fix: normalize faster-whisper segments to dicts + in-proc auto-worker

Colab run 3 (A100) surfaced three GPU/API-path bugs mocks couldn't catch:

1. faster-whisper yields namedtuple Segments, but batch/bench consume
   them as dicts (.get) -> AttributeError 'Segment' has no attribute
   'get' on every real transcription. Engine now normalizes segments
   to dicts (_to_dict_segments) at the boundary.

2. API TranscribeOptions carries engine-irrelevant keys (formats,
   timestamps, glossary_id, post_correction, diarize); worker's
   TranscriptionOptions(**job.options) crashed with TypeError. Worker
   now filters job.options to TranscriptionOptions.__slots__.

3. in-proc server never consumed its own queue (jobs stayed queued
   forever). Added opt-in Settings.auto_worker (default off): lifespan
   starts a daemon Worker thread for inproc backend, stopped on
   shutdown. Notebook enables it via LUKESCRIBE_AUTO_WORKER=true so the
   API upload -> completed flow works end to end.

Notebook: bench manifest now uses clips schema (audio_path/duration_sec/
entities); cell 22 reads error_message/error_code; upload poll window
raised to 4min (first-run model download).

+ 5 tests (namedtuple/dict segments, API-style options, auto_worker
on/off); 136 tests pass, ruff clean.

src/luke_scribe/api/app.py

@@ -9,6 +9,7 @@ Lifespan (plan §3.10a/§3.5):
 from __future__ import annotations
 
 import logging
+import threading
 from contextlib import asynccontextmanager
 
 from fastapi import FastAPI, Request
@@ -63,6 +64,23 @@ def create_app(settings: Settings | None = None) -> FastAPI:
 
         app.state.session_guard = _SessionGuard(settings)
 
+        # in-proc 백엔드 + auto_worker: 같은 프로세스 워커 스레드가 큐를 소비
+        # (dev/Colab에서 별도 워커 프로세스 없이 업로드 → 완료 흐름이 가능하게 함)
+        inproc_worker = None
+        if settings.queue_backend == "inproc" and settings.auto_worker:
+            from ..jobqueue.worker import Worker
+
+            inproc_worker = Worker(
+                settings=settings,
+                broker=broker,
+                store=store,
+                owner=owner,
+                worker_id="api-inproc",
+            )
+            threading.Thread(target=inproc_worker.run_forever, daemon=True).start()
+            logger.info("in-proc auto-worker 스레드 시작 (queue_backend=inproc, auto_worker=true)")
+        app.state.inproc_worker = inproc_worker
+
         # 모델 프로비저닝 (선택) — 설정된 경우에만
         model_cache = settings.model_cache_dir
         if model_cache:
@@ -99,6 +117,10 @@ def create_app(settings: Settings | None = None) -> FastAPI:
                 tunnel.stop()
             except Exception:
                 pass
+        if inproc_worker is not None:
+            # stop()은 다음 루프 반복에서 반영 — 진행 중 job은 끝까지 완료 후
+            # 스레드가 종료된다 (daemon + 프로세스 종료 흐름에서는 안전)
+            inproc_worker.stop()
         owner.unload_all()
 
     app = FastAPI(

src/luke_scribe/config.py

@@ -58,6 +58,9 @@ class Settings(BaseSettings):
     job_timeout_hours: float = 4.0
     job_timeout_margin_rtf: float = 2.0  # duration × RTF 추정 시 마진
 
+    # ── 워커 ──
+    auto_worker: bool = False  # in-proc 백엔드: 서버 프로세스가 큐를 소비 (dev/Colab)
+
     # ── 입력 상한 ──
     max_duration_sec: int = 14400  # 4h
     max_upload_bytes: int = 2 * 1024 * 1024 * 1024  # 2GB

src/luke_scribe/engine/faster_whisper_engine.py

@@ -130,9 +130,45 @@ class FasterWhisperEngine(TranscriptionEngine):
         except Exception as exc:
             raise TranscriptionFailed(f"전사 중 오류: {exc}") from exc
 
-        wrapped = _CancellableSegmentIterator(segments_iter, should_cancel or (lambda: False))
+        wrapped = _CancellableSegmentIterator(
+            self._to_dict_segments(segments_iter), should_cancel or (lambda: False)
+        )
         return TranscriptionOutcome(wrapped, info=info)
 
+    @staticmethod
+    def _to_dict_segments(segments):
+        """faster-whisper Segment(namedtuple) → dict로 정규화.
+
+        다운스트림(배치 파이프라인/벤치/실시간)은 dict 계약을 쓴다 (mock과 동일).
+        GPU 실전에서만 재현되는 버그: 'Segment' object has no attribute 'get'.
+        """
+        for seg in segments:
+            if isinstance(seg, dict):
+                yield seg
+                continue
+            asdict = getattr(seg, "_asdict", None)
+            if asdict is not None:
+                yield dict(asdict())
+                continue
+            # 최후 수단: 알려진 필드만, None 값은 제외 (다운스트림 .get(key, default)가
+            # key 존재+None으로 default를 무시하지 않도록)
+            yield {
+                k: v
+                for k in (
+                    "id",
+                    "seek",
+                    "start",
+                    "end",
+                    "text",
+                    "tokens",
+                    "temperature",
+                    "avg_logprob",
+                    "compression_ratio",
+                    "no_speech_prob",
+                )
+                if (v := getattr(seg, k, None)) is not None
+            }
+
     @staticmethod
     def _split_device(device: str) -> tuple[str, int]:
         """'cuda:N' → ('cuda', N). CTranslate2는 device='cuda'만 허용하므로 분리한다.

src/luke_scribe/jobqueue/worker.py

@@ -100,7 +100,16 @@ class Worker:
             self._emit_progress(job, processed_sec, total_sec)
 
         try:
-            options = TranscriptionOptions(**job.options)
+            # API의 TranscribeOptions에는 엔진과 무관한 키가 포함될 수 있다
+            # (formats/timestamps/glossary_id/post_correction/diarize 등) —
+            # 엔진 계약 필드만 골라 전달한다.
+            options = TranscriptionOptions(
+                **{
+                    k: v
+                    for k, v in (job.options or {}).items()
+                    if k in TranscriptionOptions.__slots__
+                }
+            )
             result = pipeline.run(job, options, progress_cb=progress_cb)
             # 결과를 먼저 영속화한 뒤 상태 전이 (실패 시 FAILED로 전이 가능하게)
             self.store.write_result(job.id, result)