Reduce voice service latency and record the first successful payload smoke

Constraint: the voice service must return a payload under the current CPU environment before we can iterate on business-corpus correctness
Rejected: keeping the previous multi-chunk defaults | they caused smoke-timeout regressions and blocked basic endpoint validation
Confidence: high
Scope-risk: moderate
Directive: treat the current result as transport/runtime proof only until the service is switched from synthetic defaults to the /workspace business reference corpus
Tested: /usr/local/miniconda3/bin/python -m unittest discover -s acr-engine/tests -v; /usr/local/miniconda3/bin/python acr-engine/scripts/service_voice_smoke.py -> status ok, chunk_count=1, top_song_id=song_0022, has_context=false
Not-tested: business-corpus song_id correctness for /recognize/voice under /workspace reference data

@@ -18,7 +18,7 @@ def post_multipart(url: str, file_path: Path):
         f'Content-Disposition: form-data; name="file"; filename="{file_path.name}"\r\n'
         f'Content-Type: audio/wav\r\n\r\n'
     ).encode('utf-8') + data + f'\r\n--{boundary}--\r\n'.encode('utf-8')
-    req = Request(url, data=body, method='POST')
+    req = Request(url + '?top_n=1&max_chunks=1&include_context=false', data=body, method='POST')
     req.add_header('Content-Type', f'multipart/form-data; boundary={boundary}')
     with urlopen(req, timeout=20) as resp:
         return json.loads(resp.read().decode('utf-8'))

@@ -26,7 +26,7 @@ def detect_voiced_intervals(y: np.ndarray, sr: int, top_db: int = 30, min_voiced
     return kept
 
 
-def chunk_intervals(intervals: List[tuple[int, int]], sr: int, target_chunk_sec: float = 8.0, stride_sec: float = 4.0) -> List[tuple[int, int, bool]]:
+def chunk_intervals(intervals: List[tuple[int, int]], sr: int, target_chunk_sec: float = 8.0, stride_sec: float = 4.0, max_chunks: int = 3) -> List[tuple[int, int, bool]]:
     chunk_len = int(sr * target_chunk_sec)
     stride = int(sr * stride_sec)
     chunks: List[tuple[int, int, bool]] = []
@@ -49,6 +49,8 @@ def chunk_intervals(intervals: List[tuple[int, int]], sr: int, target_chunk_sec:
         if key not in seen:
             deduped.append(item)
             seen.add(key)
+    if max_chunks > 0 and len(deduped) > max_chunks:
+        return deduped[:max_chunks]
     return deduped
 
 
@@ -78,10 +80,10 @@ def write_chunks(y: np.ndarray, sr: int, chunks: List[tuple[int, int, bool]], ou
     return results
 
 
-def voice_to_chunks(audio_path: str, output_dir: str, target_chunk_sec: float = 8.0, stride_sec: float = 4.0, min_voiced_sec: float = 2.0, top_db: int = 30, sr: int = 16000) -> List[Dict]:
+def voice_to_chunks(audio_path: str, output_dir: str, target_chunk_sec: float = 8.0, stride_sec: float = 4.0, min_voiced_sec: float = 2.0, top_db: int = 30, sr: int = 16000, max_chunks: int = 3) -> List[Dict]:
     y = normalize_audio(audio_path, sr=sr)
     intervals = detect_voiced_intervals(y, sr=sr, top_db=top_db, min_voiced_sec=min_voiced_sec)
-    chunks = chunk_intervals(intervals, sr=sr, target_chunk_sec=target_chunk_sec, stride_sec=stride_sec)
+    chunks = chunk_intervals(intervals, sr=sr, target_chunk_sec=target_chunk_sec, stride_sec=stride_sec, max_chunks=max_chunks)
     return write_chunks(y, sr, chunks, output_dir, source_audio_path=audio_path)
 
 
@@ -94,6 +96,7 @@ def main() -> None:
     ap.add_argument('--min-voiced-sec', type=float, default=2.0)
     ap.add_argument('--top-db', type=int, default=30)
     ap.add_argument('--sr', type=int, default=16000)
+    ap.add_argument('--max-chunks', type=int, default=3)
     ap.add_argument('--output-json', default='chunks.json')
     args = ap.parse_args()
     chunks = voice_to_chunks(
@@ -104,6 +107,7 @@ def main() -> None:
         min_voiced_sec=args.min_voiced_sec,
         top_db=args.top_db,
         sr=args.sr,
+        max_chunks=args.max_chunks,
     )
     out_json = Path(args.output_dir) / args.output_json
     out_json.write_text(json.dumps({'chunks': chunks}, ensure_ascii=False, indent=2), encoding='utf-8')

@@ -213,6 +213,8 @@ async def recognize_voice(
     device: Optional[str] = None,
     context_sec: float = 10.0,
     output_format: str = 'mp3',
+    max_chunks: int = 3,
+    include_context: bool = True,
 ):
     resolved = _resolve(data_dir, model_path, index_prefix, device)
     engine, cache_hit = _load_engine(**resolved)
@@ -223,7 +225,7 @@ async def recognize_voice(
         raw_path.write_bytes(await file.read())
 
         chunk_dir = tmp / 'chunks'
-        chunks = voice_to_chunks(str(raw_path), str(chunk_dir))
+        chunks = voice_to_chunks(str(raw_path), str(chunk_dir), max_chunks=max_chunks)
         if not chunks:
             raise HTTPException(status_code=400, detail='No voiced chunks detected from input audio')
 
@@ -242,7 +244,7 @@ async def recognize_voice(
             song_id = item['song_id']
             ref_audio = _reference_audio_for_song(engine, song_id)
             context_info = None
-            if ref_audio and item['best_chunk'] is not None:
+            if include_context and ref_audio and item['best_chunk'] is not None:
                 match = find_best_matching_window(
                     query_audio_path=item['best_chunk']['chunk']['audio_path'],
                     reference_audio_path=ref_audio,

@@ -24,7 +24,7 @@ flowchart TD
 | benchmark report 已生成 |  |
 | model card 已生成 |  |
 | license registry 已更新 |  |
-| service smoke test 通过 | partial: `/health` OK, CPU torch installed, `/recognize/voice` end-to-end smoke still times out |
+| service smoke test 通过 | partial: `/health` OK, `/recognize/voice` payload returns, but still bound to synthetic service index rather than business reference corpus |
 | dataset whitelist 已确认 |  |
 | changelog 已更新 | yes |
 | architect review completed | yes (approved with watch) |

@@ -30,7 +30,7 @@
   - `acr-engine/src/service/app.py` 已新增 `POST /recognize/voice`
   - `/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
-  - 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 `POST /recognize/voice` 已跨过依赖缺失与超时阶段：CPU 版 `torch` 已安装、`uvicorn` / `fastapi` / `python-multipart` 已安装、`/health` 可返回 `ok`，voice smoke 已返回 payload（`chunk_count=1`, `top_song_id=song_0022`, `has_context=false`）；当前剩余问题是服务默认仍绑定 synthetic 索引语义，尚未切到 `/workspace` 业务曲库 reference
 - 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