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:

backend/services/generation.py

@@ -17,6 +17,7 @@ Mode differences:
 from __future__ import annotations
 
 import asyncio
+import logging
 import traceback
 from typing import Literal, Optional
 
@@ -26,6 +27,22 @@ from ..database import get_db
 from ..utils.tasks import get_task_manager
 
 
+def release_generation_memory(tts_model) -> None:
+    """Best-effort post-generation memory cleanup.
+
+    Collects garbage and flushes the device allocator cache so the process heap
+    does not grow across consecutive generations (#923). Never raises: a
+    cleanup failure (e.g. a poisoned CUDA context) must not replace the
+    generation's own result or error.
+    """
+    from ..backends.base import empty_device_cache
+
+    try:
+        empty_device_cache(getattr(tts_model, "device", "cpu"))
+    except Exception as e:
+        logging.getLogger(__name__).debug("post-generation cache cleanup failed: %s", e)
+
+
 async def run_generation(
     *,
     generation_id: str,
@@ -54,7 +71,6 @@ async def run_generation(
         get_tts_backend_for_engine,
         load_engine_model,
     )
-    from ..backends.base import empty_device_cache
     from ..utils.chunked_tts import generate_chunked
     from ..utils.audio import has_tts_runaway, normalize_audio, save_audio, trim_tts_output
 
@@ -158,7 +174,7 @@ async def run_generation(
     finally:
         task_manager.complete_generation(generation_id)
         bg_db.close()
-        empty_device_cache(getattr(tts_model, "device", "cpu"))
+        release_generation_memory(tts_model)
 
 
 def _notify_speak_end(generation_id: str, *, status: str) -> None:
@@ -283,7 +299,6 @@ async def generate_audio_sync(
         get_tts_backend_for_engine,
         load_engine_model,
     )
-    from ..backends.base import empty_device_cache
     from ..utils.chunked_tts import generate_chunked
     from ..utils.audio import has_tts_runaway, normalize_audio, trim_tts_output
     from . import tts
@@ -328,7 +343,7 @@ async def generate_audio_sync(
 
         return tts.audio_to_wav_bytes(audio, sample_rate)
     finally:
-        empty_device_cache(getattr(tts_model, "device", "cpu"))
+        release_generation_memory(tts_model)
 
 
 def _save_regenerate(