test: qualify 3.5.2 SPA with isolated services [skip deploy]

docs/evidence/spa-3.5.2/README.md

@@ -0,0 +1,187 @@
+# Published 3.5.2 full-SPA follow-up
+
+The six post-login full-SPA recovery cases passed **6/6, zero failures/retries/skips**
+on product `ba9014958f7ab0dc0636a18edd1339d9e9e76e60` and published
+`@lyyzka/lingxios@3.5.2` / schema12. This closes the previously unverified **3.5.2
+SPA session/human-message recovery** portion. It does not qualify captcha login,
+live model answers, a complete sandboxed product Worker, or 3.5.2 durability.
+
+The run uses the newly built actual React application, a separate local control
+Worker and D1 session store, the signed gateway, separate PostgreSQL/Redis,
+real pinned WuKongIM and separate storage. It does not use a component recovery
+page or mock successful API responses. Synthetic post-login sessions use the
+existing test secret; ordinary session validation and authorization stay active.
+
+| Full SPA scenario | Observed result |
+| --- | --- |
+| Session restore/reload/expire/renew | Session200 →401 → restored; protected workspace hidden while expired |
+| Network fails before acceptance | One original acceptance/message and one next acceptance/message; draft survives |
+| Independent PostgreSQL container pause/unpause | Same one/one durable counts; draft survives; recovery without page refresh |
+| Independent Web process stop/start | Same one/one durable counts; lifecycle records confirm real stop/start/readiness |
+| Send rejected by expired session | Old send remains at zero acceptances/messages after renewal; next send succeeds once |
+| Typing401 races the send response | Old send remains unsent after login renewal; next send succeeds once |
+
+[Report](full-spa/report.json), [counts](full-spa/counts), [13 screenshots](full-spa/screenshots),
+[6 videos](full-spa/video), [Web lifecycle](web-lifecycle.jsonl),
+[build hashes](build-manifest.json) and [checks/log hashes](checks.json) retain
+repeatable evidence. Screenshot names include the case hash so the two expired
+session screenshots cannot overwrite each other. Request events/cookies/traces
+and raw response bodies are omitted.
+
+The runner captured the product commit while the prewritten isolated harness
+was still uncommitted at startup. The unchanged executed harness is now commit
+`ac7f9d4b2e5179de6c438ac8c5db9ae9b5d3fb15`; its dirty-start VCS fact remains in
+the report. No product files changed between the qualified main and this run.
+The SPA is freshly built from that product tree; its index and all asset hashes
+are retained, rather than reusing the earlier3.5.1 build.
+
+## Isolation and checks
+
+The explicit `E2E_SPA_ISOLATED=1` profile has fixed disposable endpoints:
+Web5280, API5281, runtime5282, supervisor5285, control8897, PG55433,
+Redis56380, WuKong55002/55201 and storage59001. D1 uses
+`.e2e/spa352-control-state`. The compose project is `lingxiloop-spa352`.
+[Container identities/images/bindings](isolated-services.json) were captured.
+
+The old helper would pause `lingxiloop-e2e-postgres-1`, which the original soak
+uses. Two safety tests first failed; the new profile now verifies the exact new
+container's label and loopback binding before pause/unpause, and rejects unknown
+profile selections. Safety2/2, existing fixture safety3/3 and E2E typecheck passed.
+The new compose services and API/control/preview processes were stopped after
+qualification to release this executor's disk space. No original soak service
+was stopped, paused, restarted, relabelled or reseeded.
+
+## Complete Worker boundary: concrete failure
+
+[The current preflight receipt](worker-preflight.json) binds public
+`releaseVersions`3.5.2/schema12 and executes the same bubblewrap command required
+by `server/src/agent-runtime/runtime.ts:startLingxiOSWorker`:
+
+```sh
+bwrap --die-with-parent --unshare-all --new-session --ro-bind / / \
+  --proc /proc --dev /dev -- /usr/bin/true
+```
+
+It exited1 with **`bwrap: setting up uid map: Read-only file system`** on kernel
+6.18.55. This cloud executor cannot satisfy the product's required production
+Worker preflight. No privilege, namespace/sysctl, AppArmor, seccomp or production
+configuration was changed. A complete product Worker was not started, and no
+model call was made. This receipt is a host preflight, not a claimed immutable
+3.5.2 Worker-image run. A second, disposable no-network container probe also failed with
+`No permissions to create new namespace`. It used the already-cached base
+image with the exact registry3.5.2 package mounted read-only, read only its public
+version export, and ran the identical preflight under the image's default UID0
+without additional privileges. The original soak container was not used. This
+corroborates the container boundary without calling the mixed diagnostic image
+an immutable3.5.2 release image. The upstream confined-image CI gate is separate
+evidence.
+
+The successful six SPA cases do not require generation: session validation,
+signed API access, human-message acceptance, actual WuKong durability, draft
+preservation and retry/session control can be tested independently. Full Worker
+restart after an action, model-assisted complete-SPA conversations and live
+provider behavior remain outside this executor's qualified results.
+
+## Four-issue code review and acceptance mapping
+
+Read-only review of product main `ba901495…` and upstream main
+`e09fc03437804ec80bc2f6d6b3258dc3e2f35540` found **no additional product/runtime
+code gap in the requested scope**. The follow-up changes only the acceptance
+harness and evidence. This finding does not turn missing environmental or
+long-duration evidence into a pass.
+
+Reviewed ownership: product `server/src/agent-runtime/runtime.ts`,
+`server/src/db/pool.ts`, shared `server/src/llm.ts`, native authorization,
+terminal delivery, published model recovery/execution/embedding/fallback,
+`src/model/child-admission.ts`, `PgModelBudgetStore.reserve/record`, terminal
+schema/outbox/read APIs and background scheduling. The verified package uses
+atomic root-budget locking, startup/closing holds, one-time usage settlement,
+current authorization/request-version checks, stable terminal identity and
+`suppressAgentWake`; failed/cancelled states are not rewritten as success.
+
+Earlier evidence below is versioned explicitly. Upstream
+[release39015169232](https://github.com/LingXi-Org/LingxiOS/actions/runs/39015169232)
+and [publication39015407648](https://github.com/LingXi-Org/LingxiOS/actions/runs/39015407648)
+qualified3.5.2 with422/422; product
+[main39016358468](https://github.com/LingXi-Org/LingxiLoop/actions/runs/39016358468)
+passed245 integration checks with21 prerequisite skips. The
+[3.5.2 development evidence](../runtime-development/README.md) records the11/11
+runtime browser fixture cases,17 owning integration passes and20/20 dedicated
+real-WuKong checks. Those11 cases use a synthetic model and fixture login and
+are not relabelled as complete-SPA/model-provider execution.
+
+| Issue / original acceptance item | Concrete coverage | Remaining condition, not a new code fix |
+| --- | --- | --- |
+| #153.1 idle/load/profile/24h | Published scheduler tests cover bounded background connections, cancellation, coalesced notifications, backoff, passive roles and periodic recovery. Prior paired image measurement is3.5.1 Web/API-only. | Qualified Worker/model-dialogue load and retained-heap measurement need an authorized suitable executor/provider; complete24h duration is still outstanding. There is no3.5.2 endurance result. |
+| #153.2 Web/Worker/DB fault E2E | This3.5.2 full SPA passes real independent Web stop/start and PG pause/unpause with durable-count and continuation assertions. | Complete sandboxed product Worker restart remains blocked by the recorded preflight. |
+| #153.3 health/operations | Readiness/liveness, pool/timeouts and gateway retry boundaries reviewed; owning tests remain qualified by the3.5.2 product CI. | Komodo entry repair, deployment/update-state and production rollout were explicitly removed by the user; they are not development blockers or implied passes. |
+| #155.1 provider fault matrix/bounds/ledger | Exact3.5.2 upstream422 gate covers classification, recovery/circuits, cancellation and metadata. Product timeout/total-deadline integration proves bounded requests, disconnection, settled usage, one terminal notice and successful next turn. | Controlled real-provider behavior requires existing authorized endpoint/model/key, prices and explicit spend/call limits. |
+| #155.2 fallback/embedding | Upstream fallback tests require current authorization, compatible capability and matching price; no attempt reset. Embedding quota/auth circuits preserve vector-space identity. | No approved live secondary route/provider is configured; none was created or enabled. |
+| #155.3 runtime+browser+real provider |3.5.2 product owning tests and11-case runtime fixture passed; this follow-up adds6 complete-SPA session/send cases. | Actual provider-assisted full-SPA conversation is not claimed; external provider configuration and qualified Worker are still required. |
+| #157.1 token/cost/deadline/concurrent admission | Atomic budget/closing holds, pending-fanout bound, rollback and settled near-limit cost/idempotency tests pass in3.5.2. | Controlled settlement uses deterministic usage/prices, not an actual supplier bill; live near-cost-limit acceptance remains external. |
+| #157.2 one root notification/results preserved | Runtime browser fanout fixture rejects two starts, creates zero child tasks, preserves the successful root answer and emits one aggregate notice without Agent wake. | No additional code blocker found; do not infer live billing from this fixture. |
+| #157.3 group mentions/new human turn | Existing group-mention integration and3.5.2 runtime browser continuation pass with no active runs afterwards. | Real multi-agent provider run needs the same provider/pricing/Worker prerequisites above. |
+| #161.1 fault matrix and durable replay |3.5.2 package/product tests cover budget/context/provider/authorization/recovery/fencing;20 real-WuKong cases retain one send/message, same result identity and unchanged action/model ledger. This SPA adds actual reconnect/session/Web/DB recovery. | Full product Worker crash-after-action and controlled real-provider matrix require a suitable authorized cloud executor. |
+| #161.2 partial/blocked/cancel/revocation | Terminal feedback remains separate from results; request/fence/audience checks and revoked-delivery tests pass. Runtime browser cancellation has no extra model call and permits the next request. | No additional code gap found in this review. |
+| #161.3 historical regressions | Package lease-recovery bounds and inherited evidence, product receipt reconciliation/FK regression remain covered; no old defect is reported as newly reproduced. | No additional issue-specific code gap found. |
+| #161.4 public API+browser+provider | Exact public3.5.2 package, product CI, runtime fixture11/11 and this complete-SPA6/6 have retained artifacts. | Real provider and complete Worker isolation remain unqualified; captcha/login itself is also excluded from these post-login tests. |
+
+[Provider prerequisites](provider-prerequisites.json) records presence checks
+only: none of the required provider/model/key/input/output-price environment
+variables was supplied. No values, secrets or prompts were exported. Open
+issues153/155/157/161 therefore retain **external configuration/environment and
+observation-window acceptance conditions**, not an identified unimplemented
+product repair. The3.5.2 durability exercise is unperformed, not silently passed.
+
+## Existing soak and independent follow-up
+
+[Isolation receipt](soak-isolation.json) confirms the original image start time,
+restart count0, unpaused state, unchanged observation start and zero violations
+after this suite. That observation remains **3.5.1 Web/API-only**, started
+2026-10-17T06:42:47.900Z, earliest completion2026-10-18T06:42:47.901Z.
+It cannot qualify3.5.2, model dialogue or a production Worker.
+
+Monitor285092 writes state/heartbeats/terminal receipt. Evidence watcher363050
+pushes initial/hourly/final snapshots to
+[codex/runtime-soak-evidence](https://github.com/LingXi-Org/LingxiLoop/tree/codex/runtime-soak-evidence/docs/evidence/runtime-measurements/soak)
+and reports terminal/stale failures. A separate existing ChatGPT Automation,
+**LingxiLoop soak 复查**, is enabled for **2026-10-18 06:50UTC**, with exact one-time
+schedule `DTSTART;TZID=Etc/UTC:20261018T065000`. It reads public remote state,
+handoff, heartbeats and terminal receipt independently of the shell watcher.
+The existing task identifier was verified and reported privately to the parent;
+no second task was created. Its schedule/enabled flag are saved; execution or
+notification delivery has not yet occurred. Missing/stale evidence or an absent
+terminal receipt must be reported incomplete. It performs no production action.
+
+## Reproduce
+
+Use Node22, the exact published dependency and existing synthetic test-service
+images (override `E2E_WUKONG_IMAGE`/`E2E_STORAGE_IMAGE` only with the matching
+already-authorized images). No developer/provider dotenv is loaded in this run.
+
+```sh
+docker compose -f e2e/acceptance/compose.yaml up -d --wait
+E2E_SPA_ISOLATED=1 node e2e/local-api.mjs --seed-only
+E2E_SPA_ISOLATED=1 node e2e/admin/start-control.mjs
+# In another terminal:
+E2E_SPA_ISOLATED=1 ACCEPTANCE_FIXTURE_ROOT="$PWD" node scripts/runtime-acceptance/spa-web-supervisor.mjs
+# Build, then preview (another terminal):
+VITE_TURNSTILE_SITE_KEY=1x00000000000000000000AA \
+LINGXILOOP_DEV_API_TARGET=http://127.0.0.1:8897 LINGXILOOP_DEV_WS_TARGET=ws://127.0.0.1:5281 \
+node node_modules/vite/bin/vite.js build --outDir .e2e/spa352-dist
+LINGXILOOP_DEV_API_TARGET=http://127.0.0.1:8897 LINGXILOOP_DEV_WS_TARGET=ws://127.0.0.1:5281 \
+node node_modules/vite/bin/vite.js preview --outDir .e2e/spa352-dist --host 127.0.0.1 --port 5280 --strictPort
+# Official engine:
+E2E_SPA_ISOLATED=1 ACCEPTANCE_FIXTURE_ROOT="$PWD" E2E_BASE_URL=http://127.0.0.1:5280 \
+E2E_BUILD_INDEX=.e2e/spa352-dist/index.html PLAYWRIGHT_BROWSERS_PATH=/workspace/.lingxi-setup/cache/ms-playwright \
+npm run test:e2e:web -- --config e2e.acceptance.config.ts --output .e2e/full-spa-acceptance --video on
+node --test e2e/acceptance/isolation-safety.test.mjs e2e/server/fixture-safety.test.mjs
+npm run test:e2e:typecheck
+```
+
+Stop only the new profile's API/control/preview processes, then run
+`docker compose -f e2e/acceptance/compose.yaml down`. Never stop the original
+soak project or use a shared PostgreSQL pause for this selection. The scoped
+repository CI may skip product suites for harness/docs-only changes; the cloud
+commands and retained results above are the actual owning checks, not a claim
+that skipped CI jobs executed the browser suite.