Refresh session handoff for the current voice service runtime state

Constraint: the handoff must reflect the real runtime state: health endpoints work, CPU torch is installed, but end-to-end voice smoke still times out
Rejected: keeping the older dependency-missing note | it no longer matches the current environment and would mislead the next session
Confidence: high
Scope-risk: narrow
Directive: keep handoff notes focused on the shortest next debugging path for /recognize/voice timeout reduction
Tested: /usr/local/miniconda3/bin/python -m unittest discover -s acr-engine/tests -v; /health endpoint returns ok under uvicorn with CPU torch installed
Not-tested: successful end-to-end /recognize/voice payload within timeout

@@ -30,7 +30,7 @@
   - `acr-engine/src/service/app.py` 已新增 `POST /recognize/voice`
   - `acr-engine/src/service/app.py` 已新增 `POST /recognize/voice`
   - `/health` 可正常启动并返回 `ok`
   - `/health` 可正常启动并返回 `ok`
   - architect review: approved with watch；当前 split（本地 FAISS / 可选 ChromaDB / 生产 pgvector）方向成立
   - architect review: approved with watch；当前 split（本地 FAISS / 可选 ChromaDB / 生产 pgvector）方向成立
-  - 当前 `POST /recognize/voice` 仍需完整 smoke 复核；此前明确阻塞点是 `torch` 缺失，现已补 CPU 版 `torch`，当前剩余现象是 voice 端到端 smoke 超时，需要继续收敛推理耗时或缓存策略
+  - 当前 `POST /recognize/voice` 已跨过依赖缺失阶段：CPU 版 `torch` 已安装、`uvicorn` / `fastapi` / `python-multipart` 已安装、`/health` 可返回 `ok`；当前剩余现象是 voice 端到端 smoke 仍超时，需要继续收敛推理耗时、chunk 数量或缓存策略
 - 当前 docs 已做第一轮简化：
 - 当前 docs 已做第一轮简化：
   - `docs/README.md` 只保留最新架构与最短阅读顺序
   - `docs/README.md` 只保留最新架构与最短阅读顺序
 
 
@@ -46,6 +46,28 @@
 3. 把哼唱评测集接入 `evaluate.py` 或独立评测脚本
 3. 把哼唱评测集接入 `evaluate.py` 或独立评测脚本
 4. 继续做 docs 第二轮收敛，只保留当前有效主文档
 4. 继续做 docs 第二轮收敛，只保留当前有效主文档
 
 
+### 最新补充（2026-06-03 voice service runtime）
+
+- 已确认当前解释器 `/usr/local/miniconda3/bin/python` 下：
+  - `torch==2.3.1+cpu`
+  - `uvicorn==0.48.0`
+  - `fastapi==0.136.3`
+  - `python-multipart==0.0.30`
+- `acr-engine/src/service/app.py` 当前可启动并通过：
+  - `GET /health`
+  - `GET /ready`
+- `POST /recognize/voice` 当前不再报缺依赖，但端到端 smoke 仍会超时。
+- 当前最可能的下一步排查点：
+  1. `voice_to_chunks` 默认切出的 chunk 数过多（当前样例可到 9 个）
+  2. 首次 `_load_engine()` + `engine.recognize()` CPU 推理耗时过长
+  3. `context_exporter` 在每个候选上再次做 reference 滑窗，进一步拉长总耗时
+- 下一 session 建议直接从：
+  - `acr-engine/src/service/app.py`
+  - `acr-engine/scripts/service_voice_smoke.py`
+  - `acr-engine/src/data/voice_chunker.py`
+  - `acr-engine/src/utils/context_exporter.py`
+  继续收敛超时问题。
+
 ### 最新交付快照（2026-06-02 16:12 UTC）
 ### 最新交付快照（2026-06-02 16:12 UTC）
 
 
 - 当前状态：先交付，后续重启继续
 - 当前状态：先交付，后续重启继续