b-re-w · Oct 3, 2026 · Oct 3, 2026
diff --git a/PROJECT.md b/PROJECT.md
@@ -32,12 +32,13 @@ INTENT → SPEC 순으로 그쪽이 이깁니다.
 | 세션 API | 부분 (결함) | 오류를 `raise` 대신 `return`, 404 대신 500, 생성 시 모델 이름 미검증 |
 | Android · 데스크톱 · 웹 클라이언트 | 구현 | `app/build.gradle.kts` |
 | iOS | 부분 | Xcode 스크립트가 없는 `:composeApp` 모듈을 호출 |
-| 모델 선택 | 부분 | 클라이언트에 하드코딩, 서버 목록 미사용 |
+| 모델 선택, 서버 주소 | 구현, 테스트 있음 | 서버의 `GET /api/models` 를 읽음 (S2.5). 서버 주소 파싱과 API 키 포함 (S2.2). Android 는 주소 설정 화면이 아직 없음 |
 | 대화 기록 | 부분 | 메모리에만 |
 | 설정 화면, 도메인 계층(Clean Architecture) | 예정 | 빈 파일 |
 | 온디바이스 추론 · 오프라인 · 동기화 | 예정 | 코드 없음 |
-| 네이티브 데스크톱 (GraalVM) | 구현 (Windows x64), 미실행 | #71. 장기적으로 compose-multiplatform-extended 가 관리 |
-| Kotlin 테스트 | **없음** | `1 + 2 == 3` 하나뿐 |
+| 네이티브 데스크톱 (GraalVM) | 예정 (SPEC S3.4) | Windows x64 빌드 경로는 develop 에 있으나(#71) 네이티브 빌드는 실행해 본 적이 없다. 장기적으로 compose-multiplatform-extended 가 관리 |
+| Kotlin 테스트 | 부분 | 서버 주소, 모델 목록 파싱, 종료 코드 매핑은 테스트 있음(CI). 화면과 뷰모델은 테스트 없음 |
+| GPU(CUDA) 추론 | 예정 (M4) | 엔진 기본값은 `device="cpu"`. torchnative CUDA 는 실행해 본 적이 없다 |
 
 ## 3. 구조
 

diff --git a/README.md b/README.md
@@ -30,7 +30,7 @@ Most AI chat apps are a thin window onto someone else's server. Gemstone starts
 
 Today that means a Kotlin [Compose Multiplatform](https://www.jetbrains.com/compose-multiplatform/)
 client for Android, iOS, desktop and the web, talking to a Python server you run yourself on your
-own GPU. Next, the same Python model code moves *into* the app through
+own machine (the engine runs on the CPU for now; GPU support is planned). Next, the same Python model code moves *into* the app through
 [Python Multiplatform](https://github.com/thisisthepy/python-multiplatform), so chatting needs no
 separate server and no network at all.
 
@@ -40,7 +40,7 @@ separate server and no network at all.
 - 🚀 **Streaming chat** — tokens arrive over a WebSocket as they are generated and render as Markdown on the fly.
 - 🧠 **Visible reasoning** — the model's `<think>` block appears as a collapsible panel with elapsed time.
 - 🔌 **Tool calling** — weather, public holidays, exchange rates, a calculator and web search, run server-side in parallel and fed back to the model.
-- 📦 **Open-weight models on one engine** — Qwen 3 0.6B served by a single transformers-based engine that runs on PyTorch or on [torchnative](https://github.com/thisisthepy/torchnative). Gemstone is becoming a local Ollama replacement: concurrent requests (continuous batching), OpenAI- and Ollama-compatible APIs and 8-bit weights are planned for November 2026.
+- 📦 **Open-weight models on one engine** — Qwen 3 0.6B served by a single transformers-based engine that runs on PyTorch or on [torchnative](https://github.com/thisisthepy/torchnative). Gemstone is becoming a local Ollama replacement: concurrent requests (continuous batching with a paged KV cache) and OpenAI- and Ollama-compatible APIs work today; 4-bit weights and GGUF are planned.
 - 🖥️ **Native desktop feel** — a JetBrains Jewel decorated window, with Dmg / Msi / Deb installers.
 
 ## 🚀 Quick start
@@ -152,16 +152,22 @@ Gemstone is an early, working prototype. The honest state of each piece:
 | Area | State |
 |---|---|
 | Streaming chat, reasoning display, tool calling | ✅ Working |
+| One transformers engine, continuous batching, paged KV cache (on CPU by default) | ✅ Working on PyTorch; 🟡 not yet verified on torchnative |
+| Several resident models, eviction, `keep_alive` | ✅ Working |
+| OpenAI-compatible API (`/v1`, tool calls passed through) | 🟡 Partial — no `n` > 1, logprobs or embeddings |
+| Ollama-compatible API (chat, generate, tags, show, pull, delete, ps, copy, create, embed) | 🟡 Partial — `push` returns 501; no Modelfile, template or blobs |
+| Security defaults (loopback bind, origin checks, API key) | ✅ Working |
 | Android, desktop and web clients | ✅ Working |
+| Model choice and server address in the client | ✅ Models are read from the server; Android has no address setting yet |
 | iOS client | 🟡 Framework targets configured; the Xcode build script needs fixing |
-| Model choice in the client | 🟡 Hard-coded list, not yet read from the server |
 | Chat history | 🟡 In memory only |
+| Automated tests | 🟡 Python tests (including a real model) and Kotlin tests for the network layer run in CI; the UI is untested |
+| Native desktop executable (GraalVM, no JVM) | ⏳ Planned — a Windows x64 build path exists but has not been run |
+| GPU (CUDA) inference | ⏳ Planned |
+| 4-bit weights, GGUF | ⏳ Planned |
 | Settings screen | ⏳ Planned |
 | On-device inference via Python Multiplatform | ⏳ Planned |
 | Offline mode, cross-device sync | ⏳ Planned |
-| OpenAI-compatible API | ⏳ Planned |
-| Native desktop executable (GraalVM, no JVM) | ⏳ In progress |
-| Automated tests | ⏳ Not yet — the first priority |
 
 ## 📖 Documentation
 

diff --git a/docs/SPEC.md b/docs/SPEC.md
@@ -16,13 +16,14 @@ a status:
 OpenAI-compatible API (S1.10), the Ollama-compatible API with its model management (S1.14,
 S1.15), the residency shared by all three, with several resident models and derived models
 (S1.14, `test_multi_resident.py`), the tool-result cache (S1.6) and the backend
-removal (S1.8). Everything else has **no
-behavioural test** yet: the only Kotlin test,
-`app/src/commonTest/kotlin/gemstone/ComposeAppCommonTest.kt`, asserts `1 + 2 == 3`. So
-`implemented` without a named test means *read in code*, never *verified by a test*.
+removal (S1.8), plus batching, queue limits, first-use download, errors, tool robustness, security
+and client compatibility (each item names its module). Real-model tests (marker `real_model`) run in
+CI. Kotlin tests in `app/src/commonTest` cover the network layer (`ModelCatalogTest`,
+`ServerAddressTest`, `ChatCloseEventTest`); the UI has none. `implemented` without a named test means
+*read in code*, never *verified by a test*.
 
-Status was established by reading the `develop` branch at commit `2be0e37` (2026-10-02). S1.11–S1.16
-and S3.4 also record the maintainer's decisions of the same day ([`serving/engine.md`](serving/engine.md)).
+Status was first established by reading `develop` at commit `2be0e37` (2026-10-02) and is updated
+with every change since. S1.11–S1.16 and S3.4 also record the maintainer's decisions of that day ([`serving/engine.md`](serving/engine.md)).
 
 ---
 
@@ -274,7 +275,8 @@ Decided (2026-10-02):
 Implemented (#84), `api/src/main/engine.py`:
 
 - `Engine(model_id, dtype=, device=, quantization=, chat_template=, ...)` loads a causal LM with
-  `from_pretrained`. `quantization="q8_0"` goes through torchnative's `TorchnativeConfig`.
+  `from_pretrained`. `quantization="q8_0"` goes through torchnative's `TorchnativeConfig`; this is the
+  November weight format but is unverified until torchnative's #15 gate measures q8_0 with batching.
 - Calling it with chat messages streams the reply as text chunks. Temperature 0 is greedy and
   equals `transformers` `generate`; `seed` makes sampling reproducible.
 - Concurrent callers are batched (S1.12). A request the batch cannot serve takes the exclusive
@@ -307,14 +309,12 @@ timeout and its 503 / 1013 mapping), `api/tests/test_engine.py`, `api/tests/test
 SmolLM2-135M. Still open for M1: the same tests on torchnative (cpu, mps) once torchnative TN-M1
 lands (2026-10-24).
 
-Proposal, pending confirmation: the engine is transformers 5.x continuous batching with a paged KV
+The 2026-10-02 proposal, now implemented (#85, #91) and awaiting verification on torchnative: the engine is transformers 5.x continuous batching with a paged KV
 cache, kernels live in torchnative, and the engine uses only the public `torch` / `transformers`
 API so a GPU server can run the same code on upstream PyTorch. Constraint: custom kernels cannot be
 registered from Gemstone, because `torch.library` registrations are no-ops on torchnative. They
 must be torchnative operators. See [`serving/engine.md`](serving/engine.md).
 
-No code on `develop`. Test: none.
-
 ### S1.12 Continuous batching — `partial` · G5
 
 Concurrent requests to a loaded model are scheduled into one running batch. A request joins and