[JayantDevkar]

Problem

Six tests in test_background_shells_router.py fail on any machine with real Claude Code data and take 40 to 50 s each, while CI stays green (#93). GET /shells calls _maybe_sync(), which scans settings.projects_dir, and the fixture isolated the DB path but not the ingest source, so the real ~/.claude/projects was ingested into the temp DB. The authors of #97 and #101 both hit this and had to explain it in their PR bodies.

Changes

• tests/api/conftest.py: an autouse fixture points settings.claude_base at an empty per-test temp dir, so no API test can read the real ~/.claude. With no projects/ subdir, an accidental sync_all_projects() returns early, as it does on CI.
• test_background_shells_router.py: the old fixture becomes shells_app with the lazy sync left live and the debounce clock reset; client stubs _maybe_sync, so every existing query test is unchanged. A new TestListShellsLazySync builds a fake projects dir with one background Bash session and asserts GET /shells ingests it (a negative control stubbing sync_all_projects fails with 0 == 1).
• background_shells is the only router that syncs lazily; admin.py syncs only on the explicit POST /admin/reindex.

Test code only. No production change.

Verification

On this machine with real data, same interpreter:

	before	after
test_background_shells_router.py	6 failed, 10 passed in 402 s	17 passed in 0.6 s
full api suite	6 failed	1929 passed, 6 skipped in 8.8 s

ruff check tests/ routers/ clean, ruff format --check clean.

Follow-ups

• Found while instrumenting: five tests in tests/api/test_set_session_title.py open the real ~/.claude_karma/metadata.db through services/session_title_cache.py, and test_set_title_success runs ensure_schema on it via get_writer_db(). No rows change (fake uuid), but it is the actual leak the memory note blamed on the app lifespan. Tracked in a separate issue.
• The pre-fix real-data sync logged NOT NULL constraint failed: background_shells.spawned_at for several real sessions; looks like a separate extractor bug on real JSONL. Same issue.