refactor: centralize cleanup in app layer

- New: application/cleanup.py — flush_cuda(), dispose_engines(), cleanup(), register_ui_cleanup()
- conversion_service.py finally: pool.dispose_all() + voice_cache.clear() + flush_cuda()
- webui/conversion_runner.py: removed gc/cuda finally block (cleanup in run_conversion)
- shutdown.py: 160→120 lines, 5 inline cleanups → 4 process-level + app_cleanup() delegation
- Fixed bugs: _PIPELINES (didn't exist), PluginManager.dispose_all() (never called),
  VoiceCache.clear() (never called in finally), duplicate cleanup removed

abogen/application/cleanup.py

@@ -0,0 +1,73 @@
+"""Application-layer cleanup — global resource disposal.
+
+Handles:
+- GPU/CUDA memory flush
+- TTS engine disposal (PluginManager)
+- UI-specific cleanup callbacks (registered by entry points)
+
+Called by shutdown.py at process exit and by run_conversion() per-conversion.
+"""
+
+from __future__ import annotations
+
+import gc
+from typing import Callable
+
+_UI_CLEANUPS: list[Callable[[], None]] = []
+
+
+def flush_cuda() -> None:
+    """Run GC and release CUDA cache. Safe to call multiple times."""
+    gc.collect()
+    try:
+        import torch
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+            torch.cuda.ipc_collect()
+    except Exception:
+        pass
+
+
+def dispose_engines() -> None:
+    """Dispose all cached TTS engines via PluginManager."""
+    try:
+        from abogen.tts_plugin.plugin_manager import get_plugin_manager
+        get_plugin_manager().dispose_all()
+    except Exception:
+        pass
+
+
+def _clear_global_voice_cache() -> None:
+    """Reset the global voice download cache state."""
+    try:
+        from abogen.voice_cache import clear_voice_cache
+        clear_voice_cache()
+    except Exception:
+        pass
+
+
+def register_ui_cleanup(fn: Callable[[], None]) -> None:
+    """Register a UI-specific cleanup callback (e.g. preview threads, temp files)."""
+    _UI_CLEANUPS.append(fn)
+
+
+def cleanup() -> None:
+    """Run all application-level cleanups. Idempotent."""
+    dispose_engines()
+    flush_cuda()
+    _clear_global_voice_cache()
+
+    for fn in _UI_CLEANUPS:
+        try:
+            fn()
+        except Exception:
+            pass
+    _UI_CLEANUPS.clear()
+
+
+__all__ = [
+    "flush_cuda",
+    "dispose_engines",
+    "register_ui_cleanup",
+    "cleanup",
+]

abogen/application/conversion_service.py

@@ -104,6 +104,9 @@ def run_conversion(
         raise
     finally:
         pool.dispose_all()
+        voice_cache.clear()
+        from abogen.application.cleanup import flush_cuda
+        flush_cuda()
 
 
 def _create_voice_resolver(

abogen/shutdown.py

@@ -1,8 +1,19 @@
-"""Graceful shutdown - single module, no over-engineering."""
+"""Graceful shutdown — process-level hooks and orchestration.
+
+Responsibilities:
+- Install atexit/signal/Qt hooks
+- Stop WebUI ConversionService (worker thread)
+- Restore sleep prevention
+- Terminate child processes (ffmpeg, etc.)
+- Delegate GPU/engine/UI cleanup to application.cleanup
+
+App-layer cleanup (GPU, engines, UI callbacks) lives in application/cleanup.py.
+Per-conversion cleanup lives in run_conversion() finally block.
+"""
+
 from __future__ import annotations
 
 import atexit
-import gc
 import signal
 import sys
 from typing import Callable
@@ -28,20 +39,11 @@ def _run_cleanups() -> None:
             pass
 
 
-# ---- Register built-in cleanup functions ----
+# ---- Process-level cleanup functions ----
 
-# 1. Restore sleep prevention
-def _restore_sleep() -> None:
-    try:
-        from abogen.utils import prevent_sleep_end
-        prevent_sleep_end()
-    except Exception:
-        pass
 
-register_cleanup(_restore_sleep)
-
-# 2. Shutdown web UI ConversionService
-def _shutdown_conversion_service() -> None:
+def _stop_conversion_service() -> None:
+    """Stop WebUI ConversionService worker thread."""
     try:
         from abogen.webui.service import get_service
         svc = get_service()
@@ -50,50 +52,18 @@ def _shutdown_conversion_service() -> None:
     except Exception:
         pass
 
-register_cleanup(_shutdown_conversion_service)
 
-# 3. Clear TTS pipelines and GPU memory
-def _cleanup_tts_pipelines() -> None:
-    # Clear web UI pipeline cache
+def _restore_sleep() -> None:
+    """Restore system sleep prevention (caffeinate/systemd-inhibit/Windows)."""
     try:
-        from abogen.webui.conversion_runner import _PIPELINES
-        _PIPELINES.clear()
+        from abogen.utils import prevent_sleep_end
+        prevent_sleep_end()
     except Exception:
         pass
 
-    # Clear PyQt conversion thread voice cache
-    try:
-        from abogen.pyqt.conversion import ConversionThread
-        if hasattr(ConversionThread, "voice_cache"):
-            ConversionThread.voice_cache.clear()
-    except Exception:
-        pass
 
-    gc.collect()
-
-    # Release CUDA cache
-    try:
-        import torch
-        if torch.cuda.is_available():
-            torch.cuda.empty_cache()
-            torch.cuda.ipc_collect()
-    except Exception:
-        pass
-
-register_cleanup(_cleanup_tts_pipelines)
-
-# 4. Clear global voice cache
-def _clear_voice_cache() -> None:
-    try:
-        from abogen.voice_cache import clear_voice_cache
-        clear_voice_cache()
-    except Exception:
-        pass
-
-register_cleanup(_clear_voice_cache)
-
-# 5. Terminate child processes (ffmpeg, etc.)
 def _terminate_subprocesses() -> None:
+    """Terminate all child processes (ffmpeg, etc.)."""
     try:
         import psutil
     except Exception:
@@ -115,6 +85,20 @@ def _terminate_subprocesses() -> None:
     except Exception:
         pass
 
+
+def _app_cleanup() -> None:
+    """Delegate to application-layer cleanup (engines, GPU, UI callbacks)."""
+    try:
+        from abogen.application.cleanup import cleanup
+        cleanup()
+    except Exception:
+        pass
+
+
+# Register in execution order
+register_cleanup(_stop_conversion_service)
+register_cleanup(_app_cleanup)
+register_cleanup(_restore_sleep)
 register_cleanup(_terminate_subprocesses)
 
 
@@ -133,7 +117,7 @@ def register_shutdown() -> None:
         except Exception:
             pass
 
-    # Qt hook
+    # Qt hook — connect AFTER QApplication is created
     try:
         from PyQt6.QtWidgets import QApplication
 

abogen/webui/conversion_runner.py

@@ -15,7 +15,6 @@ Engine converts Language → its own format internally.
 
 from __future__ import annotations
 
-import gc
 from pathlib import Path
 from typing import Any
 
@@ -191,11 +190,3 @@ def run_conversion_job(job: Job) -> None:
         job.error = str(exc)
         job.status = JobStatus.FAILED
         job.add_log(f"Job failed: {exc}", level="error")
-    finally:
-        gc.collect()
-        try:
-            import torch
-            if torch.cuda.is_available():
-                torch.cuda.empty_cache()
-        except ImportError:
-            pass