perf(ci): disable semantic search in default test fixtures

[phernandez]

Summary

semantic_search_enabled silently defaults to enabled whenever fastembed + sqlite-vec are importable — which they are in dev and CI. Neither tests/conftest.py nor test-int/conftest.py overrode it, so every test that synced a note ran the real ONNX embedding stack (~5–7s per sync). The dedicated semantic suites (test-int/semantic/) configure the flag explicitly in their own conftest and are unaffected.

On top of that, the four search performance benchmarks in test-int/test_search_performance_benchmark.py force-enable semantic themselves and were running inside every CI integration job (~85s, worst single test 41s) — they aren't collected by just test-semantic and are meant to be run on demand.

Changes

• Set semantic_search_enabled=False in both shared app_config fixtures (with constraint comments)
• Deselect benchmark-marked tests from the CI integration recipes: -m "not semantic and not benchmark"
• Fix six tests that implicitly relied on the enabled default:
  • two reindex_vectors wiring tests: stub delete_stale_vector_rows (the one call in that flow that requires the semantic stack)
  • three embedding-status tests: create the search_vector_embeddings stub table that the file's other tests already use
  • the vec0 regression test: patch the config_manager property for the status call, same pattern as its unit siblings

Measured (local, same machine)

Suite	Before	After
int SQLite	337s	110s (3.1x)
unit SQLite (2903 tests)	—	254s
tests/sync/test_watch_service.py	31s	2.9s (10.7x)
worst single int test	40.2s (benchmark)	5.0s

Test plan

• uv run pytest tests -q — 2903 passed
• uv run pytest -m "not semantic and not benchmark" test-int -q — 368 passed
• Postgres int suite running locally; CI validates the full matrix here

🤖 Generated with Claude Code

Reviewed SHA: 8c0ce6d8a7e8134acdf005453a0a9f1a8e9295ab
Verdict: approve
Status: success - BM Bossbot approved this head SHA

Summary:
Reviewed the full provided diff against the engineering style guidance and the touched search/project status code paths. The default test fixtures now explicitly disable semantic search, the semantic and benchmark suites remain opt-in through their own targets, and the latest Postgres portability fix only changes a SQLite-specific monkeypatch to be non-failing on Postgres where that method is not called. No concrete merge-blocking issues found.

Blocking findings:

• None

Non-blocking findings:

• None

BM Bossbot image choices

• Pull request: #938
• Generated asset: /home/runner/work/basic-memory/basic-memory/docs/assets/infographics/pr-938.webp
• Image model: gpt-image-2
• Size: 1536x1024
• Quality: high
• Image mode: editorial-image
• Theme source: auto

Theme / visual direction:

classic black-and-white photography: documentary field report, contact sheet, street photography, civic infrastructure, and darkroom contrast