Audit findings: potential race conditions and GPU sync gaps

[PSiKoTiC]

I missed an inconsistency in GPU sync ordering in PR #869 — I fixed _clear_accelerator_cache() but didn't scan the file for the same pattern in unload(). While investigating why I missed it and checking for similar gaps, I ran a broader audit and found a few additional areas that may warrant attention. Flagging them here for visibility — happy to submit PRs for any of these if helpful.

GPU synchronization gaps

Location	Finding
init_service_memory_basic.py:38-42	MPS path calls empty_cache() without preceding synchronize()
init_service_loader.py:127-140	After del self.model, synchronize() is in a separate try-except after empty_cache() — if it fails, cleanup is incomplete
init_service_orchestrator.py:191-200	On init failure, partial models are set to None without calling unload() — GPU memory from partially loaded models may not be freed

Thread safety

Location	Finding
i18n/i18n.py:138-155	get_i18n() checks _i18n_instance without a lock before initialization — addressed in #878
api/api_routes.py:22-86	_api_key global modified by set_api_key() without synchronization while async handlers read it — addressed in #878
api/api_routes.py:94, 153-164	diskcache.Cache accessed from concurrent async handlers without locking — addressed in #878
local_cache.py:28-40	Singleton __new__ has a double-checked lock pattern where _initialized can be stale — addressed in #878

Minor / low priority

Location	Finding
release_task_audio_paths.py:64-84	if os.close(fd) fails on temp file, file descriptor leaks — addressed in #879
api_routes.py:90-98	diskcache.Cache created at module load but never closed on shutdown — addressed in #879
runtime_helpers.py:62-73	TensorBoard subprocess may orphan on Windows if CTRL+C doesn't propagate

Tested on Windows 11, Python 3.11, RTX 5090. Findings are from code review, not observed failures — flagging proactively.

[PSiKoTiC]

During the ContextVar follow-up work for #880 (building on #878's locking), the defensive coding audit flagged a few pre-existing style items in i18n/i18n.py that were out of scope for the PR but worth noting for future cleanup:

Location	Finding
i18n/i18n.py:54-56	_load_all_translations uses print() for error logging instead of loguru.logger
i18n/i18n.py:72	set_language uses print() for warnings instead of loguru.logger
i18n/i18n.py:55	Bare except Exception as e — could be narrowed to (json.JSONDecodeError, OSError)
i18n/i18n.py:9, 32, 109, 146, 150	Uses Dict/Optional from typing instead of modern dict/`str

None of these are functional bugs — the module works correctly. Just bringing them to the same standard as the rest of the gradio UI layer. Happy to submit a cleanup PR if there's interest, or these can be picked up whenever someone's already touching the file.