perf(audit): use WAL journal to avoid fsync per commit on event loop

[hi-neason]

Summary

While reading through the engine I kept seeing self._audit(...) called directly inside the async loop — 17 of them — and each one ends up at AuditStore.append, which does an INSERT ... commit() against SQLite.

AuditStore never sets a journal mode, so SQLite runs with the defaults: DELETE journal + synchronous=FULL. That means every audit commit is an fsync. With a couple of tool calls per turn that's a handful of syscalls, but on a busy turn (multiple tools, retries, plan/approval stages) it adds up, and because the sink is invoked synchronously from inside TurnEngine._audit, each fsync parks the whole event loop — roughly 1–10ms per audited tool stage on a typical disk.

The fix is two lines at connect time:

• PRAGMA journal_mode=WAL
• PRAGMA synchronous=NORMAL

WAL means writers don't block readers and commit doesn't fsync; synchronous=NORMAL is the recommended pairing with WAL and is still durable across crashes (the WAL checkpoints to the main DB on close/rotation). Transaction atomicity is unchanged — a crash mid-commit still rolls back. No API or behavior change beyond the pragma.

Change

• coworker/audit.py — set both pragmas right after connecting, before the CREATE TABLE IF NOT EXISTS runs. Added a short comment explaining why (since the reason for the pragma is non-obvious without knowing how hot _audit is).

Tests

I noticed coworker/audit.py had no test file at all, even though it's responsible for redacting secrets before they're written to coworker.db. So along with the regression test for the pragma I added tests/test_audit.py covering the parts that are silent if they regress:

• test_audit_store_uses_wal_journal — asserts journal_mode == "wal" and synchronous == 1 (NORMAL). This is the one that actually catches the perf bug; on main it fails with assert 'delete' == 'wal'.
• test_append_and_list_roundtrip, test_list_filters — basic storage shape and the session_id/tool filters.
• test_sanitize_args_redacts_secrets — every key in _SECRET_KEYS / _BODY_KEYS plus the *_token suffix path.
• test_sanitize_args_redacts_browser_type_text — browser_type's text field gets [redacted input] (form input, not a body blob).
• test_resource_extracts_well_known_keys — url / owner-repo / zendesk subdomain / result-URL fallback.
• test_close_is_idempotent and test_summarize_truncates_long_strings.

I confirmed the new pragma test fails without the fix (stashed audit.py, saw assert 'delete' == 'wal', then popped) and passes with it.

Full suite: 1006 → 1014 passed, 1 skipped. The 9 failures present locally (test_fake_slack ×8 needs the messaging extra / slack-bolt, plus one test_ui_refresh_e2e) are pre-existing on a clean main and unrelated to this change.

No screenshots — backend-only (SQLite pragma + tests). Happy to add anything else that'd help review.

Checklist

• pytest tests -q passes
• ruff check / ruff format --check clean on the added test file
• No hardcoded API keys, tokens, or internal addresses
• No new dependencies
• Regression test added, verified to fail without the fix
• Conventional commit message