notes(asr): 切片串行 ASR 绕单文件大小限制

ASR server 直接 500 拒绝大文件 (15MB / ~15min 4.7s 即返回 500)，不是
处理超时。改成：sidecar 装 ffmpeg → /transcribe endpoint 把音频切 60s
段 → 串行调外部 ASR → 拼接 transcript。notes 主容器 call_asr 改成 POST
到 sidecar /transcribe（timeout 1h 给长录音留余地）。

- feishu sidecar Dockerfile + ffmpeg + requests
- server.py 加 TranscribeReq；fallback -c copy 失败时 re-encode AAC
- main.rs 删除 asr_url/asr_token 字段（now sidecar concern）
- k8s manifest: ASR_URL/ASR_TOKEN 从主容器移到 feishu sidecar env

apps/notes/feishu/server.py

@@ -1,17 +1,19 @@
-"""notes feishu sidecar：HTTP 包一层 markdown-to-feishu。
-
-POST /convert  {md_path, title?, existing_doc_id?}
-  → 跑 markdown-to-feishu，parse 最后那段 JSON，返回 {doc_id, url}
+"""notes 多用途 sidecar：
+  POST /transcribe — 用 ffmpeg 切片 + 串行调外部 ASR，绕过单请求大小限制
+  POST /convert    — markdown-to-feishu，把会议纪要 push 飞书 docx
 """
 
 import json
 import logging
 import os
-import re
+import shutil
 import subprocess
+import tempfile
+import uuid
 from pathlib import Path
 from typing import Optional
 
+import requests
 from fastapi import FastAPI, HTTPException
 from pydantic import BaseModel
 
@@ -27,6 +29,81 @@ def healthz():
     return {'ok': True}
 
 
+class TranscribeReq(BaseModel):
+    audio_path: str
+    chunk_seconds: int = 60  # 60s ≈ 1-1.5 MB m4a，远低于 ASR 限制
+
+
+@app.post('/transcribe')
+def transcribe(req: TranscribeReq):
+    """ffmpeg 切片 → 串行喂外部 ASR → 拼接 transcript。"""
+    src = Path(req.audio_path)
+    if not src.exists():
+        raise HTTPException(400, f'audio not found: {src}')
+    asr_url = os.environ.get('ASR_URL', '')
+    asr_token = os.environ.get('ASR_TOKEN', '')
+    if not asr_url or not asr_token:
+        raise HTTPException(500, 'ASR_URL/ASR_TOKEN not configured in sidecar')
+
+    tmp = Path(tempfile.gettempdir()) / f'transcribe-{uuid.uuid4().hex}'
+    tmp.mkdir(parents=True)
+    try:
+        # 用 ffmpeg segment：直接 copy stream（fast & 不损失质量）
+        # 个别情况下 -c copy 在某些容器格式下切不精准，回退 re-encode 到 aac
+        ext = src.suffix.lstrip('.') or 'm4a'
+        chunk_pattern = f'chunk_%03d.{ext}'
+        try:
+            subprocess.run(
+                ['ffmpeg', '-y', '-i', str(src),
+                 '-f', 'segment', '-segment_time', str(req.chunk_seconds),
+                 '-c', 'copy', '-reset_timestamps', '1',
+                 str(tmp / chunk_pattern)],
+                check=True, capture_output=True, timeout=180,
+            )
+        except subprocess.CalledProcessError:
+            # fallback: re-encode AAC，慢但稳
+            log.warning("ffmpeg -c copy 失败，回退 re-encode")
+            for p in tmp.glob(f'chunk_*.{ext}'):
+                p.unlink(missing_ok=True)
+            subprocess.run(
+                ['ffmpeg', '-y', '-i', str(src),
+                 '-f', 'segment', '-segment_time', str(req.chunk_seconds),
+                 '-c:a', 'aac', '-b:a', '64k', '-ac', '1', '-ar', '16000',
+                 '-reset_timestamps', '1',
+                 str(tmp / 'chunk_%03d.m4a')],
+                check=True, capture_output=True, timeout=600,
+            )
+            ext = 'm4a'
+
+        chunks = sorted(tmp.glob(f'chunk_*.{ext}'))
+        if not chunks:
+            raise HTTPException(500, 'ffmpeg produced 0 chunks')
+        log.info("split %s → %d chunks", src.name, len(chunks))
+
+        all_text = []
+        for i, c in enumerate(chunks, 1):
+            log.info("ASR chunk %d/%d (%s, %d KB)", i, len(chunks), c.name, c.stat().st_size // 1024)
+            with open(c, 'rb') as f:
+                r = requests.post(
+                    asr_url,
+                    headers={'Authorization': f'Bearer {asr_token}'},
+                    files={'file': (c.name, f, 'audio/mp4')},
+                    data={'model': 'qwen3-asr', 'response_format': 'json'},
+                    timeout=300,
+                )
+            if not r.ok:
+                raise HTTPException(502, f'ASR chunk {i} {r.status_code}: {r.text[:300]}')
+            try:
+                text = r.json().get('text', '').strip()
+            except Exception:
+                raise HTTPException(502, f'ASR chunk {i} bad json: {r.text[:200]}')
+            all_text.append(text)
+        full = '\n'.join(t for t in all_text if t)
+        return {'text': full, 'chunks': len(chunks)}
+    finally:
+        shutil.rmtree(tmp, ignore_errors=True)
+
+
 class ConvertReq(BaseModel):
     md_path: str
     title: Optional[str] = None