fix(honcho): fully shut down async writer + close HTTP client on exit

[btcjon]

Problem

HonchoMemoryProvider.shutdown() only called flush_all(), which leaves two things alive at interpreter teardown:

1. The honcho-async-writer background thread (owned by HonchoSessionManager) — flush_all() drains the queue but never sends _ASYNC_SHUTDOWN and never joins the thread.
2. The underlying httpx.Client inside the Honcho SDK (honcho._http._client) — its connection pool is still open when Python finalizes.

This race caused intermittent crashes during clean exit.

Fix

HonchoMemoryProvider.shutdown() now:

1. Calls HonchoSessionManager.shutdown() (which already does flush → _ASYNC_SHUTDOWN → join).
2. Explicitly closes honcho._http so the httpx pool is torn down before finalization.

Both wrapped in try/except with debug logging — shutdown must never raise.

Verification

Offline smoke test with a mocked Honcho client:

async thread: <Thread(honcho-async-writer, started daemon ...)>
alive before: True
alive after : False
http.close called: True

Full honcho test suite: 141 passed, 3 skipped.

Sibling memory plugins (byterover, hindsight, holographic) were audited for the same pattern — none affected. supermemory is not installed in this environment.

[teknium1]

Thanks for addressing deterministic Honcho teardown. The current provider still calls only flush_all() at plugins/memory/honcho/__init__.py:1421, while HonchoSessionManager.shutdown() owns the async-writer sentinel and join at plugins/memory/honcho/session.py:548-553, so the underlying target is real.

Problems

• The replacement at this PR's right-side line 707 regresses writeFrequency: "session": HonchoSessionManager.shutdown() only calls flush_all() when an async queue/thread exists (plugins/memory/honcho/session.py:550-553), but session mode has no queue and intentionally defers writes to flush_all() (plugins/memory/honcho/session.py:517-519). Preserve the provider's final flush for every mode.
• This one-file PR adds no regression tests. Existing tests cover manager async shutdown (tests/honcho_plugin/test_async_memory.py:291-336), not provider shutdown or the private client._http.close() path.
• The unchanged 5-second provider prefetch join (plugins/memory/honcho/__init__.py:1415-1417) remains relevant to the timeout mismatch documented by cross-referenced issue #33485.

Suggested changes

• Move the universal flush into HonchoSessionManager.shutdown() before its async-worker conditional, then test session-mode flush, writer termination, and idempotent HTTP close.
• Reconcile the prefetch-thread join with #33485 before closing the shared client.

Automated hermes-sweeper review.

[teknium1]

HonchoSessionManager.shutdown() currently invokes flush_all() only when an async queue/thread exists (plugins/memory/honcho/session.py:550-553). For supported writeFrequency: "session", no queue exists and writes defer until flush_all() (session.py:517-519), so this replacement removes the provider's final flush. Make manager shutdown flush all modes unconditionally, or retain that flush here before invoking the worker teardown.