Summary

• clean English FTS query terms by dropping high-frequency question glue and punctuation before FTS5 OR matching
• preserve stopword-only queries, Korean/non-ASCII terms, and non-ASCII Latin tokens
• document the MS MARCO 1M reuse improvement

1M result

On the local persistent MS MARCO 1M SQLite DB:

• before: build 0.0s, search 70.1s, MRR@10 0.462, Hit@10 30/50
• after: build 0.0s, search 9.1s, MRR@10 0.479, Hit@10 31/50

Tests

• uv run --extra dev ruff check src/synaptic/backends/sqlite.py tests/test_backend_sqlite.py
• uv run --extra dev ruff format --check src/synaptic/backends/sqlite.py tests/test_backend_sqlite.py
• uv run --extra dev pytest tests/test_backend_sqlite.py tests/test_tier1_benchmarks.py -q
• PYTHONUNBUFFERED=1 uv run --extra sqlite python examples/ablation/run_tier1_benchmarks.py --only msmarco --subset 50 --corpus-limit 1000000 --use-sqlite-graph --sqlite-db-path tests/benchmark/data/msmarco_1m.db --reuse-sqlite-db --progress-every 100000