feat: implement full-platform STT API (v2.3 consensus plan)

Batch+realtime transcription API: faster-whisper engine w/ EngineOwner
(single GPU owner, OOM downgrade chain, persisted attempted_profiles),
hardware-adaptive device manager (T0-T3 VRAM tiers), Redis/in-proc job
queue w/ leases + crash recovery, postprocess (glossary/rules/LLM w/
egress guard), privacy-first result store (UUID keys, source deleted
after transcribe), retention sweeper, API-key auth (HMAC digests,
scopes, job ownership), WebSocket realtime lane (LocalAgreement),
CLI (detect/transcribe/bench/serve/key), Docker, benchmark runner.

127 mock-based tests pass; ruff clean. Includes verification checklist
and autoplan review notes.

tests/unit/test_engine_owner.py

@@ -0,0 +1,95 @@
+"""EngineOwner 단위 테스트 — OOM 강등 체인, attempted_profiles 영속화, 우선순위."""
+
+from __future__ import annotations
+
+import pytest
+
+from luke_scribe.engine.base import TranscriptionOptions
+from luke_scribe.engine.owner import MAX_DOWNGRADES, EngineOwner, InferenceRequest
+from luke_scribe.errors import OutOfMemory
+
+
+class OOMFakeEngine:
+    """첫 N번 호출에서 OOM, 이후 성공하는 mock."""
+
+    def __init__(self, oom_calls: int = 0) -> None:
+        self.oom_calls = oom_calls
+        self.calls: list[dict] = []
+        self.closed = False
+
+    def transcribe(self, audio_path, options, should_cancel=None, download_progress=None):
+        self.calls.append({"device": options.device, "compute_type": options.compute_type})
+        if len(self.calls) <= self.oom_calls:
+            raise OutOfMemory("CUDA OOM (mock)")
+        return type("O", (), {"segments": iter([]), "info": {}})()
+
+    def unload_all(self):
+        self.closed = True
+
+
+def _opts(**kw) -> TranscriptionOptions:
+    kw.setdefault("model", "large-v3-turbo")
+    kw.setdefault("device", "cuda:0")
+    kw.setdefault("compute_type", "float16")
+    return TranscriptionOptions(**kw)
+
+
+def test_no_downgrade_on_success():
+    owner = EngineOwner.__new__(EngineOwner)
+    owner.settings = None
+    owner._engine = OOMFakeEngine(oom_calls=0)
+    owner._lock = __import__("threading").Lock()
+    owner._realtime_priority = __import__("threading").Lock()
+    owner._stats = {"jobs": 0, "downgrades": 0, "oom": 0}
+    req = InferenceRequest(audio_path="/tmp/x.wav", options=_opts())
+    out = owner.transcribe(req)
+    assert out["compute_type"] == "float16"
+    assert len(out["attempted_profiles"]) == 1
+
+
+def test_downgrade_chain_on_oom():
+    engine = OOMFakeEngine(oom_calls=1)  # 첫 프로파일에서 OOM
+    owner = EngineOwner.__new__(EngineOwner)
+    owner.settings = None
+    owner._engine = engine
+    owner._lock = __import__("threading").Lock()
+    owner._realtime_priority = __import__("threading").Lock()
+    owner._stats = {"jobs": 0, "downgrades": 0, "oom": 0}
+    req = InferenceRequest(audio_path="/tmp/x.wav", options=_opts())
+    out = owner.transcribe(req)
+    # float16 → int8_float16 (강등 1회)
+    assert out["compute_type"] == "int8_float16"
+    assert len(out["attempted_profiles"]) == 2
+
+
+def test_downgrade_capped_at_max():
+    engine = OOMFakeEngine(oom_calls=99)  # 항상 OOM
+    owner = EngineOwner.__new__(EngineOwner)
+    owner.settings = None
+    owner._engine = engine
+    owner._lock = __import__("threading").Lock()
+    owner._realtime_priority = __import__("threading").Lock()
+    owner._stats = {"jobs": 0, "downgrades": 0, "oom": 0}
+    req = InferenceRequest(audio_path="/tmp/x.wav", options=_opts())
+    with pytest.raises(OutOfMemory):
+        owner.transcribe(req)
+    # 강등 체인은 cpu까지 포함하지만 시도는 MAX_DOWNGRADES+1 이내
+    assert len(engine.calls) <= MAX_DOWNGRADES + 1
+
+
+def test_attempted_profiles_persisted_for_retry():
+    """재큐(retry) 시 attempted_profiles가 유지되어 총 시도가 캡된다."""
+    engine = OOMFakeEngine(oom_calls=0)
+    owner = EngineOwner.__new__(EngineOwner)
+    owner.settings = None
+    owner._engine = engine
+    owner._lock = __import__("threading").Lock()
+    owner._realtime_priority = __import__("threading").Lock()
+    owner._stats = {"jobs": 0, "downgrades": 0, "oom": 0}
+    # 이미 2회 시도한 이력이 Job에 영속화된 상황 → 재시도는 즉시 실패해야 함
+    req = InferenceRequest(
+        audio_path="/tmp/x.wav", options=_opts(), attempted_profiles=["float16", "int8_float16"]
+    )
+    with pytest.raises(OutOfMemory):
+        owner.transcribe(req)
+    assert len(engine.calls) == 0  # 새 시도 없음