fix(generation): make post-generation cleanup best-effort and cover /generate/stream

Wrap empty_device_cache in release_generation_memory(), which logs and
swallows cleanup failures so a poisoned CUDA context cannot replace a
finished generation's result, and call it from the streaming endpoint
too, which drives generate_chunked directly.

backend/routes/generations.py

@@ -12,7 +12,7 @@ from sqlalchemy.orm import Session
 from .. import config, models
 from ..services import history, personality, profiles, tts
 from ..database import Generation as DBGeneration, VoiceProfile as DBVoiceProfile, get_db
-from ..services.generation import run_generation
+from ..services.generation import release_generation_memory, run_generation
 from ..services.task_queue import cancel_generation as cancel_generation_job, enqueue_generation
 from ..utils.audio import load_audio
 from ..utils.tasks import get_task_manager
@@ -363,18 +363,21 @@ async def stream_speech(
 
         runaway_detector = has_tts_runaway
 
-    audio, sample_rate = await generate_chunked(
-        tts_model,
-        data.text,
-        voice_prompt,
-        language=data.language,
-        seed=data.seed,
-        instruct=data.instruct,
-        max_chunk_chars=data.max_chunk_chars,
-        crossfade_ms=data.crossfade_ms,
-        trim_fn=trim_fn,
-        runaway_detector=runaway_detector,
-    )
+    try:
+        audio, sample_rate = await generate_chunked(
+            tts_model,
+            data.text,
+            voice_prompt,
+            language=data.language,
+            seed=data.seed,
+            instruct=data.instruct,
+            max_chunk_chars=data.max_chunk_chars,
+            crossfade_ms=data.crossfade_ms,
+            trim_fn=trim_fn,
+            runaway_detector=runaway_detector,
+        )
+    finally:
+        release_generation_memory(tts_model)
 
     effects_chain_config = None
     if data.effects_chain is not None: