Merge branch 'fix/review-2138' into fix/community-integration

# Conflicts:
#	docs/dubbing/translation-engines.md

backend/api/routers/dub_core.py

@@ -1918,7 +1918,7 @@ async def dub_transcribe_stream(
                     "heuristic",
                 )
 
-        fut_diar = loop.run_in_executor(_gpu_pool, _diarize)
+        fut_diar = loop.run_in_executor(_gpu_pool, lambda: _asr_work.run(_diarize))
         async for _ping in _ping_while(fut_diar):
             yield _ping
         final_segs, diar_warning, labels_source = fut_diar.result()

backend/api/routers/dub_export.py

@@ -291,7 +291,10 @@ async def stream_task(task_id: str, after_seq: int = 0):
         finally:
             await task_manager.remove_listener(task_id, q)
 
-    return StreamingResponse(_reader(), media_type="text/event-stream")
+    return StreamingResponse(
+        _reader(), media_type="text/event-stream",
+        headers={"Cache-Control": "no-cache, no-transform", "X-Accel-Buffering": "no"},
+    )
 
 
 @router.get("/jobs")

docs/dubbing/translation-engines.md

@@ -263,3 +263,5 @@ Mixed or malformed SRT files can make a bare number indistinguishable from spoke
 If the Argos native runtime cannot load, both desktop and browser clients show localized recovery guidance: reinstall the backend or select NLLB. The API returns the stable `argos_runtime_unavailable` error code without exposing native library paths.
 
 WebVTT import separates metadata blocks from cue identifiers using the [WebVTT block-parsing rules](https://www.w3.org/TR/webvtt1/#file-parsing): a timing line immediately after an identifier makes a cue, even when that identifier is NOTE, STYLE, or REGION. Later timing examples inside metadata are ignored, and empty cues never borrow the next cue’s identifier as dialogue.
+
+Dubbing transcription emits keepalives during quiet diarization, reference-refinement, and cleanup steps. Disconnecting stops queued model work; native calls already running retain their model until they finish, then cleanup restores TTS. Task streams also request that proxies disable buffering so keepalives reach the client promptly.

tests/test_dub_transcribe.py

@@ -1024,3 +1024,69 @@ def test_stream_unload_is_single_shot_during_disconnect():
         normal.result(timeout=5)
         disconnected.result(timeout=5)
     assert calls == ["unload"]
+
+
+def test_disconnect_during_diarization_waits_before_unload_and_restore(tmp_path, monkeypatch):
+    import asyncio
+    import threading
+    from concurrent.futures import ThreadPoolExecutor
+    from api.routers import dub_core as dc
+    from services import asr_backend
+
+    started, release, cleanup_started, restored = [threading.Event() for _ in range(4)]
+    events, guarded = [], []
+    original_cleanup = dc._ASRWorkLifetime.cleanup
+    def cleanup(lifetime, fn):
+        guarded.append(lifetime._lock.locked())
+        cleanup_started.set()
+        return original_cleanup(lifetime, fn)
+    monkeypatch.setattr(dc._ASRWorkLifetime, "cleanup", cleanup)
+    def diarize(**kwargs):
+        started.set()
+        assert release.wait(5)
+        events.append("diarization finished")
+        return None, None
+    class Backend:
+        id = "fake"
+        def ensure_loaded(self): pass
+        def transcribe(self, *a, **kw):
+            return {"chunks": [{"text": "hi", "timestamp": (0., .5)}], "segments": [], "language": "en"}
+        def unload(self): events.append("unloaded")
+    async def model():
+        result = MagicMock()
+        result._asr_pipe = MagicMock()
+        return result
+    def restore():
+        events.append("restored")
+        restored.set()
+    monkeypatch.setattr(dc, "get_model", model)
+    monkeypatch.setattr(asr_backend, "get_active_asr_backend", lambda *a, **kw: Backend())
+    monkeypatch.setattr(dc, "get_diarization_pipeline", diarize)
+    monkeypatch.setattr(dc, "offload_tts_for_asr", lambda: None)
+    monkeypatch.setattr(dc, "restore_tts_after_asr", restore)
+    monkeypatch.setattr(dc, "POST_ASR_PING_S", .01)
+    audio = tmp_path / "diar.wav"
+    _make_wav(audio, seconds=1.)
+    job_id = "diar_disconnect"
+    dc._dub_jobs[job_id] = {"audio_path": str(audio), "vocals_path": None, "scene_cuts": []}
+    async def scenario():
+        response = await dc.dub_transcribe_stream(job_id)
+        try:
+            async for _ in response.body_iterator:
+                if started.is_set():
+                    break
+            await response.body_iterator.aclose()
+            assert await asyncio.to_thread(cleanup_started.wait, 5)
+            assert guarded == [True]
+            assert not restored.is_set()
+        finally:
+            release.set()
+            await response.body_iterator.aclose()
+            assert await asyncio.to_thread(restored.wait, 5)
+    try:
+        with ThreadPoolExecutor(max_workers=2) as pool:
+            monkeypatch.setattr(dc, "_gpu_pool", pool)
+            asyncio.run(asyncio.wait_for(scenario(), timeout=10))
+    finally:
+        dc._dub_jobs.pop(job_id, None)
+    assert events == ["diarization finished", "unloaded", "restored"]

tests/test_tasks_stream_keepalive.py

@@ -28,6 +28,8 @@ def test_quiet_task_stream_emits_keepalive_comments(monkeypatch):
             "error": None, "cancelled": False,
         }
         resp = await de.stream_task(task_id)
+        assert resp.headers["cache-control"] == "no-cache, no-transform"
+        assert resp.headers["x-accel-buffering"] == "no"
         frames = []
 
         async def _read():