { readHostMemory } from './host-memory.mjs';
 import { MACHINE_PROFILE_MEDIA_TYPE, profileMachine, rankRecipes, validateMachineProfile } from './machine-profile.mjs';
@@ -59,6 +63,8 @@ import { createModelMaintenanceController } from './model-maintenance-control.mj
 import { acquireRateLimitSlot, consumeRateBudget, createRateLimitRegistry } from './rate-limit.mjs';
 import { RuntimeManager, runtimeWatchdogConfig, normalizeRequestClass } from './runtime-manager.mjs';
 import { createPreferredResidencyReconciler } from './runtime-residency.mjs';
+import { createRuntimePreferenceController } from './runtime-preferences.mjs';
+import { createDashboardInstallation } from './dashboard-installation.mjs';
 import {
   applyRuntimePolicyPlan,
   createRuntimePolicyPlan,
@@ -101,7 +107,9 @@ import { ClusterCoordinator, currentNodeId, isFederatedGatewayBackend } from './
 const JSON_TYPE = 'application/json; charset=utf-8';
 const SSE_TYPE = 'text/event-stream; charset=utf-8';
 const execFileAsync = promisify(execFile);
-const longRunningMediaDispatcher = new UndiciAgent({
+// Tests may swap this per server instance via createLloomServer's
+// upstreamDispatcher option; production always uses the long-running Agent.
+let longRunningMediaDispatcher = new UndiciAgent({
   headersTimeout: 1800000,
   bodyTimeout: 1800000
 });
@@ -1385,10 +1393,10 @@ async function fetchUpstream({ backend, path, body, headers = {}, signal, dispat
         : 0;
     return await fetchWithStreamProgress(
       (progressSignal) =>
-        fetch(upstreamUrl(backend, path), {
+        undiciFetch(upstreamUrl(backend, path), {
           method: 'POST',
           headers: backendHeaders(backend, headers),
-          body: JSON.stringify(body),
+          body: JSON.stringify(path === '/v1/chat/completions' ? applyOpenRouterProviderPolicy(body, backend) : body),
           signal: progressSignal,
           dispatcher
         }),
@@ -1403,7 +1411,7 @@ async function fetchRawUpstream({ backend, path, body, headers = {}, signal, dis
   const timeoutMs = backend.timeoutMs ?? 1800000;
   const fetchSignal = upstreamSignal(signal, timeoutMs);
   try {
-    return await fetch(upstreamUrl(backend, path), {
+    return await undiciFetch(upstreamUrl(backend, path), {
       method: 'POST',
       headers: backendHeaders(backend, headers),
       body,
@@ -1682,6 +1690,7 @@ async function createLibraryPlan(config, searchParams) {
 
 function setupPlanOptionsFromQuery(searchParams) {
   return {
+    additive: queryBool(searchParams, ['additive'], false),
     recipeId: firstQueryParam(searchParams, ['recipe', 'recipe_id', 'recipe-id']),
     configPath: firstQueryParam(searchParams, ['config_out', 'config-out', 'config_path', 'config-path']),
     modelRoot: firstQueryParam(searchParams, ['model_root', 'model-root']),
@@ -1740,7 +1749,7 @@ function setupOptionsFromBody(config, body = {}) {
   return {
     recipeId: body.recipeId ?? body.recipe_id ?? body.recipe,
     ...communityOptionsFromBody(config, body),
-    configPath: body.configPath ?? body.config_path ?? body.configOut ?? body.config_out,
+    configPath: body.configPath ?? body.config_path ?? body.configOut ?? body.config_out ?? config.sourcePath,
     modelRoot: body.modelRoot ?? body.model_root,
     gatewayPort: body.gatewayPort ?? body.gateway_port ?? body.port,
     backendPortRange: body.backendPortRange ?? body.backend_port_range,
@@ -1933,7 +1942,12 @@ export async function retryRuntimeActionAfterConfigReload(action, getReloadInFli
   }
 }
 
-export function createLloomServer(config, { logger = console, runtimeManager = null, clusterCoordinator = null } = {}) {
+export function createLloomServer(
+  config,
+  { logger = console, runtimeManager = null, clusterCoordinator = null, upstreamDispatcher = null } = {}
+) {
+  // Tests install a composed mock here; production keeps the long-running Agent.
+  if (upstreamDispatcher) longRunningMediaDispatcher = upstreamDispatcher;
   const hostTelemetry = createHostTelemetry();
   const machineProfile = profileMachine().catch((error) => {
     logger.error?.(`Machine profile collection failed: ${error?.message ?? error}`);
@@ -2111,6 +2125,17 @@ export function createLloomServer(config, { logger = console, runtimeManager = n
     return retryRuntimeActionAfterConfigReload(action, () => reloadInFlight);
   }
 
+  const runtimePreferences = createRuntimePreferenceController({
+    getConfig: () => config,
+    mutateSource: (mutate) => mutateConfigSource(config, mutate),
+    reload: reloadConfig,
+    assertControl: (id, requester) => runtimeManager.assertRuntimeControl(id, requester),
+    onPersisted: (id, policy, generation) => runtimeManager.noteDesiredResidency(id, policy, generation),
+    onApplied: (id, policy, generation) => runtimeManager.settleDesiredResidency(id, policy, generation)
+  });
+  const dashboardInstallation = createDashboardInstallation({ getConfig: () => config, reload: reloadConfig });
+  const fleetProfiles = createFleetProfileController({ getConfig: () => config, reload: reloadConfig });
+
   async function routingStatus() {
     const cacheMs = Math.max(0, Number(config.cluster?.routingStatusCacheMs ?? 250));
     let runtimeStatus = routingStatusCache.value;
@@ -2576,25 +2601,32 @@ export function createLloomServer(config, { logger = console, runtimeManager = n
       if (!watchdogConfig.enabled || stream !== true) return;
       watchdogArmed = true;
       clearWatchdogTimer();
-      watchdogTimer = setTimeout(
-        () => {
-          watchdogTimer = null;
-          noteRuntimeRequestOutcome(resolved.model.runtime, {
-            id: connectionId,
-            route,
-            model: resolved.model.id,
-            status: 504,
-            ok: false,
-            durationMs: Date.now() - started,
-            runtimeDurationMs: runtimeStartedAt == null ? 0 : Date.now() - runtimeStartedAt,
-            stallDurationMs: Date.now() - lastProgressAt,
-            hadContent: watchdogHadContent,
-            responseBytes: 0,
-            stalled: true
-          });
-        },
-        watchdogHadContent ? watchdogConfig.idleContentTimeoutMs : watchdogConfig.firstContentTimeoutMs
-      );
+      const timeoutMs = watchdogHadContent ? watchdogConfig.idleContentTimeoutMs : watchdogConfig.firstContentTimeoutMs;
+      const onTimeout = () => {
+        watchdogTimer = null;
+        const stallDurationMs = Date.now() - lastProgressAt;
+        // A timer may wake slightly before the wall-clock deadline. Keep
+        // waiting rather than report a stall that classification must ignore.
+        if (stallDurationMs < timeoutMs) {
+          watchdogTimer = setTimeout(onTimeout, timeoutMs - stallDurationMs);
+          watchdogTimer.unref?.();
+          return;
+        }
+        noteRuntimeRequestOutcome(resolved.model.runtime, {
+          id: connectionId,
+          route,
+          model: resolved.model.id,
+          status: 504,
+          ok: false,
+          durationMs: Date.now() - started,
+          runtimeDurationMs: runtimeStartedAt == null ? 0 : Date.now() - runtimeStartedAt,
+          stallDurationMs,
+          hadContent: watchdogHadContent,
+          responseBytes: 0,
+          stalled: true
+        });
+      };
+      watchdogTimer = setTimeout(onTimeout, timeoutMs);
       watchdogTimer.unref?.();
     };
     const progress = (patch) => {
@@ -3711,7 +3743,7 @@ export function createLloomServer(config, { logger = console, runtimeManager = n
       }
 
       if (req.method === 'GET' && url.pathname === '/health') {
-        sendJson(res, 200, { ok: true, name: config.name ?? 'LLooM' }, {}, config);
+        sendJson(res, 200, { ok: true, name: config.name ?? 'LLooM', pid: process.pid }, {}, config);
         return;
       }
 
@@ -3721,7 +3753,17 @@ export function createLloomServer(config, { logger = console, runtimeManager = n
       }
 
       if (req.method === 'GET' && (url.pathname === '/' || url.pathname === '/gateway/dashboard')) {
-        sendHtml(res, 200, renderDashboardPage(), {}, config);
+        sendHtml(
+          res,
+          200,
+          renderDashboardPage(),
+          {
+            'x-frame-options': 'DENY',
+            'content-security-policy': "frame-ancestors 'none'; base-uri 'self'",
+            'referrer-policy': 'same-origin'
+          },
+          config
+        );
         return;
       }
 
@@ -3835,8 +3877,47 @@ export function createLloomServer(config, { logger = console, runtimeManager = n
         return;
       }
 
+      if (req.method === 'GET' && url.pathname === '/gateway/fleet/profiles') {
+        sendJson(res, 200, { ok: true, ...(await fleetProfiles.list()) });
+        return;
+      }
+      const fleetProfileMatch = url.pathname.match(/^\/gateway\/fleet\/profiles\/([^/]+)$/);
+      if (fleetProfileMatch) {
+        const name = decodeURIComponent(fleetProfileMatch[1]);
+        if (req.method === 'GET') {
+          sendJson(res, 200, { ok: true, ...(await fleetProfiles.plan(name)) });
+          return;
+        }
+        if (req.method === 'POST') {
+          const body = await readJson(req);
+          for (const key of Object.keys(body))
+            if (!['yes', 'description', 'overwrite'].includes(key))
+              throw Object.assign(new Error('Fleet profile accepts only yes, description, and overwrite.'), {
+                statusCode: 400
+              });
+          const isApply = url.searchParams.get('apply') === '1';
+          const result = isApply
+            ? await fleetProfiles.apply(name, { yes: body.yes })
+            : await fleetProfiles.save(name, {
+                yes: body.yes,
+                description: body.description,
+                overwrite: body.overwrite
+              });
+          if (isApply) {
+            reloadConfig();
+            await reloadInFlight;
+          }
+          sendJson(res, 200, { ok: true, ...result });
+          return;
+        }
+      }
+
       if (req.method === 'GET' && url.pathname === '/gateway/status') {
-        const runtimeStatus = await runtimeManager.status();
+        const runtimeStatus = await runtimeManager.status({
+          // Process sampling (ps, lsof, footprint probing) runs only when the
+          // memory scene asks for it; plain status polls skip the cost.
+          includeMemoryUsage: url.searchParams.get('memoryUsage') === '1'
+        });
         const clustered = Object.keys(config.cluster?.nodes ?? {}).length > 0;
         const localRuntimeStatus = clustered
           ? {
@@ -3861,7 +3942,9 @@ export function createLloomServer(config, { logger = console, runtimeManager = n
       if (req.method === 'GET' && url.pathname === '/gateway/node') {
         sendJson(res, 200, {
           ok: true,
-          node: await clusterCoordinator.localNodeStatus()
+          node: await clusterCoordinator.localNodeStatus({
+            includeMemoryUsage: url.searchParams.get('memoryUsage') === '1'
+          })
         });
         return;
       }
@@ -4127,7 +4210,33 @@ export function createLloomServer(config, { logger = console, runtimeManager = n
       }
 
       if (req.method === 'GET' && url.pathname === '/gateway/setup/plan') {
-        sendJson(res, 200, await createSetupPlan(config, setupPlanOptionsFromQuery(url.searchParams)));
+        const options = setupPlanOptionsFromQuery(url.searchParams);
+        sendJson(
+          res,
+          200,
+          await createSetupPlan(config, { ...options, configPath: options.configPath ?? config.sourcePath })
+        );
+        return;
+      }
+
+      if (url.pathname === '/gateway/installations' && req.method === 'GET') {
+        sendJson(res, 200, { job: dashboardInstallation.snapshot() });
+        return;
+      }
+      if (url.pathname === '/gateway/installations/plan' && req.method === 'POST') {
+        try {
+          sendJson(res, 200, await dashboardInstallation.review(await readJson(req)));
+        } catch (error) {
+          sendJson(res, error.statusCode || 400, errorBody(error.message, { code: 'installation_plan_failed' }));
+        }
+        return;
+      }
+      if (url.pathname === '/gateway/installations' && req.method === 'POST') {
+        try {
+          sendJson(res, 202, { job: dashboardInstallation.start(await readJson(req)) });
+        } catch (error) {
+          sendJson(res, error.statusCode || 400, errorBody(error.message, { code: 'installation_start_failed' }));
+        }
         return;
       }
 
@@ -4306,6 +4415,22 @@ export function createLloomServer(config, { logger = console, runtimeManager = n
         return;
       }
 
+      const residencyMatch = url.pathname.match(/^\/gateway\/runtimes\/([^/]+)\/residency$/);
+      if (req.method === 'GET' && residencyMatch) {
+        sendJson(res, 200, { ok: true, job: runtimePreferences.snapshot(decodeURIComponent(residencyMatch[1])) });
+        return;
+      }
+      if (req.method === 'POST' && residencyMatch) {
+        const body = await readJson(req);
+        const result = await runtimePreferences.request(decodeURIComponent(residencyMatch[1]), {
+          policy: body.policy,
+          yes: body.yes,
+          requestedBy: runtimeRequesterNode(req, config)
+        });
+        sendJson(res, 202, result);
+        return;
+      }
+
       const stopMatch = url.pathname.match(/^\/gateway\/runtimes\/([^/]+)\/stop$/);
       if (req.method === 'POST' && stopMatch) {
         sendJson(
diff --git a/src/setup.mjs b/src/setup.mjs
index fc149ce..62a978c 100644
--- a/src/setup.mjs
+++ b/src/setup.mjs
@@ -179,6 +179,7 @@ export async function applySetup(
     statePath,
     onProgress,
     stdio,
+    reviewedPlan,
     ...options
   } = {}
 ) {
@@ -186,11 +187,12 @@ export async function applySetup(
     throw new Error('Refusing to run setup without yes=true. Re-run with --yes after reviewing the dry-run plan.');
   }
 
-  const plan = await createSetupPlan(config, options);
+  const plan = reviewedPlan ?? (await createSetupPlan(config, options));
   if (dryRun) return plan;
 
   const init = await applyInit(config, {
     ...options,
+    reviewedPlan: reviewedPlan?.phases?.init,
     recipeId: plan.selectedRecipe.id,
     configPath: plan.configPath,
     modelRoot: plan.modelRoot,
@@ -211,6 +213,9 @@ export async function applySetup(
     recipesRoot: options.recipesRoot,
     recipeDocuments: options.recipeDocuments,
     backendCatalogPath: options.backendCatalogPath,
+    // Execute the exact reviewed bootstrap evidence that the user approved
+    // rather than re-profiling, re-selecting, or replanning.
+    reviewedPlan: plan.phases?.bootstrap,
     onProgress,
     stdio,
     ...(statePath ? { statePath } : {}),
diff --git a/test/admin-browser-security.test.mjs b/test/admin-browser-security.test.mjs
new file mode 100644
index 0000000..5092f46
--- /dev/null
+++ b/test/admin-browser-security.test.mjs
@@ -0,0 +1,42 @@
+import { test } from 'node:test';
+import assert from 'node:assert/strict';
+import { authorizeRequest } from '../src/security.mjs';
+const config = { server: { host: '127.0.0.1', port: 8100 }, security: { allowMissingAuth: true } };
+function auth(headers, method = 'POST', pathname = '/gateway/runtimes/chat/stop') {
+  return authorizeRequest({ method, headers: { host: '127.0.0.1:8100', ...headers } }, config, { method, pathname });
+}
+test('cross-site and DNS-rebound browser requests cannot read or mutate the loopback admin API', () => {
+  for (const headers of [
+    { origin: 'https://attacker.example' },
+    { origin: 'null' },
+    { host: 'rebound.example:8100' },
+    { host: '127.rebound.example:8100', origin: 'http://127.rebound.example:8100' },
+    { 'sec-fetch-site': 'cross-site' },
+    { origin: 'http://localhost:8100' },
+    { origin: 'http://127.0.0.1:8101' },
+    { origin: 'http://127.0.0.1:8100/path' },
+    { origin: 'https://127.0.0.1:8100' },
+    { host: '127.0.0.1:8100@attacker.example' }
+  ]) {
+    for (const method of ['GET', 'POST']) {
+      assert.equal(auth(headers, method).ok, false, JSON.stringify(headers));
+    }
+  }
+});
+test('same-origin UI and non-browser CLI retain the local management contract', () => {
+  assert.equal(auth({ origin: 'http://127.0.0.1:8100', 'sec-fetch-site': 'same-origin' }).ok, true);
+  assert.equal(auth({}).ok, true);
+  assert.equal(auth({ host: 'localhost:8100', origin: 'http://localhost:8100' }).ok, true);
+});
+test('public telemetry is also protected against cross-site reads while inference compatibility remains', () => {
+  const req = { method: 'GET', headers: { host: '127.0.0.1:8100', origin: 'https://attacker.example' } };
+  assert.equal(
+    authorizeRequest(
+      req,
+      { ...config, security: { ...config.security, publicTelemetry: true } },
+      { pathname: '/gateway/models' }
+    ).ok,
+    false
+  );
+  assert.equal(auth({ origin: 'https://a-client.example' }, 'POST', '/v1/chat/completions').ok, true);
+});
diff --git a/test/cluster.test.mjs b/test/cluster.test.mjs
index 7433990..9c6c065 100644
--- a/test/cluster.test.mjs
+++ b/test/cluster.test.mjs
@@ -5,7 +5,10 @@ import {
   ClusterCoordinator,
   federatedNodeConfigFromSnapshot,
   materializeFederatedNodes,
+  mergeNvidiaSyncClusterDiscovery,
   modelTargets,
+  nvidiaSyncClusterConfig,
+  nvidiaSyncDiscoverySummary,
   parseNvidiaSyncSshConfig,
   runtimeAuthority,
   runtimeControlAllowed,
@@ -22,6 +25,9 @@ import {
 } from '../src/runtime-manager.mjs';
 import { createRegistry } from '../src/registry.mjs';
 import { syncClusterSetupMembers } from '../src/setup.mjs';
+import { readFileSync } from 'node:fs';
+import path from 'node:path';
+import { fileURLToPath } from 'node:url';
 
 const syncPeers = parseNvidiaSyncSshConfig(`
 Host ennspark02-lan
@@ -32,16 +38,28 @@ Host ennspark02-lan
 Host ignored
   Hostname 192.168.1.20
 `);
-assert.deepEqual(syncPeers, [{ alias: 'ennspark02-lan', createdBySync: true, hostname: '10.100.20.2', user: 'spark' }]);
+// Legacy marked entries without `NVSyncClusterAlias` are still recognized by
+// their stable non-IP Host alias; the unmarked block is reported but ignored.
+assert.deepEqual(syncPeers, [
+  { alias: 'ennspark02-lan', createdBySync: true, hostname: '10.100.20.2', user: 'spark' },
+  { alias: 'ignored', createdBySync: false, hostname: '192.168.1.20' }
+]);
 
 const discovery = buildNvidiaSyncDiscovery({
   peers: syncPeers,
-  localAddresses: [{ interface: 'enp1s0', address: '10.100.20.1', prefixlen: 24 }],
+  localAddresses: [
+    { interface: 'enp1s0', address: '10.100.20.1', prefixlen: 24 },
+    { interface: 'enp1s1', address: '10.100.21.9', prefixlen: 24 }
+  ],
   localNodeId: 'ennspark01',
   hostname: 'spark-host'
 });
 assert.equal(discovery.provider, 'nvidia-sync');
 assert.equal(discovery.nodes.ennspark02.backendHost, '10.100.20.2');
+assert.equal(discovery.nodes.ennspark02.fabricInterface, undefined);
+assert.equal(discovery.links.length, 1);
+assert.equal(discovery.links[0].localInterface, 'enp1s0');
+assert.equal(discovery.links[0].interface, undefined);
 
 const heterogeneous = {
   cluster: {
@@ -947,4 +965,427 @@ assert.equal(
   true
 );
 
+// ---------------------------------------------------------------------------
+// NVIDIA Sync discovery: real three-node SSH fixture.
+// ---------------------------------------------------------------------------
+
+const fixturePath = path.join(path.dirname(fileURLToPath(import.meta.url)), 'fixtures', 'nvidia-sync-three-node.json');
+const fixture = JSON.parse(readFileSync(fixturePath, 'utf8'));
+const normalizeFixtureAddresses = (value) =>
+  value.flatMap((device) =>
+    (device.addr_info ?? [])
+      .filter((address) => address.family === 'inet' && address.scope !== 'host')
+      .map((address) => ({
+        interface: device.ifname,
+        address: address.local,
+        prefixlen: Number(address.prefixlen ?? 32)
+      }))
+  );
+
+const fixturePeers = parseNvidiaSyncSshConfig(fixture.sshConfig);
+assert.deepEqual(
+  fixturePeers.map((peer) => [peer.alias, peer.createdBySync, peer.clusterAlias ?? null, peer.hostname ?? null]),
+  [
+    ['10.100.16.1', false, null, null],
+    ['worker-fabric', false, null, '10.100.17.1'],
+    ['ennspark02-lan', true, 'ennspark02-lan', '10.100.192.1'],
+    ['10.100.192.1', true, 'ennspark02-lan', '10.100.192.1'],
+    ['10.100.193.1', true, 'ennspark02-lan', '10.100.193.1'],
+    ['ennspark03-lan', true, 'ennspark03-lan', '10.100.196.2'],
+    ['10.100.196.2', true, 'ennspark03-lan', '10.100.196.2'],
+    ['10.100.197.2', true, 'ennspark03-lan', '10.100.197.2']
+  ]
+);
+
+const fixtureLocalAddresses = normalizeFixtureAddresses(fixture.ipAddresses);
+const fixtureDiscovery = buildNvidiaSyncDiscovery({
+  peers: fixturePeers,
+  localAddresses: fixtureLocalAddresses,
+  localNodeId: fixture.nodeId,
+  hostname: fixture.hostname
+});
+assert.equal(fixtureDiscovery.provider, 'nvidia-sync');
+assert.equal(fixtureDiscovery.detected, true);
+assert.equal(fixtureDiscovery.nodeId, 'ennspark01');
+assert.deepEqual(Object.keys(fixtureDiscovery.nodes).sort(), ['ennspark01', 'ennspark02', 'ennspark03']);
+assert.equal(fixtureDiscovery.nodeCount, 3);
+assert.equal(fixtureDiscovery.verified, false);
+// Four distinct observed rails: two NICs per peer, with the duplicate
+// alias/address Sync blocks collapsed.
+assert.equal(fixtureDiscovery.links.length, 4);
+assert.deepEqual(
+  fixtureDiscovery.links.map((link) => link.peerAddress),
+  ['10.100.192.1', '10.100.193.1', '10.100.196.2', '10.100.197.2']
+);
+assert.deepEqual(
+  fixtureDiscovery.links.map((link) => link.localInterface),
+  ['enp1s0f0np0', 'enP2p1s0f0np0', 'enp1s0f1np1', 'enP2p1s0f1np1']
+);
+assert.equal(fixtureDiscovery.nodes.ennspark02.backendHost, '10.100.192.1');
+assert.equal(fixtureDiscovery.nodes.ennspark02.sshAlias, 'ennspark02-lan');
+assert.equal(fixtureDiscovery.nodes.ennspark03.backendHost, '10.100.196.2');
+assert.equal(fixtureDiscovery.nodes.ennspark03.sshAlias, 'ennspark03-lan');
+assert.equal(fixtureDiscovery.nodes.ennspark01.local, true);
+// Remote interface names are never asserted; only the local node has one.
+assert.equal(fixtureDiscovery.nodes.ennspark02.fabricInterface, undefined);
+assert.equal(fixtureDiscovery.nodes.ennspark03.fabricInterface, undefined);
+assert.equal(fixtureDiscovery.nodes.ennspark01.fabricInterface, 'enp1s0f0np0');
+// Discovery observes adjacency only and must not elect a leader.
+assert.equal(fixtureDiscovery.leaderNode, undefined);
+assert.equal(fixtureDiscovery.localNode, 'ennspark01');
+assert.equal(fixtureDiscovery.ambiguousPeers, undefined);
+assert.equal(fixtureDiscovery.topology, 'local-adjacency');
+
+// Permuting the SSH entries and the local address order must not change output.
+const shuffle = (list) => [...list].reverse();
+for (const peers of [shuffle(fixturePeers), [...fixturePeers].reverse()]) {
+  for (const addresses of [shuffle(fixtureLocalAddresses), [...fixtureLocalAddresses].reverse()]) {
+    const permuted = buildNvidiaSyncDiscovery({
+      peers,
+      localAddresses: addresses,
+      localNodeId: fixture.nodeId,
+      hostname: fixture.hostname
+    });
+    assert.deepEqual(permuted, fixtureDiscovery);
+    assert.deepEqual(Object.keys(permuted.nodes).sort(), Object.keys(fixtureDiscovery.nodes).sort());
+    assert.equal(permuted.nodes.ennspark02.backendHost, '10.100.192.1');
+    assert.equal(permuted.nodes.ennspark03.backendHost, '10.100.196.2');
+  }
+}
+
+const plannedFixtureCluster = nvidiaSyncClusterConfig(fixtureDiscovery, { id: 'ennspark-cluster' });
+assert.equal(plannedFixtureCluster.nodeId, 'ennspark01');
+assert.equal(plannedFixtureCluster.leaderNode, undefined);
+assert.equal(plannedFixtureCluster.nodes.ennspark01.endpoint, undefined);
+assert.equal(plannedFixtureCluster.nodes.ennspark02, undefined);
+assert.equal(plannedFixtureCluster.discovery.nodes.ennspark02.rails[0].localInterface, 'enp1s0f0np0');
+assert.equal(plannedFixtureCluster.discovery.links[0].interface, undefined);
+
+// ---------------------------------------------------------------------------
+// NVIDIA Sync discovery: legacy two-node config, malformed boundaries, refusals.
+// ---------------------------------------------------------------------------
+
+const legacyPeers = parseNvidiaSyncSshConfig(`
+Host 10.100.20.1
+  User spark
+  ### CreatedBy: NVIDIA Sync
+  ### NVSyncClusterAlias: ennspark01-lan
+  Hostname 10.100.20.1
+  User enntitysparkadmin
+`);
+assert.equal(legacyPeers.length, 1);
+assert.equal(legacyPeers[0].createdBySync, true);
+const legacyDiscovery = buildNvidiaSyncDiscovery({
+  peers: legacyPeers,
+  localAddresses: [{ interface: 'enp1s0f0np0', address: '10.100.20.2', prefixlen: 24 }],
+  localNodeId: 'ennspark02',
+  hostname: 'spark-2'
+});
+assert.deepEqual(Object.keys(legacyDiscovery.nodes).sort(), ['ennspark01', 'ennspark02']);
+assert.equal(legacyDiscovery.topology, 'direct');
+
+// Malformed SSH: wildcard/negated/multi-host Host patterns and Match blocks
+// close the previous entry instead of leaking ClusterAlias provenance forward.
+const malformedPeers = parseNvidiaSyncSshConfig(`
+Host ennspark02-lan 10.100.192.9
+  ### CreatedBy: NVIDIA Sync
+  Hostname 10.100.192.1
+Host !bad
+  ### CreatedBy: NVIDIA Sync
+  ### NVSyncClusterAlias: ennspark03-lan
+  Hostname 10.100.196.2
+Host *
+  ### CreatedBy: NVIDIA Sync
+  ### NVSyncClusterAlias: ennspark04-lan
+  Hostname 10.100.199.2
+Match host ennspark05
+  ### CreatedBy: NVIDIA Sync
+  ### NVSyncClusterAlias: ennspark05-lan
+  Hostname 10.100.200.2
+`);
+assert.deepEqual(
+  malformedPeers.map((peer) => [peer.alias, peer.createdBySync, peer.clusterAlias ?? null]),
+  []
+);
+// Conditional scopes contribute no peers.
+for (const peer of malformedPeers) assert.equal(peer.createdBySync, false);
+
+// Comments before `Host` do not become Host provenance, and the marker is only
+// honored inside the block it follows.
+assert.deepEqual(parseNvidiaSyncSshConfig('### CreatedBy: NVIDIA Sync\n### NVSyncClusterAlias: ghost\n'), []);
+
+// A Sync alias whose hostname is one of this node's own addresses is refused.
+const selfPeerDiscovery = buildNvidiaSyncDiscovery({
+  peers: [
+    {
+      alias: 'ennspark02-lan',
+      createdBySync: true,
+      clusterAlias: 'ennspark02-lan',
+      hostname: '10.100.192.2',
+      user: 'spark'
+    }
+  ],
+  localAddresses: [{ interface: 'enp1s0f0np0', address: '10.100.192.2', prefixlen: 24 }],
+  localNodeId: 'ennspark01'
+});
+assert.equal(selfPeerDiscovery, null);
+
+// An address matching two local interfaces is ambiguous: refuse instead of
+// arbitrarily picking the first rail.
+const ambiguousDiscovery = buildNvidiaSyncDiscovery({
+  peers: [
+    {
+      alias: 'ennspark02-lan',
+      createdBySync: true,
+      clusterAlias: 'ennspark02-lan',
+      hostname: '10.100.192.1',
+      user: 'spark'
+    },
+    {
+      alias: '10.100.196.2',
+      createdBySync: true,
+      clusterAlias: 'ennspark03-lan',
+      hostname: '10.100.196.2',
+      user: 'spark'
+    }
+  ],
+  localAddresses: [
+    { interface: 'enp1s0f0np0', address: '10.100.192.2', prefixlen: 24 },
+    { interface: 'enP2p1s0f0np0', address: '10.100.192.3', prefixlen: 24 },
+    { interface: 'enp1s0f1np1', address: '10.100.196.1', prefixlen: 24 }
+  ],
+  localNodeId: 'ennspark01'
+});
+assert.deepEqual(Object.keys(ambiguousDiscovery.nodes).sort(), ['ennspark01', 'ennspark03']);
+assert.deepEqual(
+  ambiguousDiscovery.ambiguousPeers.map((entry) => [entry.clusterAlias, entry.reason]),
+  [['ennspark02-lan', 'ambiguous-local-interface']]
+);
+
+// ---------------------------------------------------------------------------
+// Merge: preserve every existing node/endpoint/catalog/leader; migrate only the
+// unambiguous old-fabric endpoint; never duplicate the local hostname.
+// ---------------------------------------------------------------------------
+
+const legacyClusterConfig = {
+  cluster: {
+    id: 'spark-cluster',
+    nodeId: 'ennspark01',
+    leaderNode: 'ennspark01',
+    apiKeyEnv: 'LLOOM_ADMIN_API_KEY',
+    nodes: {
+      ennspark01: {
+        name: 'spark-8333',
+        endpoint: 'http://10.100.16.1:8100',
+        backendHost: '10.100.16.1',
+        fabricInterface: 'enp1s0f0np0',
+        apiKeyEnv: 'LOCAL_KEY',
+        labels: { role: 'leader', hardware: 'dgx-spark' },
+        resources: { memoryGb: 128, maxMemoryUtilization: 0.9 }
+      },
+      'macbook-local': {
+        name: 'MacBook Local',
+        endpoint: 'http://macbook-private:9001',
+        apiKeyEnv: 'LAB_KEY',
+        labels: { role: 'node', architecture: 'darwin-arm64', accelerator: 'metal' },
+        resources: { memoryGb: 64 },
+        proxy: {
+          models: [{ id: 'local/qwen', as: 'macbook-local/local/qwen', kind: 'chat', remoteRuntime: 'mac-qwen' }]
+        }
+      }
+    }
+  },
+  runtimes: { qwen: { enabled: true, node: 'ennspark01', memoryGb: 54 } },
+  backends: { qwen: { type: 'vllm', baseUrl: 'http://10.100.16.1:8201/v1' } },
+  models: [
+    {
+      id: 'example/Qwen',
+      backend: 'qwen',
+      runtime: 'qwen',
+      targets: [{ id: 'ennspark01', node: 'ennspark01', backend: 'qwen', runtime: 'qwen' }]
+    }
+  ]
+};
+
+const merged = mergeNvidiaSyncClusterDiscovery({
+  discovery: fixtureDiscovery,
+  cluster: legacyClusterConfig.cluster,
+  config: legacyClusterConfig
+});
+assert.equal(merged.id, 'spark-cluster');
+assert.equal(merged.nodeId, 'ennspark01');
+assert.equal(merged.leaderNode, 'ennspark01');
+assert.equal(merged.apiKeyEnv, 'LLOOM_ADMIN_API_KEY');
+assert.deepEqual(Object.keys(merged.nodes).sort(), ['ennspark01', 'macbook-local']);
+
+// An endpoint on the old fabric host migrates with that host.
+assert.equal(merged.nodes.ennspark01.name, 'spark-8333');
+assert.equal(merged.nodes.ennspark01.endpoint, 'http://10.100.192.2:8100');
+assert.equal(merged.nodes.ennspark01.apiKeyEnv, 'LOCAL_KEY');
+assert.equal(merged.nodes.ennspark01.resources.maxMemoryUtilization, 0.9);
+
+// Untouched foreign node keeps endpoint, port, catalog, credential env, labels.
+assert.equal(merged.nodes['macbook-local'].endpoint, 'http://macbook-private:9001');
+assert.equal(merged.nodes['macbook-local'].apiKeyEnv, 'LAB_KEY');
+assert.deepEqual(
+  merged.nodes['macbook-local'].proxy.models,
+  legacyClusterConfig.cluster.nodes['macbook-local'].proxy.models
+);
+assert.equal(merged.nodes['macbook-local'].labels.architecture, 'darwin-arm64');
+
+// Newly observed peers remain inventory until authenticated federation join.
+assert.equal(merged.nodes.ennspark02, undefined);
+assert.equal(merged.discovery.nodes.ennspark02.backendHost, '10.100.192.1');
+assert.equal(merged.discovery.nodes.ennspark02.fabricInterface, undefined);
+assert.equal(merged.discovery.nodes.ennspark03.backendHost, '10.100.196.2');
+assert.deepEqual(
+  merged.discovery.links.map((link) => [link.peerNodeId, link.localInterface]),
+  [
+    ['ennspark02', 'enp1s0f0np0'],
+    ['ennspark02', 'enP2p1s0f0np0'],
+    ['ennspark03', 'enp1s0f1np1'],
+    ['ennspark03', 'enP2p1s0f1np1']
+  ]
+);
+assert(Array.isArray(merged.diagnostics));
+assert(
+  merged.diagnostics.every((entry) => typeof entry === 'string' || (entry && typeof entry.message === 'string')),
+  'merge diagnostics must be actionable strings or objects with a message'
+);
+
+// Sub-migration: an old-fabric endpoint is rewritten only when its URL hostname
+// is exactly the previous backendHost of that node.
+const migrateConfig = {
+  cluster: {
+    nodeId: 'ennspark01',
+    leaderNode: 'ennspark01',
+    nodes: {
+      ennspark01: { endpoint: 'http://10.100.16.1:8100', backendHost: '10.100.16.1' },
+      ennspark02: { endpoint: 'http://10.100.16.2:9100/v1', backendHost: '10.100.16.2' }
+    }
+  }
+};
+const migrated = mergeNvidiaSyncClusterDiscovery({
+  discovery: fixtureDiscovery,
+  cluster: migrateConfig.cluster,
+  config: migrateConfig
+});
+// Preserve custom port and path when migrating the exact old hostname.
+assert.equal(migrated.nodes.ennspark02.endpoint, 'http://10.100.192.1:9100/v1');
+
+const ambiguousEndpointConfig = {
+  cluster: {
+    nodeId: 'ennspark01',
+    leaderNode: 'ennspark01',
+    nodes: {
+      ennspark01: { endpoint: 'http://10.100.16.1:8100', backendHost: '10.100.16.1' },
+      // Substring lookalike: 10.100.16.10 CONTAINS 10.100.16.1 but is not it.
+      ennspark02: { endpoint: 'http://10.100.16.10:8100', backendHost: '10.100.16.1' }
+    }
+  }
+};
+const ambiguousMerged = mergeNvidiaSyncClusterDiscovery({
+  discovery: fixtureDiscovery,
+  cluster: ambiguousEndpointConfig.cluster,
+  config: ambiguousEndpointConfig
+});
+assert.equal(ambiguousMerged.nodes.ennspark02.endpoint, 'http://10.100.16.10:8100');
+assert(
+  ambiguousMerged.diagnostics.some(
+    (entry) => typeof entry === 'object' && entry.level === 'migration-required' && entry.nodeId === 'ennspark02'
+  ),
+  'a preserved custom endpoint must be reported as migration-required'
+);
+assert.deepEqual(validateClusterConfig({ ...legacyClusterConfig, cluster: merged }), []);
+
+const fixtureSummary = nvidiaSyncDiscoverySummary(fixtureDiscovery, { currentConfig: legacyClusterConfig });
+assert.equal(fixtureSummary.observedLinks, 4);
+assert.deepEqual(fixtureSummary.discoveredNodes, ['ennspark01', 'ennspark02', 'ennspark03']);
+assert.deepEqual(fixtureSummary.newNodes, ['ennspark02', 'ennspark03']);
+assert.equal(fixtureSummary.leaderNode, 'ennspark01');
+assert(Array.isArray(fixtureSummary.diagnostics));
+
+// ---------------------------------------------------------------------------
+// Explicit runtimePlacement subsets remain honored for single/TP2/TP3.
+// ---------------------------------------------------------------------------
+
+for (const [count, expected] of [
+  [1, ['ennspark01']],
+  [2, ['ennspark01', 'ennspark02']],
+  [3, ['ennspark01', 'ennspark02', 'ennspark03']]
+]) {
+  const rendered = validateClusterConfig({ ...legacyClusterConfig, cluster: merged }).length === 0 ? '' : null;
+  assert.equal(rendered, '');
+  const nodeIds = ['ennspark01', 'ennspark02', 'ennspark03'].slice(0, count);
+  const placement = runtimePlacement(
+    { placement: { mode: 'replicated', nodes: nodeIds } },
+    { cluster: merged },
+    {
+      LLOOM_NODE_ID: 'ennspark01'
+    }
+  );
+  assert.deepEqual(placement.mode, 'replicated');
+  const members = nodeIds.map((node, index) => ({ node, runtime: `tp${count}-${index}`, order: index }));
+  const distributed = runtimePlacement(
+    {
+      placement: {
+        mode: 'distributed',
+        members: members.map((member) => ({
+          node: member.node,
+          runtime: member.runtime,
+          role: count === 1 ? 'head' : 'rank',
+          order: member.order
+        }))
+      }
+    },
+    { cluster: merged },
+    { LLOOM_NODE_ID: 'ennspark01' }
+  );
+  assert.deepEqual(distributed.nodes, expected);
+  assert.equal(distributed.members.length, count);
+}
+
+const sourceWithSettings = {
+  nodeId: '${LOCAL_NODE}',
+  leaderNode: 'ennspark02',
+  topology: 'ring',
+  links: [{ manual: true }],
+  customFlag: true,
+  nodes: {
+    ...legacyClusterConfig.cluster.nodes,
+    ennspark03: { backendHost: '100.78.22.9', endpoint: 'http://100.78.22.9:8100', apiKeyEnv: '${PEER_KEY}' }
+  }
+};
+const settingsMerged = mergeNvidiaSyncClusterDiscovery({
+  discovery: fixtureDiscovery,
+  cluster: sourceWithSettings,
+  localNodeId: 'ennspark01'
+});
+assert.equal(settingsMerged.nodeId, '${LOCAL_NODE}');
+assert.equal(settingsMerged.leaderNode, 'ennspark02');
+assert.equal(settingsMerged.topology, 'ring');
+assert.equal(settingsMerged.apiKeyEnv, undefined);
+assert.equal(settingsMerged.customFlag, true);
+assert.deepEqual(settingsMerged.links, sourceWithSettings.links);
+assert.equal(settingsMerged.nodes.ennspark03.backendHost, '100.78.22.9');
+assert.equal(settingsMerged.nodes.ennspark03.endpoint, 'http://100.78.22.9:8100');
+assert.equal(settingsMerged.nodes.ennspark03.fabricAddress, '10.100.196.2');
+assert.equal(settingsMerged.nodes.ennspark03.apiKeyEnv, '${PEER_KEY}');
+assert(!Object.hasOwn(settingsMerged.nodes, '${LOCAL_NODE}'));
+const loopbackCustom = structuredClone(migrateConfig.cluster);
+loopbackCustom.nodes.ennspark01.endpoint = 'http://127.0.0.1:8100';
+assert.equal(
+  mergeNvidiaSyncClusterDiscovery({ discovery: fixtureDiscovery, cluster: loopbackCustom }).nodes.ennspark01
+    .backendHost,
+  '10.100.16.1'
+);
+assert.throws(
+  () =>
+    mergeNvidiaSyncClusterDiscovery({
+      discovery: fixtureDiscovery,
+      cluster: legacyClusterConfig.cluster,
+      localNodeId: 'wrong-node'
+    }),
+  /identity conflicts/
+);
 console.log('cluster tests passed');
diff --git a/test/config-profiles.test.mjs b/test/config-profiles.test.mjs
new file mode 100644
index 0000000..be897f1
--- /dev/null
+++ b/test/config-profiles.test.mjs
@@ -0,0 +1,163 @@
+import assert from 'node:assert/strict';
+import fs from 'node:fs/promises';
+import os from 'node:os';
+import path from 'node:path';
+import { loadConfig } from '../src/config.mjs';
+import {
+  composeProfile,
+  createFleetProfileController,
+  listProfiles,
+  normalizeProfileDocument,
+  planProfileChanges,
+  readProfile
+} from '../src/config-profiles.mjs';
+
+const directory = await fs.mkdtemp(path.join(os.tmpdir(), 'lloom-fleet-profiles-'));
+const configPath = path.join(directory, 'config.json');
+const source = {
+  server: { host: '127.0.0.1', port: 8100 },
+  security: { allowMissingAuth: true },
+  defaults: { chatModel: 'local-model' },
+  backends: {
+    local: { type: 'openai', baseUrl: 'http://127.0.0.1:8201/v1' },
+    cloud: {
+      type: 'openai',
+      baseUrl: 'https://openrouter.ai/api/v1',
+      apiKeyEnv: 'OPENROUTER_API_KEY'
+    }
+  },
+  models: [
+    { id: 'local-model', kind: 'chat', backend: 'local', upstreamModel: 'local-model' },
+    { id: 'cloud-model', kind: 'chat', backend: 'cloud', upstreamModel: 'cloud-model' }
+  ],
+  runtimes: {
+    'local-model': { command: 'vllm', port: 8201, enabled: true, managed: true, keepWarm: true }
+  },
+  aliases: {
+    omp: {
+      members: ['local-model', 'cloud-model'],
+      activeRoute: 'local-first',
+      routeProfiles: {
+        'local-first': { members: ['local-model', 'cloud-model'] },
+        cloud: { members: ['cloud-model'] }
+      }
+    },
+    simple: { members: ['local-model'] }
+  }
+};
+
+const writeProfileFile = async (name, doc) => {
+  const dir = path.join(directory, 'profiles');
+  await fs.mkdir(dir, { recursive: true });
+  await fs.writeFile(path.join(dir, name + '.json'), `${JSON.stringify(doc, null, 2)}\n`);
+};
+
+try {
+  await fs.writeFile(configPath, `${JSON.stringify(source, null, 2)}\n`, { mode: 0o600 });
+  const config = await loadConfig(configPath, { env: { ...process.env, OPENROUTER_API_KEY: 'test' } });
+
+  // -- document normalization ------------------------------------------------
+  assert.deepEqual(normalizeProfileDocument({ routes: { omp: 'cloud' } }, 'p'), {
+    name: 'p',
+    description: '',
+    routes: { omp: 'cloud' },
+    residency: {},
+    defaults: null
+  });
+  assert.throws(() => normalizeProfileDocument({ nope: true }, 'p'), /unsupported sections/);
+  assert.throws(() => normalizeProfileDocument({ residency: { x: 'sometimes' } }, 'p'), /always, preferred, or auto/);
+
+  // -- planning ---------------------------------------------------------------
+  const localProfile = normalizeProfileDocument(
+    { routes: { omp: 'local-first', simple: 'local-model' }, residency: { 'local-model': 'preferred' } },
+    'local'
+  );
+  const cloudProfile = normalizeProfileDocument(
+    { routes: { omp: 'cloud' }, residency: { 'local-model': 'auto' }, defaults: { chatModel: 'cloud-model' } },
+    'cloud'
+  );
+  const localPlan = planProfileChanges(config, localProfile);
+  assert.equal(localPlan.unchanged.length, 2, 'omp route + residency are no-ops; simple pins members (members rewrite)');
+  const cloudPlan = planProfileChanges(config, cloudProfile);
+  assert.equal(cloudPlan.routes.length, 1);
+  assert.deepEqual(cloudPlan.routes[0], {
+    id: 'omp',
+    from: 'local-first',
+    to: 'cloud',
+    activeRoute: 'cloud',
+    members: ['cloud-model'],
+    optionalMembers: []
+  });
+  assert.deepEqual(cloudPlan.residency, [{ id: 'local-model', from: 'always', to: 'auto' }]);
+  assert.deepEqual(cloudPlan.defaults, [{ id: 'chatModel', from: 'local-model', to: 'cloud-model' }]);
+  assert.throws(() => planProfileChanges(config, normalizeProfileDocument({ routes: { ghost: 'cloud' } }, 'x')), /no alias ghost/);
+  assert.throws(
+    () => planProfileChanges(config, normalizeProfileDocument({ residency: { ghost: 'auto' } }, 'x')),
+    /no runtime ghost/
+  );
+
+  // -- compose ------------------------------------------------------------------
+  const composed = composeProfile(structuredClone(source), cloudProfile, 'cloud');
+  assert.equal(composed.fleet.activeProfile, 'cloud');
+  assert.deepEqual(composed.aliases.omp.members, ['cloud-model']);
+  assert.equal(composed.aliases.omp.activeRoute, 'cloud');
+  assert.equal(composed.runtimes['local-model'].keepWarm, false);
+  assert.equal(composed.runtimes['local-model'].preferredWarm, false);
+  assert.equal(composed.defaults.chatModel, 'cloud-model');
+
+  // Composed output must survive real config validation (validate-all-then-apply
+  // relies on loadConfig accepting the staged candidate).
+  const staged = path.join(directory, 'staged.json');
+  await fs.writeFile(staged, `${JSON.stringify(composed, null, 2)}\n`);
+  const reloaded = await loadConfig(staged, { env: { ...process.env, OPENROUTER_API_KEY: 'test' } });
+  assert.equal(reloaded.aliases.omp.activeRoute, 'cloud');
+
+  // -- controller: apply flips the file atomically and hot-marks active ----------
+  const controller = createFleetProfileController({
+    getConfig: () => config,
+    reload: () => {},
+    env: { ...process.env, OPENROUTER_API_KEY: 'test' }
+  });
+  await writeProfileFile('cloud', cloudProfile);
+  await writeProfileFile('local', localProfile);
+
+  await assert.rejects(controller.apply('cloud', { yes: false }), /confirm with yes/);
+  await assert.rejects(controller.apply('missing', { yes: true }), /ENOENT|no such file/i);
+
+  const applied = await controller.apply('cloud', { yes: true });
+  assert.equal(applied.profile, 'cloud');
+  assert.equal(applied.routes.length, 1);
+  assert.equal(applied.unchanged.length, 0, 'second plan sees no changes after apply? no: apply recomputes against raw');
+  const onDisk = JSON.parse(await fs.readFile(configPath, 'utf8'));
+  assert.equal(onDisk.fleet.activeProfile, 'cloud');
+  assert.deepEqual(onDisk.aliases.omp.members, ['cloud-model']);
+
+  // Listing reflects the active marker.
+  const listing = await listProfiles(config);
+  assert.equal(listing.active, 'cloud');
+  const cloudEntry = listing.profiles.find((profile) => profile.name === 'cloud');
+  assert.equal(cloudEntry.active, true);
+
+  // Applying the equal-content local profile is a clean no-op write.
+  const rerun = await controller.apply('local', { yes: true });
+  assert.equal(rerun.profile, 'local');
+
+  // -- save captures the live config --------------------------------------------
+  const saved = await controller.save('snapshot', { yes: true, description: 'point in time' });
+  assert.equal(saved.profile, 'snapshot');
+  const snapshotDoc = JSON.parse(await fs.readFile(saved.file, 'utf8'));
+  assert.equal(snapshotDoc.routes.omp, 'local-first', 'save records the active route profile');
+  assert.equal(snapshotDoc.residency['local-model'], 'preferred', 'the applied local profile set preferred');
+  await assert.rejects(controller.save('snapshot', { yes: true }), /already exists/);
+  await controller.save('snapshot', { yes: true, overwrite: true });
+
+  // readProfile normalizes documents for display.
+  const shown = await readProfile(config, 'snapshot');
+  assert.equal(shown.name, 'snapshot');
+  await assert.rejects(() => readProfile(config, '../escape'), /Invalid profile name/);
+  await assert.rejects(() => readProfile(config, 'missing'), /ENOENT/);
+
+  console.log('fleet profile tests passed');
+} finally {
+  await fs.rm(directory, { recursive: true, force: true });
+}
diff --git a/test/dashboard-memory.test.mjs b/test/dashboard-memory.test.mjs
new file mode 100644
index 0000000..5e14692
--- /dev/null
+++ b/test/dashboard-memory.test.mjs
@@ -0,0 +1,401 @@
+import assert from 'node:assert/strict';
+import { test } from 'node:test';
+import { buildMemoryMap } from '../src/dashboard-memory.mjs';
+
+const GiB = 1024 ** 3;
+const policy = { mode: 'enforce', minAvailableMemoryGb: 12, maxMemoryUtilization: 0.9 };
+
+function node({ total = 96 * GiB, available = 68 * GiB, reachable = true } = {}) {
+  return {
+    id: 'node-a',
+    local: true,
+    reachable,
+    telemetry: {
+      memory: {
+        totalBytes: total,
+        availableBytes: available,
+        usedBytes: total - available
+      }
+    }
+  };
+}
+
+function map(overrides = {}) {
+  const { nodeOverrides, ...rest } = overrides;
+  return buildMemoryMap({
+    node: node(nodeOverrides),
+    runtimes: {},
+    models: [],
+    memorySafety: policy,
+    ...rest
+  });
+}
+
+test('builds an exact 96 GiB local map with reserve and forecast', () => {
+  const result = map({
+    runtimes: {
+      main: {
+        id: 'main',
+        status: 'running',
+        healthy: true,
+        memoryUsage: {
+          residentBytes: 20 * GiB,
+          groupId: 'proc:1',
+          sharedRuntimeIds: ['main'],
+          loadedModelIds: ['upstream-a'],
+          residencyKnown: true
+        }
+      }
+    },
+    models: [
+      { id: 'model-a', name: 'Model A', runtime: 'main', upstreamModel: 'upstream-a', targets: [{ node: 'node-a' }] }
+    ],
+    previewModelId: 'model-a'
+  });
+
+  assert.equal(result.known, true);
+  assert.equal(result.totalBytes, 96 * GiB);
+  assert.equal(result.usedBytes, 28 * GiB);
+  assert.equal(result.availableBytes, 68 * GiB);
+  assert.equal(result.reserveBytes, 12 * GiB);
+  assert.equal(result.usableBytes, 56 * GiB);
+  assert.equal(
+    result.segments.reduce((sum, segment) => sum + segment.bytes, 0),
+    96 * GiB
+  );
+  assert.deepEqual(
+    result.segments.map((segment) => segment.kind),
+    ['runtime', 'system', 'available']
+  );
+  assert.equal(result.preview.status, 'resident');
+  assert.equal(result.preview.additionalBytes, 0);
+  assert.equal(result.preview.remainingBytes, 68 * GiB);
+});
+
+test('40 GiB incoming leaves 28 GiB above a 12 GiB reserve', () => {
+  const runtimes = {
+    incoming: { id: 'incoming', status: 'stopped', memoryGb: 40 }
+  };
+  const result = map({
+    runtimes,
+    models: [{ id: 'incoming-model', runtime: 'incoming' }],
+    previewModelId: 'incoming-model'
+  });
+  assert.equal(result.preview.status, 'fits');
+  assert.equal(result.preview.additionalBytes, 40 * GiB);
+  assert.equal(result.preview.remainingBytes, 28 * GiB);
+  assert.equal(result.preview.projectedUsedBytes, 68 * GiB);
+  assert.equal(result.preview.percent, 41.666667);
+  assert.equal(result.preview.message, 'Expected to fit');
+});
+
+test('invalid telemetry does not invent segments or fits', () => {
+  for (const memory of [
+    { totalBytes: null, availableBytes: 1, usedBytes: 1 },
+    { totalBytes: -1, availableBytes: 1, usedBytes: 1 },
+    { totalBytes: 10, availableBytes: 11, usedBytes: 1 },
+    { totalBytes: 10, availableBytes: -1, usedBytes: 1 },
+    { totalBytes: 10, availableBytes: null, usedBytes: 11 }
+  ]) {
+    const result = buildMemoryMap({
+      node: { id: 'node-a', local: true, reachable: true, telemetry: { memory } },
+      memorySafety: policy
+    });
+    assert.equal(result.known, false);
+    assert.deepEqual(result.segments, []);
+    assert.equal(result.preview, null);
+  }
+  const unreachable = map({ nodeOverrides: { reachable: false } });
+  assert.equal(unreachable.known, false);
+});
+
+test('RSS is preferred over a configured peak and oversubscription is bounded', () => {
+  const result = map({
+    runtimes: {
+      a: {
+        id: 'a',
+        status: 'running',
+        memoryGb: 40,
+        memoryUsage: { residentBytes: 5 * GiB, groupId: 'proc:1', sharedRuntimeIds: ['a'] }
+      },
+      b: {
+        id: 'b',
+        status: 'running',
+        memoryUsage: { residentBytes: 6 * GiB, groupId: 'proc:2', sharedRuntimeIds: ['b'] }
+      }
+    }
+  });
+  assert.equal(result.segments.find((segment) => segment.runtimeId === 'a').bytes, 5 * GiB);
+  assert.equal(
+    result.segments.reduce((sum, segment) => sum + segment.bytes, 0),
+    96 * GiB
+  );
+});
+
+test('oversubscribed estimates are scaled within measured used space', () => {
+  const result = map({
+    nodeOverrides: { total: 10 * GiB, available: 9 * GiB },
+    runtimes: {
+      a: { id: 'a', status: 'starting', memoryGb: 6 },
+      b: { id: 'b', status: 'starting', memoryGb: 6 }
+    }
+  });
+  const runtimeBytes = result.segments
+    .filter((segment) => segment.kind === 'runtime')
+    .reduce((sum, segment) => sum + segment.bytes, 0);
+  assert.ok(runtimeBytes <= result.usedBytes);
+  assert.equal(
+    result.segments.reduce((sum, segment) => sum + segment.bytes, 0),
+    result.totalBytes
+  );
+  assert.ok(result.segments.filter((segment) => segment.kind === 'runtime').every((segment) => segment.estimated));
+});
+
+test('shared runtime groups and confirmed empty lazy lists are honest', () => {
+  const result = map({
+    runtimes: {
+      'shared-a': {
+        id: 'shared-a',
+        status: 'running',
+        command: 'ollama',
+        memoryGb: 8,
+        memoryUsage: {
+          residentBytes: 8 * GiB,
+          groupId: 'proc:9',
+          sharedRuntimeIds: ['shared-a', 'shared-b'],
+          loadedModelIds: [],
+          residencyKnown: true
+        }
+      },
+      'shared-b': {
+        id: 'shared-b',
+        status: 'running',
+        command: 'ollama',
+        memoryGb: 8,
+        memoryUsage: {
+          residentBytes: 8 * GiB,
+          groupId: 'proc:9',
+          sharedRuntimeIds: ['shared-a', 'shared-b'],
+          loadedModelIds: [],
+          residencyKnown: true
+        }
+      }
+    },
+    models: [
+      { id: 'shared-model', runtime: 'shared-a', upstreamModel: 'shared-upstream', targets: [{ node: 'node-a' }] }
+    ],
+    previewModelId: 'shared-model'
+  });
+  const runtimeSegments = result.segments.filter((segment) => segment.kind === 'runtime');
+  assert.equal(runtimeSegments.length, 1);
+  assert.equal(runtimeSegments[0].bytes, 8 * GiB);
+  assert.equal(result.preview.status, 'fits');
+  assert.equal(result.preview.additionalBytes, 8 * GiB);
+});
+
+test('missing cold estimate is unknown rather than zero or fits', () => {
+  const result = map({
+    runtimes: { cold: { id: 'cold', status: 'stopped' } },
+    models: [{ id: 'cold-model', runtime: 'cold', targets: [{ node: 'node-a' }] }],
+    previewModelId: 'cold-model'
+  });
+  assert.equal(result.preview.status, 'unknown');
+  assert.equal(result.preview.additionalBytes, null);
+  assert.equal(result.preview.percent, null);
+});
+
+test('remote runtimes are excluded and previews remain node-local', () => {
+  const result = map({
+    runtimes: {
+      remote: {
+        id: 'remote',
+        remote: true,
+        node: 'node-b',
+        status: 'running',
+        memoryUsage: { residentBytes: 20 * GiB, groupId: 'proc:remote', sharedRuntimeIds: ['remote'] }
+      }
+    },
+    models: [{ id: 'remote-model', runtime: 'remote', targets: [{ node: 'node-b' }] }],
+    previewModelId: 'remote-model'
+  });
+  assert.deepEqual(
+    result.segments.filter((segment) => segment.kind === 'runtime'),
+    []
+  );
+  assert.equal(result.preview.status, 'other-node');
+  assert.equal(result.preview.nodeId, 'node-b');
+  assert.equal(result.preview.additionalBytes, 0);
+});
+
+test('distributed wrappers and remote resources do not double count', () => {
+  const result = map({
+    runtimes: {
+      member: {
+        id: 'member',
+        status: 'running',
+        memoryUsage: { residentBytes: 4 * GiB, groupId: 'proc:member', sharedRuntimeIds: ['member'] }
+      },
+      distributed: { id: 'distributed', distributed: true, members: [{ runtime: 'member' }], status: 'running' }
+    }
+  });
+  assert.deepEqual(
+    result.segments.filter((segment) => segment.kind === 'runtime').map((segment) => segment.runtimeId),
+    ['member']
+  );
+});
+
+test('remote reserve is not pooled into the local reserve', () => {
+  const result = map({
+    runtimes: { remote: { id: 'remote', remote: true, node: 'node-b', runtimeManager: { memorySafety: policy } } }
+  });
+  assert.equal(result.reserveBytes, 12 * GiB);
+});
+
+test('runtime colors are stable and order independent', () => {
+  const first = map({ runtimes: { a: { id: 'a', status: 'running' }, b: { id: 'b', status: 'running' } } });
+  const second = map({ runtimes: { b: { id: 'b', status: 'running' }, a: { id: 'a', status: 'running' } } });
+  const colors = (result) =>
+    Object.fromEntries(
+      result.segments
+        .filter((segment) => segment.kind === 'runtime')
+        .map((segment) => [segment.runtimeId, segment.colorIndex])
+    );
+  assert.deepEqual(colors(first), colors(second));
+  assert.ok(Object.values(colors(first)).every((value) => value >= 0 && value <= 7));
+});
+
+test('the pure builder is serializable into a browser function', () => {
+  const source = buildMemoryMap.toString();
+  const browserFunction = new Function(`return (${source})`)();
+  const result = browserFunction({
+    node: {
+      id: 'node-a',
+      local: true,
+      reachable: true,
+      telemetry: { memory: { totalBytes: 10, availableBytes: 7, usedBytes: 3 } }
+    }
+  });
+  assert.equal(result.known, true);
+  assert.equal(result.usedBytes, 3);
+});
+
+test('a peer uses its own observed runtime and reserve, never pooled capacity', () => {
+  const result = map({
+    node: {
+      id: 'peer',
+      local: false,
+      reachable: true,
+      telemetry: { memory: { totalBytes: 128 * GiB, availableBytes: 32 * GiB } },
+      runtimeManager: {
+        memorySafety: { ...policy, minAvailableMemoryGb: 20 },
+        runtimes: {
+          remote: {
+            status: 'running',
+            healthy: true,
+            memoryUsage: { residentBytes: 16 * GiB, residencyKnown: true, loadedModelIds: ['resident'] }
+          }
+        }
+      }
+    },
+    runtimes: { remote: { remote: true, node: 'peer', memoryGb: 16 } },
+    models: [{ id: 'resident', name: 'Resident on peer', runtime: 'remote' }],
+    previewModelId: 'resident'
+  });
+  assert.equal(result.preview.status, 'resident');
+  assert.equal(result.reserveBytes, 20 * GiB);
+  assert.equal(result.segments.find((s) => s.kind === 'runtime').bytes, 16 * GiB);
+  assert.equal(
+    result.segments.reduce((sum, s) => sum + s.percent, 0),
+    100
+  );
+});
+
+test('maintenance and missing runtime state cannot promise a free model', () => {
+  const base = { models: [{ id: 'paused', runtime: 'model' }], previewModelId: 'paused' };
+  assert.equal(
+    map({
+      ...base,
+      runtimes: { model: { healthy: true, status: 'running', maintenance: { state: 'suspended' }, memoryGb: 32 } }
+    }).preview.status,
+    'paused'
+  );
+  assert.equal(map(base).preview.status, 'unknown');
+});
+
+test('a healthy but empty cache still requires incoming memory', () => {
+  const result = map({
+    runtimes: {
+      a: {
+        command: 'ollama',
+        status: 'running',
+        healthy: true,
+        memoryGb: 8,
+        memoryUsage: { residentBytes: GiB, residencyKnown: true, loadedModelIds: [] }
+      }
+    },
+    models: [{ id: 'cold', runtime: 'a' }],
+    previewModelId: 'cold'
+  });
+  assert.equal(result.preview.status, 'fits');
+  assert.equal(result.preview.additionalBytes, 8 * GiB);
+});
+
+test('a healthy single-model service may still need to allocate its weights', () => {
+  const result = map({
+    runtimes: {
+      a: {
+        status: 'running',
+        healthy: true,
+        memoryGb: 32,
+        memoryUsage: { residentBytes: GiB / 8, residencyKnown: false }
+      }
+    },
+    models: [{ id: 'image', runtime: 'a' }],
+    previewModelId: 'image'
+  });
+  assert.equal(result.preview.status, 'fits');
+  assert.equal(result.preview.additionalBytes, 31.875 * GiB);
+});
+
+test('replica identifiers are not machine identifiers', () => {
+  const result = map({
+    models: [{ id: 'cloud', targets: [{ id: 'default', backend: 'provider' }] }],
+    previewModelId: 'cloud'
+  });
+  assert.equal(result.preview.status, 'external');
+});
+
+test('a federated target resolves the observed runtime on its peer', () => {
+  const result = map({
+    node: {
+      id: 'peer',
+      reachable: true,
+      local: false,
+      telemetry: { memory: { totalBytes: 96 * GiB, availableBytes: 48 * GiB } },
+      runtimeManager: {
+        memorySafety: policy,
+        runtimes: {
+          audio: {
+            status: 'running',
+            healthy: true,
+            memoryGb: 8,
+            memoryUsage: { residentBytes: 4 * GiB, residencyKnown: true, loadedModelIds: ['voice'] }
+          }
+        }
+      }
+    },
+    models: [
+      {
+        id: 'peer/voice',
+        name: 'Peer voice',
+        upstreamModel: 'voice',
+        targets: [{ id: 'replica', node: 'peer', remoteRuntime: 'audio' }],
+        federated: true
+      }
+    ],
+    previewModelId: 'peer/voice'
+  });
+  assert.equal(result.preview.status, 'resident');
+  assert.equal(result.segments[0].runtimeId, 'audio');
+  assert.deepEqual(result.segments[0].modelIds, ['peer/voice']);
+});
diff --git a/test/first-run.test.mjs b/test/first-run.test.mjs
new file mode 100644
index 0000000..476c398
--- /dev/null
+++ b/test/first-run.test.mjs
@@ -0,0 +1,390 @@
+import { test } from 'node:test';
+import assert from 'node:assert/strict';
+import http from 'node:http';
+import fs from 'node:fs/promises';
+import os from 'node:os';
+import path from 'node:path';
+import { createHash } from 'node:crypto';
+import { createFirstRunServer, shouldOpenSetup } from '../src/first-run.mjs';
+import { recipeSupportsWorkload, createReviewedSetupRunner } from '../src/browser-setup.mjs';
+import { applyInit } from '../src/init.mjs';
+
+async function fixture(overrides = {}) {
+  let applications = 0;
+  const app = createFirstRunServer({
+    planBuilder: async ({ workloadId, recipeId }) => {
+      if (recipeId && recipeId !== 'safe-recipe') throw new Error('Unknown recipe');
+      return {
+        plan: { configPath: '/owned/config', marker: 'reviewed' },
+        view: { workloadId, selected: { id: 'safe-recipe' } }
+      };
+    },
+    applyRunner: async (plan) => {
+      applications++;
+      assert.equal(plan.marker, 'reviewed');
+      return { ok: true, configPath: plan.configPath };
+    },
+    gatewayStarter: async () => ({ status: 'started', url: 'http://127.0.0.1:8100' }),
+    gatewayProbe: async () => ({ healthy: true, inferenceVerified: true }),
+    logger: { error() {} },
+    ...overrides
+  });
+  await app.listen();
+  const bootstrap = new URL(app.bootstrapUrl()),
+    origin = bootstrap.origin;
+  const token = new URLSearchParams(bootstrap.hash.slice(1)).get('setup');
+  const headers = { authorization: 'Bearer ' + token, origin, 'content-type': 'application/json' };
+  const get = async (suffix) => {
+    const response = await fetch(origin + '/gateway/first-run/' + suffix, { headers });
+    return { response, body: await response.json() };
+  };
+  const post = async (body) => {
+    const response = await fetch(origin + '/gateway/first-run/apply', {
+      method: 'POST',
+      headers,
+      body: JSON.stringify(body)
+    });
+    return { response, body: await response.json() };
+  };
+  return { app, origin, headers, get, post, applications: () => applications };
+}
+test('browser setup remains optional for scripts and explicit CLI plans', () => {
+  assert.equal(shouldOpenSetup([], { interactive: true }), true);
+  assert.equal(shouldOpenSetup(['--browser']), true);
+  for (const flag of ['--json', '--offline', '--no-browser', '--go', '--apply', '--recipe', '--format'])
+    assert.equal(shouldOpenSetup(['--browser', flag], { interactive: true }), false);
+  assert.equal(shouldOpenSetup(['--browser'], { installed: true }), false);
+});
+test('workload filters do not recommend chat recipes for images or unsupported hardware intents', () => {
+  const chat = { capabilities: ['chat', 'tools'] };
+  assert.equal(recipeSupportsWorkload(chat, 'code'), true);
+  assert.equal(recipeSupportsWorkload(chat, 'images'), false);
+  assert.equal(recipeSupportsWorkload({ capabilities: ['audio-transcription'] }, 'voice'), true);
+});
+test('fresh bootstrap token, exact origin and host protect all setup data and writes', async () => {
+  const f = await fixture();
+  try {
+    assert.equal((await fetch(f.origin + '/gateway/first-run/job')).status, 401);
+    assert.equal(
+      (await fetch(f.origin + '/gateway/first-run/job', { headers: { ...f.headers, origin: 'https://evil.example' } }))
+        .status,
+      403
+    );
+    const reboundStatus = await new Promise((resolve, reject) => {
+      http
+        .get(f.origin + '/', { headers: { host: 'rebound.example' } }, (response) => {
+          response.resume();
+          resolve(response.statusCode);
+        })
+        .on('error', reject);
+    });
+    assert.equal(reboundStatus, 403);
+    const html = await fetch(f.origin + '/').then((r) => r.text());
+    assert(!html.includes(f.headers.authorization.split(' ')[1]));
+    assert.equal((await f.post({ planId: 'x', yes: false })).response.status, 400);
+    assert.equal(f.applications(), 0);
+  } finally {
+    await f.app.close();
+  }
+});
+test('apply binds the reviewed plan and repeated clicks run once; gateway starts after install', async () => {
+  const calls = [];
+  let release;
+  const barrier = new Promise((resolve) => (release = resolve));
+  const f = await fixture({
+    applyRunner: async (plan) => {
+      calls.push('apply');
+      assert.equal(plan.marker, 'reviewed');
+      await barrier;
+      return { ok: true, configPath: plan.configPath };
+    },
+    gatewayStarter: async () => {
+      calls.push('gateway');
+      return { status: 'started' };
+    }
+  });
+  try {
+    const plan = (await f.get('plan?workload=code')).body.plan;
+    assert.equal((await f.post({ planId: plan.planId, yes: true, configPath: '/attacker' })).response.status, 400);
+    assert.equal((await f.post({ planId: plan.planId, yes: true, workloadId: 'voice' })).response.status, 409);
+    const started = await f.post({ planId: plan.planId, yes: true, workloadId: 'code' });
+    assert.equal(started.response.status, 202);
+    const duplicate = await f.post({ planId: plan.planId, yes: true });
+    assert.equal(duplicate.body.job.id, started.body.job.id);
+    assert.deepEqual(calls, ['apply']);
+    assert.equal((await f.get('plan?workload=chat')).response.status, 409);
+    release();
+    await f.app.close();
+    assert.deepEqual(calls, ['apply', 'gateway']);
+  } finally {
+    release();
+    if (f.app.server.listening) await f.app.close();
+  }
+});
+test('expired plans and unknown recipe selections cannot install', async () => {
+  let clock = 100;
+  const f = await fixture({ now: () => clock, planTtlMs: 10 });
+  try {
+    const plan = (await f.get('plan')).body.plan;
+    clock = 111;
+    assert.equal((await f.post({ planId: plan.planId, yes: true })).response.status, 410);
+    assert.equal((await f.get('plan?recipe=unknown')).response.status, 500);
+    assert.equal(f.applications(), 0);
+  } finally {
+    await f.app.close();
+  }
+});
+test('health alone never marks inference ready; install failure never starts the gateway', async () => {
+  for (const fail of [false, true]) {
+    let started = false;
+    const f = await fixture({
+      applyRunner: async () => {
+        if (fail) throw new Error('disk full');
+        return { ok: true };
+      },
+      gatewayStarter: async () => {
+        started = true;
+        return { status: 'started' };
+      },
+      gatewayProbe: async () => ({ healthy: true, inferenceVerified: false, detail: 'Model still loading' })
+    });
+    try {
+      const plan = (await f.get('plan')).body.plan;
+      await f.post({ planId: plan.planId, yes: true });
+      const job = (await f.get('job')).body.job;
+      assert.equal(job.ready, false);
+      assert.equal(job.inferenceVerified, false);
+      assert.equal(started, !fail);
+      if (fail) {
+        assert.equal(job.status, 'failed');
+        assert.match(job.error, /disk full/);
+      }
+    } finally {
+      await f.app.close();
+    }
+  }
+});
+
+test('first-run resumes only its unchanged configuration after a partial bootstrap failure', async () => {
+  const dir = await fs.mkdtemp(path.join(os.tmpdir(), 'lloom-first-run-owned-'));
+  const file = path.join(dir, 'config.json');
+  const raw = { models: [], runtimes: {}, backends: {} };
+  const evidence = { recipe: { id: 'reviewed' }, backendCatalog: {} };
+  const fingerprint = (value) => createHash('sha256').update(JSON.stringify(value)).digest('hex');
+  const plan = {
+    configPath: file,
+    modelRoot: path.join(dir, 'models'),
+    selectedRecipe: { id: 'reviewed' },
+    setup: { phases: { init: { config: raw } } },
+    browserSetup: { fingerprint: fingerprint(evidence), options: {} }
+  };
+  let applies = 0,
+    bootstraps = 0;
+  const run = createReviewedSetupRunner(
+    {},
+    {
+      evidence: async () => evidence,
+      apply: async (_config, options) => {
+        applies++;
+        assert.equal(options.reviewedPlan, plan.setup);
+        assert.equal(options.exclusiveConfig, true);
+        await fs.writeFile(file, JSON.stringify(raw), { flag: 'wx' });
+        await options.onConfigWritten(file, raw);
+        throw new Error('download interrupted');
+      },
+      bootstrap: async () => {
+        bootstraps++;
+        return { ok: true };
+      }
+    }
+  );
+  try {
+    await assert.rejects(run(plan, {}), /download interrupted/);
+    assert.equal((await run(plan, {})).ok, true);
+    assert.equal(applies, 1);
+    assert.equal(bootstraps, 1);
+    await fs.writeFile(file, JSON.stringify({ ...raw, changed: true }));
+    await assert.rejects(run(plan, {}), /will not overwrite/);
+    evidence.recipe.id = 'changed';
+    await assert.rejects(run(plan, {}), /recipe or backend catalog changed/);
+  } finally {
+    await fs.rm(dir, { recursive: true, force: true });
+  }
+});
+
+test('exclusive first-run publication cannot replace a file created after review', async () => {
+  const dir = await fs.mkdtemp(path.join(os.tmpdir(), 'lloom-first-run-exclusive-'));
+  const file = path.join(dir, 'config.json');
+  try {
+    await fs.writeFile(file, 'original');
+    await assert.rejects(
+      applyInit(
+        {},
+        { dryRun: false, yes: true, exclusiveConfig: true, reviewedPlan: { configPath: file, config: { models: [] } } }
+      ),
+      { code: 'EEXIST' }
+    );
+    assert.equal(await fs.readFile(file, 'utf8'), 'original');
+    assert.deepEqual(await fs.readdir(dir), ['config.json']);
+  } finally {
+    await fs.rm(dir, { recursive: true, force: true });
+  }
+});
+
+test('failed HTTP setup retries reuse the immutable reviewed plan instead of a fresh timestamp', async () => {
+  let plans = 0,
+    attempts = 0;
+  const f = await fixture({
+    planBuilder: async ({ workloadId }) => ({
+      plan: { configPath: '/owned/config', generatedAt: ++plans },
+      view: { workloadId, selected: { id: 'safe-recipe' } }
+    }),
+    applyRunner: async (plan) => {
+      assert.equal(plan.generatedAt, 1);
+      if (++attempts === 1) throw new Error('interrupted download');
+      return { ok: true, configPath: plan.configPath };
+    }
+  });
+  try {
+    const first = (await f.get('plan')).body.plan;
+    await f.post({ planId: first.planId, yes: true });
+    assert.equal((await f.get('job')).body.job.status, 'failed');
+    const retry = (await f.get('plan')).body.plan;
+    assert.notEqual(retry.planId, first.planId);
+    assert.equal(plans, 1);
+    await f.post({ planId: retry.planId, yes: true });
+    assert.equal((await f.get('job')).body.job.status, 'succeeded');
+    await f.post({ planId: retry.planId, yes: true });
+    assert.equal(attempts, 2, 'completed apply remains idempotent');
+  } finally {
+    await f.app.close();
+  }
+});
+
+test('reviewed bootstrap executes pinned step plans without reopening source catalogs', async () => {
+  const { applyBootstrap } = await import('../src/bootstrap.mjs');
+  const dir = await fs.mkdtemp(path.join(os.tmpdir(), 'lloom-frozen-bootstrap-'));
+  const reviewed = {
+    reviewedRecipe: { id: 'frozen', name: 'Frozen recipe', backend: { id: 'frozen-backend' } },
+    reviewedBackend: { id: 'frozen-backend', name: 'Frozen backend' },
+    modelRoot: path.join(dir, 'models'),
+    backend: {
+      platformSupported: true,
+      steps: [{ id: 'approved-backend', action: 'command', command: ['approved', 'backend'] }]
+    },
+    recipe: {
+      validationErrors: [],
+      steps: [{ id: 'approved-model', action: 'command', command: ['approved', 'model'] }]
+    }
+  };
+  try {
+    const result = await applyBootstrap(
+      { server: { host: '127.0.0.1', port: 8100 }, security: {}, models: [], runtimes: {} },
+      {
+        reviewedPlan: reviewed,
+        dryRun: true,
+        home: dir,
+        statePath: path.join(dir, 'state.json'),
+        recipesRoot: '/must-not-reopen-recipes',
+        backendCatalogPath: '/must-not-reopen-catalog',
+        clientId: 'all'
+      }
+    );
+    assert.deepEqual(result.backend.results[0].command, ['approved', 'backend']);
+    assert.deepEqual(result.recipe.results[0].command, ['approved', 'model']);
+    assert.equal(result.selectedRecipe.id, 'frozen');
+  } finally {
+    await fs.rm(dir, { recursive: true, force: true });
+  }
+});
+
+test('reviewed downloads retain their executable when the environment changes', async () => {
+  const { pinDownloadCommands, applyRecipe } = await import('../src/installer.mjs');
+  const dir = await fs.mkdtemp(path.join(os.tmpdir(), 'lloom-reviewed-download-'));
+  try {
+    const executable = path.join(dir, 'reviewed-hf');
+    await fs.writeFile(
+      executable,
+      '#!/bin/sh\nfor last do :; done\nmkdir -p "$last"\nprintf "{}" > "$last/config.json"\nprintf "weights" > "$last/model.safetensors"\n',
+      { mode: 0o755 }
+    );
+    const destination = path.join(dir, 'model');
+    const reviewed = await pinDownloadCommands(
+      {
+        validationErrors: [],
+        steps: [
+          {
+            id: 'download',
+            action: 'download-model',
+            provider: 'huggingface',
+            model: 'test/model',
+            destination,
+            command: ['hf', 'download', 'test/model', '--local-dir', destination]
+          }
+        ]
+      },
+      { env: { ...process.env, LLOOM_HF_BIN: executable } }
+    );
+    assert.equal(reviewed.steps[0].command[0], executable);
+    const report = await applyRecipe(
+      { id: 'reviewed-test' },
+      {},
+      {
+        reviewedPlan: reviewed,
+        dryRun: false,
+        yes: true,
+        statePath: path.join(dir, 'state.json'),
+        env: { ...process.env, LLOOM_HF_BIN: '/must-not-run', HF_HUB_CLI: '/must-not-run' }
+      }
+    );
+    assert.equal(report.results[0].status, 'completed', JSON.stringify(report.results));
+    assert.equal(report.results[0].command[0], executable);
+    assert.equal(await fs.readFile(path.join(destination, 'model.safetensors'), 'utf8'), 'weights');
+  } finally {
+    await fs.rm(dir, { recursive: true, force: true });
+  }
+});
+
+test('legacy command downloads retain the reviewed executable and all vendor arguments', async () => {
+  const { pinDownloadCommands, applyRecipe } = await import('../src/installer.mjs');
+  const dir = await fs.mkdtemp(path.join(os.tmpdir(), 'lloom-reviewed-legacy-'));
+  try {
+    const executable = path.join(dir, 'hf');
+    await fs.writeFile(
+      executable,
+      '#!/bin/sh\nfor last do :; done\nmkdir -p "$last"\nprintf "reviewed" > "$last/model.gguf"\n',
+      { mode: 0o755 }
+    );
+    const destination = path.join(dir, 'model');
+    const args = [
+      'download',
+      'owner/model',
+      'model.gguf',
+      '--revision',
+      '1234567890abcdef1234567890abcdef12345678',
+      '--local-dir',
+      destination
+    ];
+    const reviewed = await pinDownloadCommands(
+      { validationErrors: [], steps: [{ id: 'legacy', action: 'command', command: ['hf', ...args] }] },
+      { env: { PATH: dir + ':/bin:/usr/bin' } }
+    );
+    assert.deepEqual(reviewed.steps[0].command, [executable, ...args]);
+    const report = await applyRecipe(
+      { id: 'legacy-reviewed' },
+      {},
+      {
+        reviewedPlan: reviewed,
+        dryRun: false,
+        yes: true,
+        statePath: path.join(dir, 'state.json'),
+        env: { PATH: '/bin:/usr/bin', LLOOM_HF_BIN: '/must-not-run' }
+      }
+    );
+    assert.equal(report.results[0].status, 'completed', JSON.stringify(report.results));
+    assert.deepEqual(report.results[0].command, [executable, ...args]);
+    assert.equal(await fs.readFile(path.join(destination, 'model.gguf'), 'utf8'), 'reviewed');
+  } finally {
+    await fs.rm(dir, { recursive: true, force: true });
+  }
+});
diff --git a/test/fixtures/nvidia-sync-three-node.json b/test/fixtures/nvidia-sync-three-node.json
new file mode 100644
index 0000000..bb0c5b4
--- /dev/null
+++ b/test/fixtures/nvidia-sync-three-node.json
@@ -0,0 +1,62 @@
+{
+  "hostname": "spark-test-host",
+  "nodeId": "ennspark01",
+  "sshConfig": "Host 10.100.16.1\n  User sparkadmin\nHost worker-fabric\n  HostName 10.100.17.1\n  User sparkadmin\nHost ennspark02-lan\n  ### CreatedBy: NVIDIA Sync\n  ### NVSyncClusterAlias: ennspark02-lan\n  Hostname 10.100.192.1\n  User sparkadmin\nHost 10.100.192.1\n  ### CreatedBy: NVIDIA Sync\n  ### NVSyncClusterAlias: ennspark02-lan\n  Hostname 10.100.192.1\n  User sparkadmin\nHost 10.100.193.1\n  ### CreatedBy: NVIDIA Sync\n  ### NVSyncClusterAlias: ennspark02-lan\n  Hostname 10.100.193.1\n  User sparkadmin\nHost ennspark03-lan\n  ### CreatedBy: NVIDIA Sync\n  ### NVSyncClusterAlias: ennspark03-lan\n  Hostname 10.100.196.2\n  User sparkadmin\nHost 10.100.196.2\n  ### CreatedBy: NVIDIA Sync\n  ### NVSyncClusterAlias: ennspark03-lan\n  Hostname 10.100.196.2\n  User sparkadmin\nHost 10.100.197.2\n  ### CreatedBy: NVIDIA Sync\n  ### NVSyncClusterAlias: ennspark03-lan\n  Hostname 10.100.197.2\n  User sparkadmin\n",
+  "ipAddresses": [
+    {
+      "ifname": "lo",
+      "addr_info": [
+        {
+          "family": "inet",
+          "local": "127.0.0.1",
+          "prefixlen": 8,
+          "scope": "host"
+        }
+      ]
+    },
+    {
+      "ifname": "enp1s0f0np0",
+      "addr_info": [
+        {
+          "family": "inet",
+          "local": "10.100.192.2",
+          "prefixlen": 24,
+          "scope": "global"
+        }
+      ]
+    },
+    {
+      "ifname": "enp1s0f1np1",
+      "addr_info": [
+        {
+          "family": "inet",
+          "local": "10.100.196.1",
+          "prefixlen": 24,
+          "scope": "global"
+        }
+      ]
+    },
+    {
+      "ifname": "enP2p1s0f0np0",
+      "addr_info": [
+        {
+          "family": "inet",
+          "local": "10.100.193.2",
+          "prefixlen": 24,
+          "scope": "global"
+        }
+      ]
+    },
+    {
+      "ifname": "enP2p1s0f1np1",
+      "addr_info": [
+        {
+          "family": "inet",
+          "local": "10.100.197.1",
+          "prefixlen": 24,
+          "scope": "global"
+        }
+      ]
+    }
+  ]
+}
diff --git a/test/host-memory.test.mjs b/test/host-memory.test.mjs
index 78eb850..3ad1958 100644
--- a/test/host-memory.test.mjs
+++ b/test/host-memory.test.mjs
@@ -23,6 +23,34 @@ assert.equal(mac.availableBytes, Math.round(96 * gibibyte * 0.44));
 assert(Math.abs(mac.utilization - 56) < 0.000001);
 assert.equal(mac.source, 'macos-memory-pressure');
 
+// Full memory_pressure output: page counts (free+inactive+speculative+purgeable)
+// must win over the opaque free percentage.
+const paged = parseMacMemoryPressure(
+  `The system has 103079215104 (6291456 pages with a page size of 16384).
+
+Stats:
+Pages free: 100000
+Pages purgeable: 20000
+Pages purged: 349093517
+
+Page Q counts:
+Pages active: 2130087
+Pages inactive: 30000
+Pages speculative: 5000
+Pages throttled: 0
+Pages wired down: 495904
+
+System-wide memory free percentage: 1%`,
+  96 * gibibyte
+);
+assert.equal(paged.source, 'macos-memory-pages');
+assert.equal(paged.availableBytes, 155000 * 16384);
+
+// Percentage-only output keeps the legacy path.
+const percentOnly = parseMacMemoryPressure('System-wide memory free percentage: 44%', 96 * gibibyte);
+assert.equal(percentOnly.source, 'macos-memory-pressure');
+assert.equal(percentOnly.availableBytes, Math.round(96 * gibibyte * 0.44));
+
 const sampledMac = await readHostMemory({
   platform: 'darwin',
   totalBytes: 96 * gibibyte,
diff --git a/test/installation-jobs.test.mjs b/test/installation-jobs.test.mjs
new file mode 100644
index 0000000..6254f59
--- /dev/null
+++ b/test/installation-jobs.test.mjs
@@ -0,0 +1,227 @@
+import { test } from 'node:test';
+import assert from 'node:assert/strict';
+import fs from 'node:fs/promises';
+import os from 'node:os';
+import path from 'node:path';
+import { createInstallationJobs } from '../src/installation-jobs.mjs';
+import { createDashboardInstallation } from '../src/dashboard-installation.mjs';
+import { loadConfig } from '../src/config.mjs';
+import { installImportedModelAssets } from '../src/model-installation.mjs';
+import { MODEL_ACQUISITION_MANIFEST } from '../src/model-acquisition.mjs';
+
+test('review is read-only; apply is bound, serialized and idempotent across refresh', async () => {
+  let release,
+    calls = 0,
+    clock = 0;
+  const barrier = new Promise((resolve) => (release = resolve));
+  const jobs = createInstallationJobs({
+    now: () => clock,
+    ttlMs: 10,
+    plan: async (input) => ({ input, view: { summary: 'Review' } }),
+    apply: async (prepared, progress) => {
+      calls++;
+      assert.equal(prepared.input.recipeId, 'vendor');
+      progress({ message: 'Downloading' });
+      await barrier;
+      return { ok: true };
+    }
+  });
+  const a = await jobs.review({ recipeId: 'vendor' });
+  const b = await jobs.review({ recipeId: 'vendor' });
+  assert.equal(calls, 0);
+  assert.throws(() => jobs.start({ planId: a.planId, yes: false }), /reviewed/);
+  assert.throws(() => jobs.start({ planId: a.planId, yes: true, recipeId: 'other' }), /reviewed/);
+  const first = jobs.start({ planId: a.planId, yes: true });
+  assert.equal(jobs.start({ planId: a.planId, yes: true }).id, first.id);
+  assert.throws(() => jobs.start({ planId: b.planId, yes: true }), /already/);
+  await Promise.resolve();
+  assert.equal(jobs.snapshot().detail, 'Downloading');
+  release();
+  assert.equal((await jobs.wait()).status, 'succeeded');
+  assert.equal(calls, 1);
+  clock = 11;
+  assert.throws(() => jobs.start({ planId: b.planId, yes: true }), /fresh/);
+});
+
+test('installation failures remain visible and a fresh reviewed plan can retry', async () => {
+  const jobs = createInstallationJobs({
+    plan: async () => ({ view: {} }),
+    apply: async () => {
+      throw new Error('disk full');
+    }
+  });
+  const plan = await jobs.review({});
+  jobs.start({ planId: plan.planId, yes: true });
+  const failed = await jobs.wait();
+  assert.equal(failed.status, 'failed');
+  assert.equal(failed.error, 'disk full');
+  assert.notEqual((await jobs.review({})).planId, plan.planId);
+});
+
+test('dashboard installs an unmanaged endpoint without processes and preserves raw secrets', async () => {
+  const dir = await fs.mkdtemp(path.join(os.tmpdir(), 'lloom-installation-'));
+  try {
+    const file = path.join(dir, 'config.json');
+    await fs.writeFile(
+      file,
+      JSON.stringify({ models: [], backends: {}, runtimes: {}, security: { apiKeys: ['${UNEXPANDED_TEST_KEY}'] } })
+    );
+    let config = await loadConfig(file),
+      reloads = 0;
+    const jobs = createDashboardInstallation({
+      getConfig: () => config,
+      reload: async () => {
+        config = await loadConfig(file);
+        reloads++;
+      }
+    });
+    const input = { modelRef: 'openai:http://127.0.0.1:9999/v1#test-model' };
+    const plan = await jobs.review(input);
+    jobs.start({ planId: plan.planId, yes: true });
+    const result = await jobs.wait();
+    assert.equal(result.status, 'succeeded', result.error);
+    assert.equal(reloads, 1);
+    const raw = JSON.parse(await fs.readFile(file, 'utf8'));
+    assert.equal(raw.security.apiKeys[0], '${UNEXPANDED_TEST_KEY}');
+    assert.equal(raw.models.length, 1);
+    const stale = await jobs.review({ modelRef: 'openai:http://127.0.0.1:9998/v1#another-model' });
+    raw.extra = 'concurrent edit';
+    await fs.writeFile(file, JSON.stringify(raw));
+    jobs.start({ planId: stale.planId, yes: true });
+    assert.equal((await jobs.wait()).status, 'failed');
+    assert.equal(JSON.parse(await fs.readFile(file, 'utf8')).extra, 'concurrent edit');
+  } finally {
+    await fs.rm(dir, { recursive: true, force: true });
+  }
+});
+
+test('dashboard installation routes enforce browser origin and publish the reviewed endpoint', async () => {
+  const { createLloomServer } = await import('../src/server.mjs');
+  const dir = await fs.mkdtemp(path.join(os.tmpdir(), 'lloom-install-http-'));
+  let app;
+  try {
+    const file = path.join(dir, 'config.json');
+    await fs.writeFile(
+      file,
+      JSON.stringify({
+        server: { host: '127.0.0.1', port: 0 },
+        security: { allowMissingAuth: true },
+        models: [],
+        runtimes: {},
+        backends: {}
+      })
+    );
+    app = createLloomServer(await loadConfig(file), { logger: { log() {}, error() {} } });
+    await app.listen();
+    const origin = 'http://127.0.0.1:' + app.server.address().port;
+    const post = (suffix, body, site = origin) =>
+      fetch(origin + '/gateway/installations' + suffix, {
+        method: 'POST',
+        headers: { origin: site, 'content-type': 'application/json' },
+        body: JSON.stringify(body)
+      });
+    const input = { modelRef: 'openai:http://127.0.0.1:9999/v1#http-test' };
+    assert.equal((await post('/plan', input, 'https://foreign.example')).status, 403);
+    const response = await post('/plan', input);
+    assert.equal(response.status, 200);
+    const plan = await response.json();
+    assert.equal((await post('', { planId: plan.planId, yes: true })).status, 202);
+    let job;
+    for (let i = 0; i < 50; i++) {
+      job = (await fetch(origin + '/gateway/installations', { headers: { origin } }).then((r) => r.json())).job;
+      if (job.status !== 'running') break;
+      await new Promise((resolve) => setTimeout(resolve, 10));
+    }
+    assert.equal(job.status, 'succeeded', job.error);
+    const catalog = await fetch(origin + '/gateway/models', { headers: { origin } }).then((r) => r.json());
+    assert(catalog.models.some((model) => model.id === 'http-test'));
+  } finally {
+    if (app) await app.close();
+    await fs.rm(dir, { recursive: true, force: true });
+  }
+});
+
+test('browser imports require an immutable Hugging Face revision before creating a plan', async () => {
+  const dir = await fs.mkdtemp(path.join(os.tmpdir(), 'lloom-import-review-'));
+  try {
+    const file = path.join(dir, 'config.json');
+    await fs.writeFile(file, JSON.stringify({ models: [], runtimes: {}, paths: { modelRoot: dir } }));
+    const config = await loadConfig(file);
+    const executable = path.join(dir, 'hf');
+    const jobs = createDashboardInstallation({
+      getConfig: () => config,
+      reload: async () => {},
+      env: { PATH: dir, LLOOM_HF_BIN: executable }
+    });
+    for (const modelRef of ['owner/model-gguf', 'https://huggingface.co/owner/model-gguf/tree/main'])
+      await assert.rejects(jobs.review({ modelRef }), /pinned to a commit/);
+    const revision = '1234567890abcdef1234567890abcdef12345678';
+    await assert.rejects(
+      jobs.review({ modelRef: 'https://huggingface.co/owner/model-gguf/resolve/' + revision + '/model.gguf' }),
+      /downloader is not installed/
+    );
+    await fs.writeFile(executable, '#!/bin/sh\nexit 0\n', { mode: 0o755 });
+    const reviewed = await jobs.review({
+      modelRef: 'https://huggingface.co/owner/model-gguf/resolve/' + revision + '/model.gguf'
+    });
+    assert.equal(reviewed.details.download.acquisition.revision, revision);
+    assert.deepEqual(reviewed.details.download.acquisition.include, ['model.gguf']);
+    assert(path.isAbsolute(reviewed.details.download.command[0]));
+    assert.deepEqual(
+      (await fs.readdir(dir)).sort(),
+      ['config.json', 'hf'],
+      'review must not create download directories'
+    );
+  } finally {
+    await fs.rm(dir, { recursive: true, force: true });
+  }
+});
+
+test('import acquisition keeps failed payloads staged and verifies before publishing', async () => {
+  const dir = await fs.mkdtemp(path.join(os.tmpdir(), 'lloom-import-acquisition-'));
+  try {
+    const destination = path.join(dir, 'model');
+    const executable = path.join(dir, 'hf');
+    const revision = '1234567890abcdef1234567890abcdef12345678';
+    const acquisition = {
+      provider: 'huggingface',
+      model: 'owner/model',
+      revision,
+      destination,
+      include: ['model.gguf']
+    };
+    const command = [
+      executable,
+      'download',
+      acquisition.model,
+      '--revision',
+      revision,
+      '--include',
+      'model.gguf',
+      '--local-dir',
+      destination
+    ];
+    const plan = { additions: {}, reference: { type: 'huggingface' }, download: { command, acquisition } };
+    // A successful subprocess that writes the wrong payload must not publish it.
+    await fs.writeFile(executable, '#!/bin/sh\nfor last do :; done\nprintf "partial" > "$last/wrong.gguf"\n', {
+      mode: 0o755
+    });
+    await assert.rejects(installImportedModelAssets(plan), /verification failed/);
+    await assert.rejects(fs.access(destination), { code: 'ENOENT' });
+    await fs.access(destination + '.incomplete/wrong.gguf');
+    await fs.writeFile(executable, '#!/bin/sh\nfor last do :; done\nprintf "weights" > "$last/model.gguf"\n', {
+      mode: 0o755
+    });
+    await installImportedModelAssets(plan);
+    await assert.rejects(fs.access(destination + '.incomplete'), { code: 'ENOENT' });
+    const manifest = JSON.parse(await fs.readFile(path.join(destination, MODEL_ACQUISITION_MANIFEST), 'utf8'));
+    assert.equal(manifest.revision, revision);
+    assert.deepEqual(manifest.include, ['model.gguf']);
+    assert.equal(await fs.readFile(path.join(destination, 'model.gguf'), 'utf8'), 'weights');
+    // Verified data is reusable even after the original downloader disappears.
+    await fs.unlink(executable);
+    await installImportedModelAssets(plan);
+  } finally {
+    await fs.rm(dir, { recursive: true, force: true });
+  }
+});
diff --git a/test/long-prefill-transport.test.mjs b/test/long-prefill-transport.test.mjs
index 85d3d25..41e8fb4 100644
--- a/test/long-prefill-transport.test.mjs
+++ b/test/long-prefill-transport.test.mjs
@@ -1,6 +1,7 @@
 import assert from 'node:assert/strict';
 import http from 'node:http';
-import { Agent, getGlobalDispatcher, setGlobalDispatcher } from 'undici';
+// Dispatcher and fetch must come from the same undici copy.
+import { Agent, fetch as undiciFetch, getGlobalDispatcher, setGlobalDispatcher } from 'undici';
 import { createLloomServer } from '../src/server.mjs';
 const listen = (server) =>
   new Promise((resolve) => server.listen(0, '127.0.0.1', () => resolve(server.address().port)));
@@ -47,7 +48,7 @@ try {
     );
     const port = await listen(app.server);
     try {
-      const r = await fetch(`http://127.0.0.1:${port}/v1/chat/completions`, {
+      const r = await undiciFetch(`http://127.0.0.1:${port}/v1/chat/completions`, {
         method: 'POST',
         dispatcher: client,
         headers: { 'content-type': 'application/json' },
diff --git a/test/openrouter-provider.test.mjs b/test/openrouter-provider.test.mjs
new file mode 100644
index 0000000..25388e6
--- /dev/null
+++ b/test/openrouter-provider.test.mjs
@@ -0,0 +1,489 @@
+import assert from 'node:assert/strict';
+import { Agent, MockAgent, getGlobalDispatcher, setGlobalDispatcher } from 'undici';
+import { createLloomServer } from '../src/server.mjs';
+import {
+  applyOpenRouterProviderPolicy,
+  isOpenRouterBackend,
+  normalizeOpenRouterProviderPolicy
+} from '../src/protocol/openrouter-provider.mjs';
+
+// The gateway consults this for its long-running upstream dispatcher while a
+// mock session is active.
+let gatewayDispatcher = null;
+
+// ---------------------------------------------------------------------------
+// Pure policy unit coverage
+// ---------------------------------------------------------------------------
+
+const openRouterBackend = (extra = {}) => ({
+  id: 'openrouter-lane',
+  type: 'openai',
+  baseUrl: 'https://openrouter.ai/api/v1',
+  ...extra
+});
+
+function testLookalikeHosts() {
+  assert.equal(isOpenRouterBackend(openRouterBackend()), true);
+  assert.equal(isOpenRouterBackend({ baseUrl: 'https://openrouter.ai/api/v1' }), true);
+  for (const baseUrl of [
+    'https://openrouter.ai.evil.example/api/v1',
+    'https://notopenrouter.ai/api/v1',
+    'https://openrouter.ai.example.com/api/v1',
+    'https://api.openrouter.ai/api/v1',
+    'http://127.0.0.1:8200/v1',
+    'not a url',
+    '',
+    undefined
+  ]) {
+    assert.equal(isOpenRouterBackend({ baseUrl }), false, `expected lookalike to be rejected: ${baseUrl}`);
+  }
+}
+
+function testNoConfigLeavesBodyUntouched() {
+  const body = {
+    model: 'z-ai/glm-5.2',
+    messages: [{ role: 'user', content: 'hi' }],
+    provider: { order: ['together'], allow_fallbacks: true }
+  };
+  const frozen = JSON.parse(JSON.stringify(body));
+  // No policy key at all.
+  assert.equal(applyOpenRouterProviderPolicy(body, openRouterBackend()), body);
+  // Policy explicitly null is "not configured", not malformed.
+  assert.equal(applyOpenRouterProviderPolicy(body, openRouterBackend({ openrouterProvider: null })), body);
+  // Non-OpenRouter host is untouched. A policy there is never enforced, so a
+  // plain (not-openrouter) backend must not be rejected for carrying one.
+  const local = { id: 'local', baseUrl: 'http://127.0.0.1:8201/v1', openrouterProvider: { only: ['z-ai'] } };
+  assert.equal(applyOpenRouterProviderPolicy(body, local), body);
+  assert.deepEqual(body, frozen);
+}
+
+function testValidPolicyEnforced() {
+  const body = { model: 'z-ai/glm-5.2', messages: [{ role: 'user', content: 'hi' }] };
+  const applied = applyOpenRouterProviderPolicy(body, openRouterBackend({ openrouterProvider: { only: ['z-ai'] } }));
+  assert.deepEqual(applied.provider, { only: ['z-ai'], allow_fallbacks: false });
+  // Default allow_fallbacks is false (fail closed) when omitted.
+  assert.equal(applied.provider.allow_fallbacks, false);
+  // Original body is not mutated.
+  assert.equal(body.provider, undefined);
+
+  const explicit = applyOpenRouterProviderPolicy(
+    body,
+    openRouterBackend({ openrouterProvider: { only: ['z-ai', 'z-ai-intl'], allow_fallbacks: true } })
+  );
+  assert.deepEqual(explicit.provider, { only: ['z-ai', 'z-ai-intl'], allow_fallbacks: true });
+
+  // Whitespace is trimmed and surrounding body fields are preserved.
+  const trimmed = applyOpenRouterProviderPolicy(
+    { model: 'z-ai/glm-5.2', temperature: 0.4, provider: { order: ['x'] } },
+    openRouterBackend({ openrouterProvider: { only: [' z-ai '] } })
+  );
+  assert.deepEqual(trimmed.provider.only, ['z-ai']);
+  assert.equal(trimmed.temperature, 0.4);
+  // Caller provider fields survive alongside the enforced keys.
+  assert.deepEqual(trimmed.provider.order, ['x']);
+}
+
+function testCallerOverrideAttemptsLose() {
+  const backend = openRouterBackend({ openrouterProvider: { only: ['z-ai'], allow_fallbacks: false } });
+  const cases = [
+    { provider: { only: ['openai'], allow_fallbacks: true } },
+    { provider: { allow_fallbacks: true } },
+    { provider: { only: [] } },
+    { provider: { only: ['any'] } }
+  ];
+  for (const extra of cases) {
+    const applied = applyOpenRouterProviderPolicy({ model: 'm', messages: [], ...extra }, backend);
+    assert.deepEqual(applied.provider, { only: ['z-ai'], allow_fallbacks: false }, JSON.stringify(extra));
+  }
+
+  const permissive = openRouterBackend({ openrouterProvider: { only: ['z-ai'], allow_fallbacks: true } });
+  const applied = applyOpenRouterProviderPolicy({ provider: { allow_fallbacks: false } }, permissive);
+  assert.deepEqual(applied.provider, { only: ['z-ai'], allow_fallbacks: true });
+}
+
+function testNonObjectBodyUntouched() {
+  const backend = openRouterBackend({ openrouterProvider: { only: ['z-ai'] } });
+  for (const body of [null, undefined, 'not-an-object', 42, ['array']]) {
+    assert.equal(applyOpenRouterProviderPolicy(body, backend), body);
+  }
+}
+
+function testMalformedPoliciesThrow() {
+  const base = { id: 'openrouter-lane', baseUrl: 'https://openrouter.ai/api/v1' };
+  const malformed = [
+    {}, // only is required
+    { only: [] },
+    { only: 'z-ai' },
+    { only: ['z-ai', ''] },
+    { only: ['z-ai', '   '] },
+    { only: [null] },
+    { only: ['z-ai'], allow_fallbacks: 'no' },
+    { only: ['z-ai'], allow_fallbacks: 1 },
+    { only: ['z-ai'], order: ['x'] }, // unknown key
+    { only: ['z-ai'], allow_fallbacks: false, ignore: ['a'] },
+    { only: ['z-ai'], dataCollection: 'deny' },
+    'z-ai',
+    ['z-ai'],
+    42
+  ];
+  for (const openrouterProvider of malformed) {
+    assert.throws(
+      () => normalizeOpenRouterProviderPolicy(openrouterProvider, { backendId: 'openrouter-lane' }),
+      /openrouterProvider/,
+      `expected malformed policy to throw: ${JSON.stringify(openrouterProvider)}`
+    );
+    assert.throws(
+      () => applyOpenRouterProviderPolicy({ model: 'm' }, { ...base, openrouterProvider }),
+      /openrouterProvider/,
+      `expected apply to fail closed for: ${JSON.stringify(openrouterProvider)}`
+    );
+  }
+  // No configured value (undefined/null) is not malformed.
+  assert.equal(normalizeOpenRouterProviderPolicy(undefined), null);
+  assert.equal(normalizeOpenRouterProviderPolicy(null), null);
+}
+
+// ---------------------------------------------------------------------------
+// Gateway integration: the whole /v1/chat/completions path is exercised with a
+// mocked undici dispatcher so no network or TLS egress occurs.
+// ---------------------------------------------------------------------------
+
+function gatewayConfig(backend) {
+  return {
+    server: { host: '127.0.0.1', port: 0 },
+    security: { allowMissingAuth: true, apiKeys: [] },
+    defaults: { chatModel: 'z-ai/glm-5.2' },
+    backends: { 'openrouter-lane': backend },
+    models: [
+      {
+        id: 'z-ai/glm-5.2',
+        backend: 'openrouter-lane',
+        upstreamModel: 'z-ai/glm-5.2',
+        kind: 'chat',
+        contextWindow: 200000,
+        maxPromptTokens: 100000
+      }
+    ],
+    runtimes: {}
+  };
+}
+
+async function withMockedDispatcher(startGatewayFn, fn) {
+  // The gateway's upstream calls go through its long-running undici Agent.
+  // Undici 8 composes mocks by wrapping that agent instead of patching
+  // prototypes, so the composed dispatcher is installed as the package global
+  // (the package fetch consults it) and handed to the gateway explicitly.
+  // Loopback stays open so client requests to the local gateway pass through.
+  const previous = getGlobalDispatcher();
+  const longRunningAgent = new Agent({ headersTimeout: 1800000, bodyTimeout: 1800000 });
+  const mockAgent = new MockAgent({ agent: longRunningAgent });
+  mockAgent.disableNetConnect();
+  mockAgent.enableNetConnect(
+    (host) => typeof host === 'string' && (host.startsWith('127.0.0.1') || host.startsWith('localhost'))
+  );
+  setGlobalDispatcher(mockAgent);
+  gatewayDispatcher = mockAgent;
+  try {
+    await startGatewayFn();
+    return await fn(mockAgent);
+  } finally {
+    gatewayDispatcher = null;
+    setGlobalDispatcher(previous);
+    await mockAgent.close();
+  }
+}
+
+async function readRequestBody(body) {
+  if (!body || typeof body[Symbol.asyncIterator] !== 'function') return body;
+  const chunks = [];
+  for await (const chunk of body) chunks.push(chunk);
+  return Buffer.concat(chunks).toString('utf8');
+}
+
+function intercept(mockAgent, { status = 200, contentType = 'application/json', payload, onBody }) {
+  const pool = mockAgent.get('https://openrouter.ai');
+  return pool.intercept({ path: '/api/v1/chat/completions', method: 'POST' }).reply(
+    status,
+    async (opts) => {
+      onBody(await readRequestBody(opts.body));
+      return payload;
+    },
+    { headers: { 'content-type': contentType } }
+  );
+}
+
+const openAiChatPayload = {
+  id: 'chatcmpl-1',
+  object: 'chat.completion',
+  choices: [{ index: 0, message: { role: 'assistant', content: 'ok' }, finish_reason: 'stop' }],
+  usage: { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 }
+};
+
+const openAiStreamPayload = [
+  'data: {"id":"chatcmpl-1","object":"chat.completion.chunk","choices":[{"index":0,"delta":{"role":"assistant","content":"ok"},"finish_reason":null}]}',
+  '',
+  'data: {"id":"chatcmpl-1","object":"chat.completion.chunk","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}',
+  '',
+  'data: [DONE]',
+  '',
+  ''
+].join('\n');
+
+let gatewayServer;
+let gatewayPort;
+
+async function startGateway(backend) {
+  const app = createLloomServer(gatewayConfig(backend), {
+    logger: { error() {}, warn() {}, info() {}, log() {} },
+    upstreamDispatcher: gatewayDispatcher ?? undefined
+  });
+  await new Promise((resolve) => app.server.listen(0, '127.0.0.1', resolve));
+  gatewayServer = app.server;
+  gatewayPort = app.server.address().port;
+  return gatewayPort;
+}
+
+async function stopGateway() {
+  if (!gatewayServer) return;
+  const server = gatewayServer;
+  gatewayServer = null;
+  await new Promise((resolve) => server.close(resolve));
+}
+
+async function testGatewayChatBuffered() {
+  const seen = [];
+  try {
+    await withMockedDispatcher(
+      async () => startGateway(openRouterBackend({ openrouterProvider: { only: ['z-ai'], allow_fallbacks: false } })),
+      async (mockAgent) => {
+      intercept(mockAgent, { payload: openAiChatPayload, onBody: (body) => seen.push(body) });
+      const res = await fetch(`http://127.0.0.1:${gatewayPort}/v1/chat/completions`, {
+        method: 'POST',
+        headers: { 'content-type': 'application/json' },
+        body: JSON.stringify({
+          model: 'z-ai/glm-5.2',
+          messages: [{ role: 'user', content: 'hi' }],
+          provider: { only: ['openai'], allow_fallbacks: true, order: ['x'] }
+        })
+      });
+      assert.equal(res.status, 200);
+      const json = await res.json();
+      assert.equal(json.choices[0].message.content, 'ok');
+    });
+    assert.equal(seen.length, 1, 'expected exactly one upstream chat call');
+    const outbound = JSON.parse(seen[0]);
+    // Caller override attempt is defeated; other caller provider fields survive.
+    assert.deepEqual(outbound.provider, { order: ['x'], only: ['z-ai'], allow_fallbacks: false });
+    assert.equal(outbound.model, 'z-ai/glm-5.2');
+    assert.deepEqual(outbound.messages, [{ role: 'user', content: 'hi' }]);
+  } finally {
+    await stopGateway();
+  }
+}
+
+async function testGatewayChatStream() {
+  const seen = [];
+  try {
+    await withMockedDispatcher(
+      async () => startGateway(openRouterBackend({ openrouterProvider: { only: ['z-ai'] } })),
+      async (mockAgent) => {
+      intercept(mockAgent, {
+        contentType: 'text/event-stream',
+        payload: openAiStreamPayload,
+        onBody: (body) => seen.push(body)
+      });
+      const res = await fetch(`http://127.0.0.1:${gatewayPort}/v1/chat/completions`, {
+        method: 'POST',
+        headers: { 'content-type': 'application/json' },
+        body: JSON.stringify({
+          model: 'z-ai/glm-5.2',
+          messages: [{ role: 'user', content: 'hi' }],
+          stream: true
+        })
+      });
+      assert.equal(res.status, 200);
+      const text = await res.text();
+      assert.match(text, /data: \[DONE\]/);
+    });
+    assert.equal(seen.length, 1);
+    const outbound = JSON.parse(seen[0]);
+    assert.deepEqual(outbound.provider, { only: ['z-ai'], allow_fallbacks: false });
+    assert.equal(outbound.stream, true);
+  } finally {
+    await stopGateway();
+  }
+}
+
+async function testGatewayResponsesBridge(stream = false) {
+  const seen = [];
+  try {
+    await withMockedDispatcher(
+      async () => startGateway(openRouterBackend({ openrouterProvider: { only: ['z-ai'] } })),
+      async (mockAgent) => {
+      intercept(mockAgent, {
+        payload: stream ? openAiStreamPayload : openAiChatPayload,
+        contentType: stream ? 'text/event-stream' : 'application/json',
+        onBody: (body) => seen.push(body)
+      });
+      const res = await fetch(`http://127.0.0.1:${gatewayPort}/v1/responses`, {
+        method: 'POST',
+        headers: { 'content-type': 'application/json' },
+        body: JSON.stringify({ stream, model: 'z-ai/glm-5.2', input: 'hi', provider: { only: ['openai'] } })
+      });
+      assert.equal(res.status, 200);
+      if (stream) assert.match(await res.text(), /response.completed/);
+      else assert.equal((await res.json()).object, 'response');
+    });
+    assert.equal(seen.length, 1);
+    const outbound = JSON.parse(seen[0]);
+    assert.deepEqual(outbound.provider, { only: ['z-ai'], allow_fallbacks: false });
+  } finally {
+    await stopGateway();
+  }
+}
+
+async function testGatewayAnthropicBridge(stream = false) {
+  const seen = [];
+  try {
+    await withMockedDispatcher(
+      async () => startGateway(openRouterBackend({ openrouterProvider: { only: ['z-ai'], allow_fallbacks: true } })),
+      async (mockAgent) => {
+      intercept(mockAgent, {
+        payload: stream ? openAiStreamPayload : openAiChatPayload,
+        contentType: stream ? 'text/event-stream' : 'application/json',
+        onBody: (body) => seen.push(body)
+      });
+      const res = await fetch(`http://127.0.0.1:${gatewayPort}/v1/messages`, {
+        method: 'POST',
+        headers: { 'content-type': 'application/json', 'anthropic-version': '2023-06-01' },
+        body: JSON.stringify({
+          stream,
+          model: 'z-ai/glm-5.2',
+          max_tokens: 32,
+          messages: [{ role: 'user', content: 'hi' }]
+        })
+      });
+      assert.equal(res.status, 200);
+      if (stream) assert.match(await res.text(), /message_stop/);
+      else assert.equal((await res.json()).type, 'message');
+    });
+    assert.equal(seen.length, 1);
+    const outbound = JSON.parse(seen[0]);
+    assert.deepEqual(outbound.provider, { only: ['z-ai'], allow_fallbacks: true });
+  } finally {
+    await stopGateway();
+  }
+}
+
+async function testGatewayMalformedPolicyFailsClosed() {
+  // The gateway must not silently send an unconstrained request to OpenRouter.
+  try {
+    await withMockedDispatcher(
+      async () => startGateway(openRouterBackend({ openrouterProvider: { only: [] } })),
+      async () => {
+      const res = await fetch(`http://127.0.0.1:${gatewayPort}/v1/chat/completions`, {
+        method: 'POST',
+        headers: { 'content-type': 'application/json' },
+        body: JSON.stringify({ model: 'z-ai/glm-5.2', messages: [{ role: 'user', content: 'hi' }] })
+      });
+      assert.notEqual(res.status, 200);
+      await res.text();
+    });
+  } finally {
+    await stopGateway();
+  }
+}
+
+async function testGatewayNoPolicyUntouched() {
+  const seen = [];
+  try {
+    await withMockedDispatcher(
+      async () => startGateway({ id: 'openrouter-lane', type: 'openai', baseUrl: 'https://openrouter.ai/api/v1' }),
+      async (mockAgent) => {
+      intercept(mockAgent, { payload: openAiChatPayload, onBody: (body) => seen.push(body) });
+      const res = await fetch(`http://127.0.0.1:${gatewayPort}/v1/chat/completions`, {
+        method: 'POST',
+        headers: { 'content-type': 'application/json' },
+        body: JSON.stringify({
+          model: 'z-ai/glm-5.2',
+          messages: [{ role: 'user', content: 'hi' }],
+          provider: { only: ['openai'], allow_fallbacks: true }
+        })
+      });
+      assert.equal(res.status, 200);
+      await res.text();
+    });
+    const outbound = JSON.parse(seen[0]);
+    assert.deepEqual(outbound.provider, { only: ['openai'], allow_fallbacks: true });
+  } finally {
+    await stopGateway();
+  }
+}
+
+async function testGatewayLookalikeHostUntouched() {
+  const seen = [];
+  const previous = getGlobalDispatcher();
+  const longRunningAgent = new Agent({ headersTimeout: 1800000, bodyTimeout: 1800000 });
+  const mockAgent = new MockAgent({ agent: longRunningAgent });
+  mockAgent.disableNetConnect();
+  mockAgent.enableNetConnect(
+    (host) => typeof host === 'string' && (host.startsWith('127.0.0.1') || host.startsWith('localhost'))
+  );
+  setGlobalDispatcher(mockAgent);
+  gatewayDispatcher = mockAgent;
+  try {
+    mockAgent
+      .get('https://notopenrouter.ai')
+      .intercept({ path: '/api/v1/chat/completions', method: 'POST' })
+      .reply(
+        200,
+        async (opts) => {
+          seen.push(await readRequestBody(opts.body));
+          return openAiChatPayload;
+        },
+        { headers: { 'content-type': 'application/json' } }
+      );
+    await startGateway({
+      id: 'openrouter-lane',
+      type: 'openai',
+      baseUrl: 'https://notopenrouter.ai/api/v1',
+      openrouterProvider: { only: ['z-ai'] }
+    });
+    const res = await fetch(`http://127.0.0.1:${gatewayPort}/v1/chat/completions`, {
+      method: 'POST',
+      headers: { 'content-type': 'application/json' },
+      body: JSON.stringify({ model: 'z-ai/glm-5.2', messages: [{ role: 'user', content: 'hi' }] })
+    });
+    assert.equal(res.status, 200);
+    await res.text();
+  } finally {
+    gatewayDispatcher = null;
+    setGlobalDispatcher(previous);
+    await mockAgent.close();
+    await stopGateway();
+  }
+  const outbound = JSON.parse(seen[0]);
+  // Lookalike host never receives the OpenRouter provider policy.
+  assert.equal(outbound.provider, undefined);
+}
+
+// ---------------------------------------------------------------------------
+
+testLookalikeHosts();
+testNoConfigLeavesBodyUntouched();
+testValidPolicyEnforced();
+testCallerOverrideAttemptsLose();
+testNonObjectBodyUntouched();
+testMalformedPoliciesThrow();
+
+await testGatewayChatBuffered();
+await testGatewayChatStream();
+await testGatewayResponsesBridge();
+await testGatewayResponsesBridge(true);
+await testGatewayAnthropicBridge();
+await testGatewayAnthropicBridge(true);
+await testGatewayMalformedPolicyFailsClosed();
+await testGatewayNoPolicyUntouched();
+await testGatewayLookalikeHostUntouched();
+
+console.log('openrouter-provider: ok');
diff --git a/test/runtime-memory-safety.test.mjs b/test/runtime-memory-safety.test.mjs
new file mode 100644
index 0000000..38baa61
--- /dev/null
+++ b/test/runtime-memory-safety.test.mjs
@@ -0,0 +1,373 @@
+import test from 'node:test';
+import assert from 'node:assert/strict';
+import fs from 'node:fs/promises';
+import os from 'node:os';
+import path from 'node:path';
+import http from 'node:http';
+import { spawn } from 'node:child_process';
+import { setTimeout as delay } from 'node:timers/promises';
+import { RuntimeManager } from '../src/runtime-manager.mjs';
+import {
+  memorySafetyPolicy,
+  assertMemorySafety,
+  createMemorySafetyGuard,
+  MemorySafetyConfigError
+} from '../src/runtime-memory-safety.mjs';
+import { readHostMemory } from '../src/host-memory.mjs';
+import { terminateProcessTree } from '../src/process-control.mjs';
+
+const GiB = 1024 ** 3;
+const snapshot = (available) => ({ totalBytes: 96 * GiB, availableBytes: available * GiB });
+const healthy = snapshot(60);
+const pressure = snapshot(3);
+const alive = (pid) => {
+  try {
+    process.kill(pid, 0);
+    return true;
+  } catch {
+    return false;
+  }
+};
+async function until(fn) {
+  for (let i = 0; i < 100; i++) {
+    if (await fn()) return;
+    await delay(25);
+  }
+  throw new Error('Timed out waiting for small test process');
+}
+
+async function fixture(t, { warmup = false, yolo = false } = {}) {
+  const dir = await fs.mkdtemp(path.join(os.tmpdir(), 'lloom-memory-guard-'));
+  const receipt = path.join(dir, 'receipt.json');
+  const marker = path.join(dir, 'warmup');
+  const code = path.join(dir, 'backend.mjs');
+  const reserve = http.createServer();
+  await new Promise((resolve) => reserve.listen(0, '127.0.0.1', resolve));
+  const port = reserve.address().port;
+  await new Promise((resolve) => reserve.close(resolve));
+  await fs.writeFile(
+    code,
+    `
+    import http from 'node:http';
+    import fs from 'node:fs';
+    import {spawn} from 'node:child_process';
+    const worker=spawn(process.execPath,['-e',"process.on('SIGTERM',()=>{});setInterval(()=>{},1000)"],{stdio:'ignore'});
+    fs.writeFileSync(process.argv[2],JSON.stringify({backend:process.pid,worker:worker.pid}));
+    http.createServer((req,res)=>{
+      if(req.url==='/warmup'){fs.writeFileSync(process.argv[3],'warming');return;}
+      res.writeHead(${warmup || yolo ? 200 : 503});res.end('{}');
+    }).listen(${port},'127.0.0.1');
+  `
+  );
+  const config = {
+    runtimePolicy: {
+      enabled: false,
+      reserveMemoryGb: 12,
+      memorySafety: { mode: yolo ? 'yolo' : 'enforce', pollIntervalMs: 50 }
+    },
+    runtimes: {
+      demo: {
+        enabled: true,
+        command: process.execPath,
+        args: [code, receipt, marker],
+        port,
+        healthUrl: `http://127.0.0.1:${port}/health`,
+        startupTimeoutMs: 5000,
+        ...(warmup ? { warmup: { url: `http://127.0.0.1:${port}/warmup`, method: 'POST', body: {} } } : {})
+      }
+    }
+  };
+  const manager = new RuntimeManager(config, { logger: { error() {} }, memorySampler: async () => healthy });
+  const ids = async () => JSON.parse(await fs.readFile(receipt, 'utf8'));
+  t.after(async () => {
+    const pids = await ids().catch(() => null);
+    if (pids) await terminateProcessTree([pids.backend, pids.worker], { termTimeoutMs: 100 });
+    const supervisor = manager.processes.get('demo');
+    if (supervisor?.pid) await terminateProcessTree([supervisor.pid], { termTimeoutMs: 100 });
+    await fs.rm(dir, { recursive: true, force: true });
+  });
+  return { manager, receipt, marker, ids, port };
+}
+
+test('hard limits include the host reserve, reject malformed telemetry, and use inclusive boundaries', () => {
+  const policy = memorySafetyPolicy({ runtimePolicy: { reserveMemoryGb: 12 } }, 96);
+  assert.equal(policy.minAvailableMemoryGb, 12);
+  assert.throws(() => assertMemorySafety(policy, snapshot(12)), /hard reserve/);
+  assert.throws(() => assertMemorySafety(policy, { totalBytes: 1, availableBytes: 2 }), /could not be measured/);
+  assert.throws(() => assertMemorySafety(policy, null), /could not be measured/);
+  assert.doesNotThrow(() => assertMemorySafety(policy, healthy));
+  assert.equal(
+    memorySafetyPolicy({ runtimePolicy: { memorySafety: { minAvailableMemoryGb: 20 } } }, 96).minAvailableMemoryGb,
+    20
+  );
+  assert.throws(
+    () => memorySafetyPolicy({ runtimePolicy: { memorySafety: { mode: 'YOLO-ish' } } }),
+    (error) => error instanceof MemorySafetyConfigError && /Invalid/.test(error.message) && error.statusCode === 500
+  );
+});
+
+test('strict Mac telemetry fails closed instead of substituting free memory', async () => {
+  await assert.rejects(
+    readHostMemory({
+      platform: 'darwin',
+      strict: true,
+      execFileImpl: async () => {
+        throw new Error('sampler unavailable');
+      }
+    }),
+    /sampler unavailable/
+  );
+  await assert.rejects(
+    readHostMemory({ platform: 'darwin', strict: true, execFileImpl: async () => ({ stdout: 'invalid' }) }),
+    /unavailable/
+  );
+});
+
+test('force and disabled predictive policy cannot bypass hard preflight', async (t) => {
+  const f = await fixture(t);
+  f.manager.memorySampler = async () => pressure;
+  await assert.rejects(f.manager.start('demo', { force: true }), (e) => e.code === 'runtime_memory_safety_abort');
+  assert.equal(f.manager.processes.size, 0);
+  assert.equal(
+    await fs.access(f.receipt).then(
+      () => true,
+      () => false
+    ),
+    false
+  );
+});
+
+for (const warmup of [false, true])
+  test(`pressure during blocked ${warmup ? 'warmup' : 'health'} kills only the new process tree`, async (t) => {
+    const f = await fixture(t, { warmup });
+    const sentinel = spawn(process.execPath, ['-e', 'setInterval(()=>{},1000)'], { stdio: 'ignore' });
+    t.after(() => terminateProcessTree([sentinel.pid], { termTimeoutMs: 100 }));
+    let trip = false;
+    f.manager.memorySampler = async () => (trip ? pressure : healthy);
+    const starting = f.manager.start('demo', { reason: 'model-request' });
+    const rejected = assert.rejects(starting, (e) => e.code === 'runtime_memory_safety_abort' && !e.temporary);
+    await until(() =>
+      fs.access(warmup ? f.marker : f.receipt).then(
+        () => true,
+        () => false
+      )
+    );
+    const pids = await f.ids();
+    assert(alive(pids.backend));
+    trip = true;
+    await rejected;
+    await until(() => !alive(pids.backend) && !alive(pids.worker));
+    assert(alive(sentinel.pid), 'unrelated process must survive');
+    assert.equal(f.manager.stateFor('demo').status, 'failed');
+    assert.match(f.manager.stateFor('demo').lastError, /protect this machine/);
+    f.manager.memorySampler = async () => healthy;
+    await assert.rejects(
+      f.manager.start('demo', { reason: 'model-request' }),
+      (e) => e.code === 'runtime_memory_safety_abort'
+    );
+    assert.equal(f.manager.stateFor('demo').starts, 1, 'automatic retry must not spawn again');
+  });
+
+test('a reused healthy endpoint remains usable under pressure', async (t) => {
+  const server = http.createServer((req, res) => res.end('{}'));
+  await new Promise((resolve) => server.listen(0, '127.0.0.1', resolve));
+  t.after(() => new Promise((resolve) => server.close(resolve)));
+  const manager = new RuntimeManager(
+    { runtimes: { existing: { enabled: true, healthUrl: `http://127.0.0.1:${server.address().port}/health` } } },
+    { memorySampler: async () => pressure }
+  );
+  const result = await manager.start('existing', { warmup: false });
+  assert.equal(result.reason, 'already-healthy');
+  assert(server.listening);
+});
+
+test('explicit YOLO is the only guard bypass', async (t) => {
+  const f = await fixture(t, { yolo: true });
+  f.manager.memorySampler = async () => {
+    throw new Error('must not be sampled in YOLO');
+  };
+  const result = await f.manager.start('demo');
+  assert.equal(result.healthy, true);
+  assert.equal((await f.manager.status()).memorySafety.mode, 'yolo');
+  assert(alive((await f.ids()).backend));
+});
+
+test('a stale sample cannot abort a completed operation', async () => {
+  let release;
+  let aborts = 0;
+  const guard = createMemorySafetyGuard({
+    policy: memorySafetyPolicy({}),
+    sample: () =>
+      new Promise((resolve) => {
+        release = resolve;
+      }),
+    onAbort: () => aborts++
+  });
+  const checking = guard.check();
+  await until(() => Boolean(release));
+  const stopping = guard.stop();
+  release(pressure);
+  await Promise.all([checking, stopping]);
+  assert.equal(aborts, 0);
+});
+
+test('the independent supervisor enforces the transmitted reserve without gateway sampling', async (t) => {
+  const dir = await fs.mkdtemp(path.join(os.tmpdir(), 'lloom-independent-guard-'));
+  const source = path.resolve('src');
+  for (const name of ['runtime-supervisor.mjs', 'runtime-memory-safety.mjs'])
+    await fs.copyFile(path.join(source, name), path.join(dir, name));
+  // Replace only the OS telemetry dependency in this isolated subprocess test.
+  await fs.writeFile(
+    path.join(dir, 'host-memory.mjs'),
+    `import fs from 'node:fs/promises'; export async function readHostMemory(){ return JSON.parse(await fs.readFile(process.env.SAMPLE_FILE,'utf8')); }`
+  );
+  const sampleFile = path.join(dir, 'sample.json');
+  const receipt = path.join(dir, 'receipt.json');
+  await fs.writeFile(sampleFile, JSON.stringify(healthy));
+  const backend = `const fs=require('fs'),{spawn}=require('child_process');const worker=spawn(process.execPath,['-e','setInterval(()=>{},1000)'],{stdio:'ignore'});fs.writeFileSync(process.argv[1],JSON.stringify({backend:process.pid,worker:worker.pid}));setInterval(()=>{},1000);`;
+  const child = spawn(
+    process.execPath,
+    [path.join(dir, 'runtime-supervisor.mjs'), process.execPath, '-e', backend, receipt],
+    {
+      env: {
+        ...process.env,
+        SAMPLE_FILE: sampleFile,
+        LLOOM_MEMORY_SAFETY_POLICY: JSON.stringify({
+          mode: 'enforce',
+          minAvailableMemoryGb: 20,
+          maxMemoryUtilization: 0.9,
+          pollIntervalMs: 50
+        })
+      },
+      stdio: ['ignore', 'ignore', 'pipe', 'ipc']
+    }
+  );
+  let stderr = '';
+  child.stderr.on('data', (data) => {
+    stderr += data;
+  });
+  let aborted;
+  child.on('message', (message) => {
+    if (message.type === 'memory-safety-abort') aborted = message;
+  });
+  const exited = new Promise((resolve) => child.once('exit', (code, signal) => resolve({ code, signal })));
+  t.after(async () => {
+    const pids = await fs.readFile(receipt, 'utf8').then(JSON.parse, () => ({}));
+    await terminateProcessTree([child.pid, pids.backend, pids.worker].filter(Boolean), { termTimeoutMs: 100 });
+    await fs.rm(dir, { recursive: true, force: true });
+  });
+  await until(() =>
+    fs.access(receipt).then(
+      () => true,
+      () => false
+    )
+  );
+  const pids = JSON.parse(await fs.readFile(receipt, 'utf8'));
+  await fs.writeFile(sampleFile, JSON.stringify(snapshot(18)));
+  const result = await Promise.race([
+    exited,
+    delay(3000).then(() => {
+      throw new Error('Independent cutoff did not fire');
+    })
+  ]);
+  assert.equal(result.code, 78, stderr);
+  assert.match(aborted?.message ?? '', /20.0 GB hard reserve/);
+  await until(() => !alive(pids.backend) && !alive(pids.worker));
+});
+
+test('owned Docker startup is killed by immutable container ID on a hard breach', async (t) => {
+  const dir = await fs.mkdtemp(path.join(os.tmpdir(), 'lloom-docker-memory-'));
+  const started = path.join(dir, 'started');
+  const killed = path.join(dir, 'killed');
+  const docker = path.join(dir, 'docker');
+  await fs.writeFile(
+    docker,
+    `#!/usr/bin/env node
+const fs=require('fs');const started=${JSON.stringify(started)},killed=${JSON.stringify(killed)};
+switch(process.argv[2]) {
+case 'inspect': console.log(JSON.stringify({Id:'owned-id',State:{Running:fs.existsSync(started)&&!fs.existsSync(killed),Status:'created'}})); break;
+case 'start': fs.writeFileSync(started,'started'); break;
+case 'kill': if(process.argv[3]!=='owned-id')process.exit(3);fs.writeFileSync(killed,'killed');break;
+default: process.exit(4);
+}
+`,
+    { mode: 0o755 }
+  );
+  const oldPath = process.env.PATH;
+  process.env.PATH = dir + path.delimiter + oldPath;
+  t.after(async () => {
+    process.env.PATH = oldPath;
+    await fs.rm(dir, { recursive: true, force: true });
+  });
+  const manager = new RuntimeManager(
+    {
+      runtimePolicy: { memorySafety: { pollIntervalMs: 50 } },
+      runtimes: {
+        demo: {
+          enabled: true,
+          adapter: 'docker',
+          management: 'managed',
+          containerName: 'test-name',
+          healthUrl: 'http://127.0.0.1:1/health',
+          healthTimeoutMs: 50
+        }
+      }
+    },
+    {
+      logger: { error() {} },
+      memorySampler: async () =>
+        await fs.access(started).then(
+          () => pressure,
+          () => healthy
+        )
+    }
+  );
+  await assert.rejects(manager.start('demo'), (e) => e.code === 'runtime_memory_safety_abort');
+  assert.equal(await fs.readFile(killed, 'utf8'), 'killed');
+});
+
+test('supervisor preflight rejection cannot race an immediate health success', async (t) => {
+  const f = await fixture(t);
+  const totalBytes = os.totalmem();
+  f.manager.config.runtimePolicy.memorySafety.minAvailableMemoryGb = totalBytes / GiB;
+  f.manager.memorySampler = async () => ({ totalBytes: totalBytes * 4, availableBytes: totalBytes * 3 });
+  let probed = false;
+  f.manager.waitForHealth = async () => {
+    probed = true;
+    return { healthy: true };
+  };
+  await assert.rejects(f.manager.start('demo'), (e) => e.code === 'runtime_memory_safety_abort');
+  assert.equal(probed, false);
+  assert.equal(
+    await fs.access(f.receipt).then(
+      () => true,
+      () => false
+    ),
+    false
+  );
+});
+
+test('late health success after a safety abort cannot become a successful load', async (t) => {
+  const f = await fixture(t);
+  let release;
+  f.manager.waitForHealth = () =>
+    new Promise((resolve) => {
+      release = resolve;
+    });
+  let trip = false;
+  f.manager.memorySampler = async () => (trip ? pressure : healthy);
+  const starting = f.manager.start('demo');
+  const rejected = assert.rejects(starting, (e) => e.code === 'runtime_memory_safety_abort');
+  await until(() =>
+    fs.access(f.receipt).then(
+      () => Boolean(release),
+      () => false
+    )
+  );
+  const pids = await f.ids();
+  trip = true;
+  await until(() => !alive(pids.backend));
+  release({ healthy: true });
+  await rejected;
+  assert.equal(f.manager.stateFor('demo').status, 'failed');
+});
diff --git a/test/runtime-memory-usage.test.mjs b/test/runtime-memory-usage.test.mjs
new file mode 100644
index 0000000..e68881b
--- /dev/null
+++ b/test/runtime-memory-usage.test.mjs
@@ -0,0 +1,397 @@
+import assert from 'node:assert/strict';
+import { test } from 'node:test';
+import { createRuntimeMemoryUsageSampler } from '../src/runtime-memory-usage.mjs';
+
+const KiB = 1024;
+
+function response(body, { ok = true, status = 200 } = {}) {
+  return {
+    ok,
+    status,
+    text: async () => JSON.stringify(body)
+  };
+}
+
+function fakeClock() {
+  let value = 0;
+  return {
+    now: () => value,
+    advance(amount) {
+      value += amount;
+    }
+  };
+}
+
+function sampler(options = {}) {
+  const calls = { ps: 0, listeners: 0, fetches: [] };
+  return {
+    calls,
+    sampler: createRuntimeMemoryUsageSampler({
+      fetchImpl: options.fetchImpl ?? null,
+      clock: options.clock,
+      scheduleTimeout: () => 0,
+      cacheTtlMs: options.cacheTtlMs ?? 5000,
+      footprintReader: async () => ({}),
+      ...options,
+      nowIso: () => new Date(options.clock?.now?.() ?? 0).toISOString(),
+      psReader: async () => {
+        calls.ps++;
+        return options.psReader ? options.psReader() : '';
+      },
+      listenerReader: async () => {
+        calls.listeners++;
+        return options.listenerReader ? options.listenerReader() : '';
+      }
+    })
+  };
+}
+
+test('parses one unique process tree from listener fallback', async () => {
+  const { calls, sampler: usage } = sampler({
+    psReader: async () => ['  101     1       100', '  102   101        50', '  103   102        20'].join('\n'),
+    listenerReader: async () => 'lloom  101  user  12u  IPv4 100  TCP 127.0.0.1:8201 (LISTEN)'
+  });
+  const result = await usage.sample({
+    runtimes: { runtime: { id: 'runtime', status: 'running', port: 8201 } },
+    runtimeIds: ['runtime']
+  });
+  assert.equal(result.runtime.residentBytes, 170 * KiB);
+  assert.equal(result.runtime.groupId, 'proc:101');
+  assert.deepEqual(result.runtime.sharedRuntimeIds, ['runtime']);
+  assert.equal(result.runtime.loadedModelIds, null);
+  assert.equal(result.runtime.residencyKnown, false);
+  assert.match(result.runtime.sampledAt, /^\d{4}-\d{2}-\d{2}T/);
+  assert.equal(calls.ps, 1);
+  assert.equal(calls.listeners, 1);
+});
+
+test('shared runtimes are grouped and their process tree is counted once', async () => {
+  const { sampler: usage } = sampler({
+    psReader: async () => ['  101     1        50', '  102   101        25'].join('\n'),
+    listenerReader: async () => 'a 101 1u TCP 127.0.0.1:8201 (LISTEN)'
+  });
+  const runtimes = {
+    a: { id: 'a', status: 'running', port: 8201 },
+    b: { id: 'b', status: 'running', port: 8201 }
+  };
+  const result = await usage.sample({ runtimes, runtimeIds: ['b', 'a'] });
+  assert.equal(result.a.residentBytes, 75 * KiB);
+  assert.equal(result.b.residentBytes, 75 * KiB);
+  assert.equal(result.a.groupId, 'proc:101');
+  assert.equal(result.b.groupId, 'proc:101');
+  assert.deepEqual(result.a.sharedRuntimeIds, ['a', 'b']);
+  assert.deepEqual(result.b.sharedRuntimeIds, ['a', 'b']);
+});
+
+test('shared backend observations use one residency request', async () => {
+  const fetches = [];
+  const { sampler: usage } = sampler({
+    psReader: async () => '  101     1        80',
+    listenerReader: async () => 'ollama 101 1u TCP localhost:11434 (LISTEN)',
+    fetchImpl: async (url) => {
+      fetches.push(url.toString());
+      return response({ models: [{ model: 'llama3' }, { name: 'qwen' }] });
+    }
+  });
+  const result = await usage.sample({
+    runtimes: {
+      a: { id: 'a', status: 'running', command: 'ollama', port: 8201, healthUrl: 'http://localhost:8201' },
+      b: { id: 'b', status: 'running', command: 'ollama', port: 8201, healthUrl: 'http://localhost:8201' }
+    },
+    runtimeIds: ['a', 'b']
+  });
+  assert.deepEqual([...fetches], ['http://localhost:8201/api/ps']);
+  assert.deepEqual(result.a.loadedModelIds, ['llama3', 'qwen']);
+  assert.deepEqual(result.b.loadedModelIds, ['llama3', 'qwen']);
+  assert.equal(result.a.residencyKnown, true);
+});
+
+test('confirmed empty Ollama residency stays empty and is known', async () => {
+  const { sampler: usage } = sampler({
+    psReader: async () => '',
+    listenerReader: async () => '',
+    fetchImpl: async () => response({ models: [] })
+  });
+  const result = await usage.sample({
+    runtimes: { ollama: { id: 'ollama', status: 'running', command: '/usr/local/bin/ollama', port: 8201 } },
+    runtimeIds: ['ollama']
+  });
+  assert.equal(result.ollama.residentBytes, null);
+  assert.deepEqual(result.ollama.loadedModelIds, []);
+  assert.equal(result.ollama.residencyKnown, true);
+});
+
+test('audio server residency reads only its configured loopback health endpoint', async () => {
+  const fetches = [];
+  const { sampler: usage } = sampler({
+    psReader: async () => '',
+    listenerReader: async () => '',
+    fetchImpl: async (url) => {
+      fetches.push(url.toString());
+      return response({ tts_loaded: [{ id: 'voice' }], stt_loaded: ['ears'] });
+    }
+  });
+  const result = await usage.sample({
+    runtimes: {
+      audio: {
+        id: 'audio',
+        status: 'running',
+        command: '/usr/local/bin/lloom-audio-server',
+        healthUrl: 'http://127.0.0.1:8300/'
+      }
+    },
+    runtimeIds: ['audio']
+  });
+  assert.deepEqual(fetches, ['http://127.0.0.1:8300/health']);
+  assert.deepEqual(result.audio.loadedModelIds, ['ears', 'voice']);
+});
+
+test('sampling failures and failed residency calls remain unknown and do not throw', async () => {
+  const { sampler: usage } = sampler({
+    psReader: async () => {
+      throw new Error('ps failed');
+    },
+    listenerReader: async () => {
+      throw new Error('lsof failed');
+    },
+    fetchImpl: async () => {
+      throw new Error('model endpoint failed');
+    }
+  });
+  const result = await usage.sample({
+    runtimes: { ollama: { id: 'ollama', status: 'running', command: 'ollama', port: 8201 } },
+    runtimeIds: ['ollama']
+  });
+  assert.equal(result.ollama.residentBytes, null);
+  assert.equal(result.ollama.loadedModelIds, null);
+  assert.equal(result.ollama.residencyKnown, false);
+});
+
+test('concurrent samples share the process and residency caches', async () => {
+  const clock = fakeClock();
+  const { calls, sampler: usage } = sampler({
+    clock,
+    psReader: async () => '  101     1        20',
+    listenerReader: async () => 'a 101 1u TCP 127.0.0.1:8201 (LISTEN)',
+    fetchImpl: async () => response({ models: [] })
+  });
+  const request = { runtimes: { a: { id: 'a', status: 'running', command: 'ollama', port: 8201 } }, runtimeIds: ['a'] };
+  const [first, second] = await Promise.all([usage.sample(request), usage.sample(request)]);
+  assert.deepEqual(first, second);
+  assert.equal(calls.ps, 1);
+  assert.equal(calls.listeners, 1);
+});
+
+test('cached samples expire after the configured TTL', async () => {
+  const clock = fakeClock();
+  const { calls, sampler: usage } = sampler({
+    clock,
+    psReader: async () => '  101     1        20',
+    listenerReader: async () => ''
+  });
+  const request = { runtimes: { a: { id: 'a', status: 'running', pid: 101 } }, runtimeIds: ['a'] };
+  await usage.sample(request);
+  clock.advance(4999);
+  await usage.sample(request);
+  clock.advance(1);
+  await usage.sample(request);
+  assert.equal(calls.ps, 2);
+  assert.equal(calls.listeners, 2);
+});
+
+test('remote and distributed runtimes are not attributed locally', async () => {
+  const { calls, sampler: usage } = sampler({
+    psReader: async () => '  101     1        20',
+    listenerReader: async () => ''
+  });
+  const result = await usage.sample({
+    runtimes: {
+      remote: { id: 'remote', remote: true, node: 'node-b', status: 'running', pid: 101 },
+      distributed: { id: 'distributed', distributed: true, status: 'running', pid: 101 }
+    },
+    runtimeIds: ['remote', 'distributed']
+  });
+  assert.deepEqual(result, {});
+  assert.equal(calls.ps, 0);
+  assert.equal(calls.listeners, 0);
+});
+
+test('Docker container PIDs are used only on Linux; macOS VM memory stays unattributed', async () => {
+  const processRows = '   999       1        30\n   201       1        10';
+  const listeners = 'docker 201 1u TCP 127.0.0.1:8201 (LISTEN)';
+  const runtimes = {
+    docker: {
+      id: 'docker',
+      status: 'running',
+      adapter: 'docker',
+      container: { pid: 999 },
+      port: 8201
+    }
+  };
+  const darwin = sampler({
+    platform: 'darwin',
+    psReader: async () => processRows,
+    listenerReader: async () => listeners
+  });
+  const darwinResult = await darwin.sampler.sample({ runtimes, runtimeIds: ['docker'] });
+  assert.equal(darwinResult.docker.residentBytes, null);
+
+  const linux = sampler({
+    platform: 'linux',
+    psReader: async () => processRows,
+    listenerReader: async () => listeners
+  });
+  const linuxResult = await linux.sampler.sample({ runtimes, runtimeIds: ['docker'] });
+  assert.equal(linuxResult.docker.residentBytes, 40 * KiB);
+});
+
+test('lazy model calls are not made for unconfigured or remote hosts', async () => {
+  const fetches = [];
+  const { sampler: usage } = sampler({
+    psReader: async () => '',
+    listenerReader: async () => '',
+    fetchImpl: async (url) => {
+      fetches.push(url.toString());
+      return response({ models: [] });
+    }
+  });
+  const result = await usage.sample({
+    runtimes: {
+      remote: { id: 'remote', remote: true, command: 'ollama', port: 8201 },
+      external: {
+        id: 'external',
+        status: 'running',
+        command: 'ollama',
+        port: 8202,
+        healthUrl: 'https://example.invalid'
+      }
+    },
+    runtimeIds: ['remote', 'external']
+  });
+  assert.deepEqual(fetches, []);
+  assert.equal(result.external.loadedModelIds, null);
+});
+
+test('overlapping parent and child ownership is one memory group', async () => {
+  const { sampler: usage } = sampler({
+    psReader: async () => '101 1 20\n102 101 40\n103 102 10'
+  });
+  const result = await usage.sample({
+    runtimes: {
+      parent: { status: 'running', pid: 101 },
+      child: { status: 'running', pid: 102 }
+    },
+    runtimeIds: ['child', 'parent']
+  });
+  assert.equal(result.parent.groupId, result.child.groupId);
+  assert.equal(result.parent.residentBytes, 70 * KiB);
+  assert.deepEqual(result.parent.sharedRuntimeIds, ['child', 'parent']);
+});
+
+test('malformed residency is unknown even when process memory is known', async () => {
+  const { sampler: usage } = sampler({
+    psReader: async () => '101 1 20',
+    fetchImpl: async (_url, options) => {
+      assert.equal(options.redirect, 'error');
+      return response({ ok: true });
+    }
+  });
+  const result = await usage.sample({
+    runtimes: { a: { pid: 101, status: 'running', command: 'ollama', port: 8201 } },
+    runtimeIds: ['a']
+  });
+  assert.equal(result.a.residentBytes, 20 * KiB);
+  assert.equal(result.a.loadedModelIds, null);
+  assert.equal(result.a.residencyKnown, false);
+});
+
+test('a slow residency endpoint is aborted and becomes unknown', async () => {
+  const usage = createRuntimeMemoryUsageSampler({
+    psReader: async () => '',
+    listenerReader: async () => '',
+    modelTimeoutMs: 20,
+    scheduleTimeout: (callback, ms) => setTimeout(callback, ms),
+    fetchImpl: async (_url, { signal }) =>
+      new Promise((_, reject) => signal.addEventListener('abort', () => reject(signal.reason), { once: true }))
+  });
+  const result = await usage.sample({
+    runtimes: { a: { status: 'running', command: 'ollama', port: 8201 } },
+    runtimeIds: ['a']
+  });
+  assert.equal(result.a.residencyKnown, false);
+});
+
+test('routing status avoids dashboard-only sampling', async () => {
+  const { RuntimeManager } = await import('../src/runtime-manager.mjs');
+  let calls = 0;
+  const manager = new RuntimeManager(
+    { runtimes: { a: { enabled: false } }, models: [], cluster: {} },
+    {
+      memoryUsageSampler: {
+        sample: async () => {
+          calls++;
+          return { a: { residentBytes: 12 } };
+        }
+      }
+    }
+  );
+  await manager.status();
+  assert.equal(calls, 0);
+  const status = await manager.status({ includeMemoryUsage: true });
+  assert.equal(calls, 1);
+  assert.equal(status.runtimes.a.memoryUsage.residentBytes, 12);
+});
+
+test('Darwin footprint includes charged memory omitted by RSS and falls back cleanly', async () => {
+  const request = { runtimes: { a: { status: 'running', pid: 101 } }, runtimeIds: ['a'] };
+  const measured = sampler({
+    platform: 'darwin',
+    psReader: async () => '101 1 20',
+    footprintReader: async () => ({ 101: 8 * 1024 ** 3 })
+  });
+  const result = await measured.sampler.sample(request);
+  assert.equal(result.a.residentBytes, 8 * 1024 ** 3);
+  assert.equal(result.a.source, 'process-footprint');
+  const fallback = sampler({
+    platform: 'darwin',
+    psReader: async () => '101 1 20',
+    footprintReader: async () => {
+      throw Error('unavailable');
+    }
+  });
+  const absent = await fallback.sampler.sample(request);
+  assert.equal(absent.a.residentBytes, 20 * 1024);
+  assert.equal(absent.a.source, 'process-rss');
+});
+
+test('an endpoint that ignores abort cannot hold status forever', async () => {
+  const usage = createRuntimeMemoryUsageSampler({
+    psReader: async () => '',
+    listenerReader: async () => '',
+    modelTimeoutMs: 10,
+    fetchImpl: async () => new Promise(() => {})
+  });
+  const result = await usage.sample({
+    runtimes: { a: { status: 'running', command: 'ollama', port: 8201 } },
+    runtimeIds: ['a']
+  });
+  assert.equal(result.a.residencyKnown, false);
+});
+
+test('runtime status exposes inherited maintenance to the dashboard', async () => {
+  const { RuntimeManager } = await import('../src/runtime-manager.mjs');
+  const manager = new RuntimeManager({
+    cluster: {},
+    models: [],
+    runtimes: {
+      a: { enabled: false },
+      group: {
+        enabled: false,
+        placement: { mode: 'distributed', members: [{ runtime: 'a' }] },
+        maintenance: { state: 'suspended' }
+      }
+    }
+  });
+  const result = await manager.status();
+  assert.equal(result.runtimes.a.maintenance.state, 'suspended');
+});
diff --git a/test/runtime-policy.test.mjs b/test/runtime-policy.test.mjs
index 6f38003..e8bc7ba 100644
--- a/test/runtime-policy.test.mjs
+++ b/test/runtime-policy.test.mjs
@@ -1,5 +1,14 @@
 import assert from 'node:assert/strict';
-import { applyRuntimePolicyPlan, createRuntimePolicyPlan, runtimeAdmissionBlockers } from '../src/runtime-policy.mjs';
+import {
+  applyRuntimePolicyPlan as applyPolicy,
+  createRuntimePolicyPlan as createPolicy,
+  runtimeAdmissionBlockers
+} from '../src/runtime-policy.mjs';
+
+// Keep synthetic capacity tests independent of the developer machine's load.
+const profile = { totalMemoryGb: 128, availableMemoryGb: 128 };
+const createRuntimePolicyPlan = (config, options) => createPolicy(config, { profile, ...options });
+const applyRuntimePolicyPlan = (config, manager, options) => applyPolicy(config, manager, { profile, ...options });
 
 const runtimePolicyConfig = {
   runtimePolicy: {
@@ -248,6 +257,64 @@ const predictiveConfig = {
     requested: { enabled: true, memoryGb: 64 }
   }
 };
+
+const macReserveConfig = {
+  runtimePolicy: { enabled: true, autoEvict: false, reserveMemoryGb: 12 },
+  runtimes: { bonsai: { enabled: true, memoryGb: 16 } }
+};
+const macIdleStatus = { runtimes: { bonsai: { healthy: false, status: 'idle' } } };
+for (const runtimePolicy of [macReserveConfig.runtimePolicy, { enabled: true, memoryBudgetGb: 84 }]) {
+  const pressured = await createRuntimePolicyPlan(
+    { ...macReserveConfig, runtimePolicy },
+    {
+      requestedRuntimeId: 'bonsai',
+      status: macIdleStatus,
+      profile: { totalMemoryGb: 96, availableMemoryGb: 16 }
+    }
+  );
+  assert.equal(pressured.admission.predictive, true);
+  assert.equal(pressured.admission.projectedMemoryGb, 96);
+  assert.equal(pressured.admission.allowed, false, 'other applications count without a percentage policy');
+}
+const macEstimateFits = await createRuntimePolicyPlan(macReserveConfig, {
+  requestedRuntimeId: 'bonsai',
+  status: macIdleStatus,
+  profile: { totalMemoryGb: 96, availableMemoryGb: 35 }
+});
+assert.equal(macEstimateFits.admission.projectedMemoryGb, 77);
+assert.equal(macEstimateFits.admission.allowed, true, 'live startup guard must catch an underestimated load later');
+
+const clusteredReserve = await createRuntimePolicyPlan(
+  {
+    ...macReserveConfig,
+    cluster: { nodeId: 'mac', leaderNode: 'mac', nodes: { mac: { resources: { memoryGb: 96 } } } },
+    runtimes: { bonsai: { enabled: true, node: 'mac', memoryGb: 16 } }
+  },
+  {
+    requestedRuntimeId: 'bonsai',
+    requesterNode: 'mac',
+    status: {
+      ...macIdleStatus,
+      cluster: {
+        nodes: {
+          mac: {
+            local: true,
+            reachable: true,
+            telemetry: { memory: { totalBytes: 96 * 1024 ** 3, availableBytes: 16 * 1024 ** 3 } }
+          }
+        }
+      }
+    },
+    profile: { totalMemoryGb: 96, availableMemoryGb: 16 }
+  }
+);
+assert.equal(
+  clusteredReserve.admission.allowed,
+  false,
+  'local cluster node also counts host pressure without a percentage'
+);
+assert.equal(clusteredReserve.admission.nodes.mac.predictive, true);
+
 const predictiveStatus = {
   runtimes: {
     loaded: { healthy: true, status: 'running', activeRequests: 0 },
diff --git a/test/runtime-preferences.test.mjs b/test/runtime-preferences.test.mjs
new file mode 100644
index 0000000..0510193
--- /dev/null
+++ b/test/runtime-preferences.test.mjs
@@ -0,0 +1,223 @@
+import { test } from 'node:test';
+import assert from 'node:assert/strict';
+import fs from 'node:fs/promises';
+import os from 'node:os';
+import path from 'node:path';
+import { loadConfig } from '../src/config.mjs';
+import { mutateConfigSource } from '../src/config-mutation.mjs';
+import { RuntimeManager, reconfigureRuntimeIds } from '../src/runtime-manager.mjs';
+import { createRuntimePreferenceController, updateRuntimePreference } from '../src/runtime-preferences.mjs';
+
+const config = () => ({
+  runtimes: { chat: { enabled: true, command: 'unused', memoryGb: 8 } },
+  models: []
+});
+
+test('readiness changes do not restart a running runtime or erase its request state', async () => {
+  const previous = config();
+  const manager = new RuntimeManager(previous);
+  const current = manager.stateFor('chat');
+  current.activeRequests = 1;
+  manager.stop = () => {
+    throw new Error('must not stop a running request');
+  };
+  manager.admit = () => {
+    throw new Error('policy changes must not force a load');
+  };
+  for (const policy of ['always', 'preferred', 'auto']) {
+    const next = structuredClone(manager.config);
+    updateRuntimePreference(next, 'chat', policy);
+    assert.deepEqual(reconfigureRuntimeIds(manager.config, next), []);
+    const result = await manager.reconfigure(next);
+    assert.deepEqual(result.changed, []);
+    assert.equal(manager.stateFor('chat').activeRequests, 1);
+  }
+});
+
+test('unknown, external, disabled, maintained and non-owned runtimes cannot be updated', () => {
+  assert.throws(() => updateRuntimePreference(config(), '__proto__', 'auto'), /Unknown/);
+  assert.throws(() => updateRuntimePreference(config(), 'chat', 'other'), /Choose/);
+  for (const extra of [{ enabled: false }, { managed: false }, { management: 'external' }, { maintenance: {} }]) {
+    const raw = config();
+    Object.assign(raw.runtimes.chat, extra);
+    assert.throws(() => updateRuntimePreference(raw, 'chat', 'always'));
+  }
+  const raw = config();
+  raw.cluster = { nodeId: 'here' };
+  raw.runtimes.chat.node = 'elsewhere';
+  assert.throws(() => updateRuntimePreference(raw, 'chat', 'always'), /owns/);
+});
+
+test('readiness waits for active admission and preserves an independent pause', async () => {
+  const manager = new RuntimeManager(config());
+  manager.pausedRuntimes.add('chat');
+  let release;
+  let entered;
+  const started = new Promise((resolve) => (entered = resolve));
+  const barrier = new Promise((resolve) => (release = resolve));
+  const admission = manager.withAdmissionLock(
+    async (signal) => {
+      entered();
+      await barrier;
+      assert.equal(signal?.aborted ?? false, false);
+      assert.notEqual(manager.config.runtimes.chat.keepWarm, true);
+    },
+    { runtimeId: 'chat', preemptible: true }
+  );
+  await started;
+  const next = structuredClone(manager.config);
+  updateRuntimePreference(next, 'chat', 'always');
+  const reload = manager.reconfigure(next);
+  release();
+  await Promise.all([admission, reload]);
+  assert.equal(manager.config.runtimes.chat.keepWarm, true);
+  assert.equal(manager.pausedRuntimes.has('chat'), true);
+});
+
+test('concurrent persisted preferences preserve other fields and literal environment placeholders', async () => {
+  const dir = await fs.mkdtemp(path.join(os.tmpdir(), 'lloom-preferences-'));
+  try {
+    const file = path.join(dir, 'config.json');
+    const raw = config();
+    raw.runtimes.chat.evictable = false;
+    raw.runtimes.chat.policy = { evictable: false, priority: 77 };
+    raw.runtimes.chat.env = { SAMPLE: '${A_PRIVATE_ENV_VALUE}' };
+    await fs.writeFile(file, JSON.stringify(raw), { mode: 0o600 });
+    let current = await loadConfig(file);
+    let reloadCount = 0;
+    const controller = createRuntimePreferenceController({
+      getConfig: () => current,
+      mutateSource: (fn) => mutateConfigSource(current, fn),
+      reload: async () => {
+        current = await loadConfig(file);
+        reloadCount++;
+      }
+    });
+    await assert.rejects(controller.set('chat', { policy: 'auto' }), /confirm/);
+    assert.equal(reloadCount, 0);
+    const results = await Promise.all([
+      controller.set('chat', { policy: 'preferred', yes: true }),
+      controller.set('chat', { policy: 'auto', yes: true })
+    ]);
+    assert.deepEqual(
+      results.map((r) => r.policy),
+      ['preferred', 'auto']
+    );
+    const stored = JSON.parse(await fs.readFile(file, 'utf8'));
+    assert.equal(stored.runtimes.chat.keepWarm, false);
+    assert.equal(stored.runtimes.chat.preferredWarm, false);
+    assert.equal(stored.runtimes.chat.evictable, undefined);
+    assert.deepEqual(stored.runtimes.chat.policy, { priority: 77 });
+    assert.equal(stored.runtimes.chat.env.SAMPLE, '${A_PRIVATE_ENV_VALUE}');
+  } finally {
+    await fs.rm(dir, { recursive: true, force: true });
+  }
+});
+
+test('a failed reload can be retried without losing the saved preference', async () => {
+  const raw = config();
+  raw.sourcePath = '/test/config.json';
+  let fail = true;
+  const controller = createRuntimePreferenceController({
+    getConfig: () => raw,
+    mutateSource: async (fn) => {
+      fn(raw);
+      return { changed: false };
+    },
+    reload: async () => {
+      if (fail) throw new Error('reload failed');
+    }
+  });
+  await assert.rejects(controller.set('chat', { policy: 'preferred', yes: true }), /reload failed/);
+  fail = false;
+  assert.equal((await controller.set('chat', { policy: 'preferred', yes: true })).policy, 'preferred');
+});
+
+test('accepted readiness stays observable while admission is busy and rejects queued preferred restore', async () => {
+  const raw = config();
+  raw.sourcePath = '/test/config';
+  raw.runtimes.chat.preferredWarm = true;
+  const manager = new RuntimeManager(structuredClone(raw));
+  let finish;
+  const barrier = new Promise((resolve) => {
+    finish = resolve;
+  });
+  const controller = createRuntimePreferenceController({
+    getConfig: () => manager.config,
+    mutateSource: async (fn) => {
+      fn(raw);
+      return { changed: true };
+    },
+    reload: async () => {
+      await barrier;
+      await manager.reconfigure(structuredClone(raw));
+    },
+    onPersisted: (id, policy, generation) => manager.noteDesiredResidency(id, policy, generation),
+    onApplied: (id, policy, generation) => manager.settleDesiredResidency(id, policy, generation)
+  });
+  const result = await controller.request('chat', { policy: 'auto', yes: true });
+  assert.equal(result.status, 'pending');
+  assert.equal(manager.config.runtimes.chat.preferredWarm, true);
+  assert.equal(manager.resolveAdmissionGuard({ runtimeId: 'chat', reason: 'preferred-warm' }).ok, false);
+  assert.equal(controller.snapshot('chat').status, 'pending');
+  finish();
+  await new Promise((resolve) => setImmediate(resolve));
+  assert.equal(controller.snapshot('chat').status, 'succeeded');
+  assert.equal(manager.config.runtimes.chat.preferredWarm, false);
+  assert.equal(manager.desiredResidencyPolicy('chat'), undefined);
+});
+
+test('pending Always ready protects a resident model from admission eviction', async () => {
+  const { applyRuntimePolicyPlan } = await import('../src/runtime-policy.mjs');
+  const raw = {
+    runtimePolicy: { enabled: true, memoryBudgetGb: 40, protectActiveRequests: true },
+    runtimes: { resident: { enabled: true, memoryGb: 30 }, incoming: { enabled: true, memoryGb: 30 } }
+  };
+  const manager = new RuntimeManager(raw);
+  manager.noteDesiredResidency('resident', 'always', 1);
+  manager.status = async () => ({
+    runtimes: {
+      resident: { status: 'running', healthy: true, activeRequests: 0 },
+      incoming: { status: 'idle', healthy: false }
+    }
+  });
+  manager.stop = async () => {
+    throw new Error('must not evict pending pin');
+  };
+  await assert.rejects(
+    applyRuntimePolicyPlan(raw, manager, { requestedRuntimeId: 'incoming', dryRun: false, yes: true }),
+    (error) => error.code === 'runtime_keep_warm_conflict'
+  );
+  manager.noteDesiredResidency('resident', 'auto', 2);
+  manager.settleDesiredResidency('resident', 'always', 1);
+  assert.equal(manager.desiredResidencyPolicy('resident'), 'auto', 'older completion cannot erase newer intent');
+});
+
+test('hard-pin startup keeps admission enabled even when the general policy is disabled', async () => {
+  const { applyRuntimePolicyPlan } = await import('../src/runtime-policy.mjs');
+  const raw = {
+    runtimePolicy: { enabled: false, memoryBudgetGb: 40 },
+    runtimes: {
+      resident: { enabled: true, keepWarm: true, memoryGb: 30 },
+      incoming: { enabled: true, keepWarm: true, memoryGb: 30 }
+    }
+  };
+  const manager = new RuntimeManager(raw);
+  manager.status = async () => ({
+    runtimes: { resident: { healthy: true, status: 'running' }, incoming: { healthy: false, status: 'idle' } }
+  });
+  manager.start = async () => {
+    throw new Error('must not bypass admission');
+  };
+  await assert.rejects(
+    applyRuntimePolicyPlan(raw, manager, {
+      requestedRuntimeId: 'incoming',
+      reason: 'keep-warm',
+      dryRun: false,
+      yes: true
+    }),
+    (error) => error.code === 'runtime_keep_warm_conflict'
+  );
+  manager.noteDesiredResidency('incoming', 'auto', 1);
+  assert.equal(manager.resolveAdmissionGuard({ runtimeId: 'incoming', reason: 'keep-warm' }).ok, false);
+});
diff --git a/test/runtime-watchdog.test.mjs b/test/runtime-watchdog.test.mjs
index 8a938cb..d251fd9 100644
--- a/test/runtime-watchdog.test.mjs
+++ b/test/runtime-watchdog.test.mjs
@@ -398,6 +398,7 @@ for (const mode of ['observe', 'defer', 'force', 'disable-during-drain', 'recove
   const idleStall = observations.find((item) => item.stalled);
   assert.ok(idleStall, 'shorter idle budget detects silence after content');
   assert.equal(idleStall.hadContent, true);
+  assert.ok(idleStall.stallDurationMs >= 20, 'timer must not report a stall before its elapsed-time threshold');
   assert.ok(idleStall.stallDurationMs < idleStall.runtimeDurationMs);
   assert.equal(
     classifyRuntimeWatchdogOutcome(config.runtimes['watchdog-runtime'], idleStall).kind,
diff --git a/test/security.test.mjs b/test/security.test.mjs
index 5c650a4..f74aa8e 100644
--- a/test/security.test.mjs
+++ b/test/security.test.mjs
@@ -18,6 +18,9 @@ import {
 } from '../src/security.mjs';
 
 assert.equal(isLoopbackAddress('127.0.0.1'), true);
+assert.equal(isLoopbackAddress('127.2.3.4'), true);
+assert.equal(isLoopbackAddress('127.attacker.example'), false);
+assert.equal(isLoopbackAddress('127.1.example:8100'), false);
 assert.equal(isLoopbackAddress('::1'), true);
 assert.equal(isLoopbackAddress('localhost'), true);
 assert.equal(isLoopbackAddress('0.0.0.0'), false);
@@ -176,10 +179,10 @@ assert.equal(
 
 const remoteAdminAllowed = {
   ...remoteConfig,
-  security: { ...remoteConfig.security, allowRemoteAdmin: true }
+  security: { ...remoteConfig.security, allowRemoteAdmin: true, adminApiKeys: ['sk-admin-only'] }
 };
 assert.equal(
-  authorizeRequest({ headers: { authorization: 'Bearer sk-lloom-local' } }, remoteAdminAllowed, {
+  authorizeRequest({ headers: { authorization: 'Bearer sk-admin-only' } }, remoteAdminAllowed, {
     method: 'POST',
     pathname: '/gateway/setup/apply'
   }).ok,
@@ -193,6 +196,22 @@ assert.equal(
   false
 );
 
+assert.equal(
+  authorizeRequest({ headers: { authorization: 'Bearer sk-lloom-local' } }, remoteAdminAllowed, {
+    method: 'POST',
+    pathname: '/gateway/setup/apply'
+  }).ok,
+  false
+);
+assert.equal(
+  authorizeRequest(
+    { headers: { authorization: 'Bearer sk-lloom-local' } },
+    { ...remoteAdminAllowed, security: { ...remoteAdminAllowed.security, adminApiKeys: [] } },
+    { method: 'POST', pathname: '/gateway/setup/apply' }
+  ).code,
+  'admin_key_required'
+);
+
 assert.equal(hasValidApiKey({ headers: { authorization: 'Bearer sk-lloom-local' } }, loopbackConfig), true);
 assert.equal(hasValidApiKey({ headers: {} }, loopbackConfig), false);
 
diff --git a/test/smoke.mjs b/test/smoke.mjs
index 103e2d6..f213b99 100644
--- a/test/smoke.mjs
+++ b/test/smoke.mjs
@@ -1080,11 +1080,10 @@ const libraryCli = await runCommand(process.execPath, [
 const libraryJson = JSON.parse(libraryCli.stdout);
 assert.equal(libraryJson.index.id, 'lloom-community-recipes');
 assert.equal(libraryJson.recipes[0].id, 'apple-silicon-flux2-klein-4b');
-if (process.platform === 'darwin' && process.arch === 'arm64') {
-  assert.equal(libraryJson.selected.recipeId, 'apple-silicon-qwen36-35b-a3b-optiq');
-} else {
-  assert.equal(libraryJson.selected, null);
-}
+// Selection depends on the host's live memory and already-running backends.
+// Verify the ranked decision without requiring a particular local model lane.
+const compatibleLibraryCandidates = libraryJson.candidates.filter((candidate) => candidate.selectable);
+assert.equal(libraryJson.selected?.recipeId ?? null, compatibleLibraryCandidates[0]?.recipeId ?? null);
 const addModelCli = await runCommand(process.execPath, [
   path.join(process.cwd(), 'bin', 'lloom.mjs'),
   'add-model',
@@ -1928,7 +1927,7 @@ assert(helpCli.includes('lloom integrate'));
 assert(!helpCli.includes('recipe-submit <pack-file-or-url>'));
 const helpFlagCli = (await runCommand(process.execPath, [path.join(process.cwd(), 'bin', 'lloom.mjs'), '--help']))
   .stdout;
-assert(helpFlagCli.includes('Start installed LLooM; preview setup on first run'));
+assert(helpFlagCli.includes('Start installed LLooM; guided setup on first run'));
 const advancedHelpCli = (
   await runCommand(process.execPath, [path.join(process.cwd(), 'bin', 'lloom.mjs'), 'help', 'advanced'])
 ).stdout;
@@ -5960,11 +5959,8 @@ if (listened) {
     assert.equal(libraryResponse.status, 200);
     const libraryPlanJson = await libraryResponse.json();
     assert.equal(libraryPlanJson.index.id, 'lloom-community-recipes');
-    if (process.platform === 'darwin' && process.arch === 'arm64') {
-      assert.equal(libraryPlanJson.selected.recipeId, 'apple-silicon-qwen36-35b-a3b-optiq');
-    } else {
-      assert.equal(libraryPlanJson.selected, null);
-    }
+    const compatible = libraryPlanJson.candidates.filter((candidate) => candidate.selectable);
+    assert.equal(libraryPlanJson.selected?.recipeId ?? null, compatible[0]?.recipeId ?? null);
     assert.equal(
       libraryPlanJson.recipes.find((recipe) => recipe.id === 'apple-silicon-qwen36-35b-a3b-optiq')?.commands
         .installApply,
@@ -8686,7 +8682,11 @@ if (mockListened) {
       assert(metricsJson.host?.memory?.freeBytes >= 0);
       assert(metricsJson.host?.memory?.availableBytes >= 0);
       assert.equal(typeof metricsJson.host?.memory?.pressureUtilization, 'number');
-      assert(['linux-memavailable', 'macos-memory-pressure', 'os-freemem'].includes(metricsJson.host?.memory?.source));
+      assert(
+        ['linux-memavailable', 'macos-memory-pages', 'macos-memory-pressure', 'os-freemem'].includes(
+          metricsJson.host?.memory?.source
+        )
+      );
       assert.equal(typeof metricsJson.host?.cpu?.logicalCpus, 'number');
       const modelMetrics = metricsJson.models.find(
         (model) => model.id === 'Youssofal/Qwen3.6-27B-MTPLX-Optimized-Speed'