mouth: sweep orphaned espeak temp dirs at startup

[BigpapaWarren]

Sweep orphaned espeak temp dirs at startup

The problem

phonemizer copies the espeak shared library into a fresh mkdtemp() for every
backend it builds — espeak-ng keeps its state in globals, and dlopen won't load
the same file twice. Kokoro builds several backends, so one warm() leaves
several of these directories behind.

Cleanup is registered two different ways, and only one of them works reliably.
From phonemizer/backend/espeak/api.py:

# But... weakref implementation does not work on windows so we register
# the cleanup with atexit. This means that, on Windows, all the
# temporary directories created by EspeakAPI instances will remain on
# disk until the Python process exit.
if sys.platform == 'win32':
    atexit.register(self._delete_win32)
else:
    weakref.finalize(self, self._delete, self._library, self._tempdir)

atexit does not run when a process is killed rather than exited. Anything
that launches backtalk from a wrapper and stops it by terminating the process — a
launcher script, a service wrapper, a supervisor, a stop button that calls
taskkill — leaks every directory it ever created.

This is not hypothetical: 60 orphaned directories had accumulated on the
machine where I found it. The count grows by one per backend per start and never
falls. Each holds a ~410 KB copy of espeak-ng.dll.

Why not fix it in phonemizer

Two reasons:

1. It's arguably not phonemizer's bug — its comment is accurate about what it
   can guarantee, and atexit genuinely cannot cover SIGKILL-equivalent
   termination.
2. Even as a local workaround it doesn't survive: any install step that runs
   uv sync or pip install -U on launch overwrites site-packages.

Sweeping at backtalk's own startup bounds the total at one run's worth,
regardless of how the previous run ended.

The change

One function, called from warm() immediately before Kokoro creates this run's
directories. 78 lines added, nothing modified or removed.

Why it's safe to point at the shared temp directory

Two independent guarantees, neither of them a heuristic:

1. Shape check — only a directory whose sole entry is an espeak shared
   library is removed. phonemizer's scratch dirs look exactly like that, and
   nothing else in a temp directory does.
2. The OS enforces the rest — Windows refuses to delete a loaded DLL, so a
   concurrently running instance's directory fails the rmtree and is skipped.
   The code doesn't have to detect concurrent instances; it cannot win that race
   because the filesystem won't let it.

shutil.rmtree failures are caught and ignored: a directory still in use is left
for the next start, which is the correct outcome.

Testing

Run against a live instance with four in-use directories present:

synthetic orphans removed : 3/3     <- want 3
decoy (multi-file) kept   : True    <- want True
decoy (non-espeak) kept   : True    <- want True
decoy (empty dir) kept    : True    <- want True
LIVE voice-line dirs kept : 4/4     <- want all (DLL is loaded)

The last line is the one that matters: the sweep ran with four in-use
directories sitting in the same temp folder and left every one alone.

Confirmed in production across three restarts — the log line appears, the count
returns to one run's worth each time instead of climbing:

11:45:11 [mouth] swept 4 orphaned espeak temp dir(s)
12:00:20 [mouth] swept 4 orphaned espeak temp dir(s)

Notes

• Windows is where the leak is unbounded, but the sweep is platform-neutral and
  harmless on POSIX, where weakref.finalize already keeps the count near zero.
• It logs only when it actually removes something, so a clean system stays quiet.
• No new dependencies; shutil and tempfile are stdlib.