Fix HNSW boot-integrity check: compare element count, not label_map size

[danielhertz1999-bit]

The bug

HippoDB._initialize_hnsw_index decides whether to rebuild the HNSW index by comparing the in-memory _label_map size against the SQLite row count:

active_label_count = len(self._label_map)
if active_label_count != sqlite_count:
    ... rebuild ...

But _label_map is repopulated from SQLite (_repopulate_label_map_from_sqlite()) before this check runs, so len(self._label_map) already equals sqlite_count — the condition is effectively always false and the rebuild never fires.

When the on-disk HNSW file is stale relative to SQLite (e.g. an interrupted write), the daemon boots with the HNSW index holding fewer elements than the label map / SQLite believe exist. The integrity check that's supposed to catch and repair this silently passes. A subsequent knn_query(k=2) against a 1-element index then raises RuntimeError on every insert.

Impact (observed on Windows)

This blocks the deferred-capture drain completely: with HNSW count=1 but _label_map/SQLite count=2, every capture insert crashes, so a backlog of session files never drains. After the fix, the same store drains cleanly (a 66-file / 322-turn backlog went from N=2 to draining normally).

The fix

Compare the actual HNSW element count against SQLite, so a real divergence triggers the rebuild:

hnsw_count = self._hnsw.get_current_count()
if hnsw_count != sqlite_count:
    ... rebuild ...

4 lines, no behavioral change when the index is already consistent. Surfaced while validating the new Windows support (v1.2.0) end-to-end.

🤖 Generated with Claude Code

[danielhertz1999-bit]

Heads up: the red build & test (macOS) here is pre-existing on main, not caused by this change.

This PR's branch differs from main by exactly 4 lines in one file (src/iai_mcp/hippo/_db.py), and the failures are all in unrelated suites:

• tests/test_mcp_tools.py:36 → NameError: name 'tmp_path' is not defined — a module-scoped daemon_sock fixture references the function-scoped tmp_path fixture (which isn't a parameter). This is in main as-is.
• OSError: AF_UNIX path too long (test_capture_transcript_no_spawn_defer, test_concurrency) — the macOS runner's pytest-of-runner/... tmp path exceeds the 104-char sun_path limit.
• daemon socket-bind timeouts (test_bridge_socket_first, test_concurrency) — the known flaky daemon-subprocess tests.

The same job on the Release v1.2.0 push to main (run 28267001306) fails with the identical test_mcp_tools.py:36 NameError — i.e. the gate was already red before this branch existed. The HNSW boot tests this change actually affects (test_daemon_hippo_boot, test_hnsw_reuse_recall_parity) pass.

Happy to send a separate PR fixing the macOS-CI issues (the tmp_path→tmp_path_factory fixture bug is a quick one; the AF_UNIX path-length one needs a short socket dir) if that's useful — just didn't want to scope-creep this 4-line correctness fix.

🤖 via Claude Code