refactor(index): close IndexStore connections deterministically

[Hotragn]

Summary

IndexStore is the only store in the codebase that does not close its SQLite connections. All eight of its methods rely on sqlite3.Connection.__exit__, which scopes a transaction and returns without closing.

Adds open_index, which nests the existing transaction context inside a try/finally that closes. Behaviour is otherwise identical.

Why

Every other store routes through codealmanac.database.open_local_database, which closes in a finally and carries a comment explaining exactly why it exists:

sqlite3.Connection's own context manager scopes a transaction and never closes; long agent runs record hundreds of events and leaked descriptors until sqlite3.connect failed with "unable to open database file".

IndexStore bypasses that helper, so it never received the fix:

with connect_index(db_path) as connection:   # x8 — commits, does not close

Confirmed by grep that repositories/store.py, runs/store.py, runs/events.py, runs/worker_locks.py and telemetry/store.py all use open_local_database, and services/index/store.py is the sole holdout.

It matters more here than elsewhere because every query command reindexes implicitly, so this is the hot path rather than a rare one.

Scope note — what I am not claiming

I could not reproduce descriptor exhaustion from this path, and psutil's open-file reporting on Windows was not reliable enough to measure it (200 deliberately-retained connections reported as 3 handles, unchanged after gc.collect()). So please read this as deterministic resource hygiene and consistency with the pattern the repo already chose everywhere else, not as a fix for an observed crash. I would rather say that up front than have you find it in review.

What is verifiable is the contract: after this change the connection is closed when the block exits, and the test proves it.

Verification

# Windows 11, Python 3.13 (uv-managed), uv 0.12.0
uv run pytest tests/test_database.py tests/test_read_model.py \
              tests/test_architecture.py -q
# 72 passed

uv run pytest -q          # 13 failed, 551 passed, 1 skipped
# baseline on main:         13 failed, 550 passed, 1 skipped   (+1 = the new test)

uv run ruff check .       # All checks passed!
git diff --check          # clean

The new test fails on the previous code:

FAILED tests/test_database.py::test_index_store_connections_close_after_every_use

Docs and wiki

• Not applicable — internal resource handling; no user-facing behaviour or output change.

Notes for reviewers

• Nesting with connection: rather than replacing it is the load-bearing detail. None of the eight callers commits explicitly — IndexStore has zero commit() calls and replace_documents does not commit either. So they depend entirely on __exit__ for durability, and swapping it for a plain closing context manager would silently stop persisting the index. Nesting preserves commit-on-success and rollback-on-error exactly; the only change is that the handle is released.
• connect_index is deliberately kept, both because open_index builds on it and because tests/test_architecture.py:187 pins "def connect_index" in schema. I did not want to quietly weaken an architecture test to make a refactor fit.
• store.py line count is unchanged at 131, so the <= 160 budget in test_architecture.py still holds with room to spare.
• The test mirrors the existing one for RepositoryStore (test_store_connections_close_after_every_use) rather than inventing a new style — same tracking-connect monkeypatch, same closed-connection assertion.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}