feat(DRIVE): hybrid RAG (keyword + e5-large embedding) for NL edit grounding

[dancinlife]

What

DRIVE test-drive REPL now grounds the local model on curated fix-recipes via hybrid retrieval — keyword-precision first, embedding-recall fallback.

STAGE 1 keyword substring  →  exact, 0.02s, no model load (queries quote keys)
STAGE 2 in-process fastembed (multilingual-e5-large, query:/passage: prefixes)
        →  paraphrase fallback ONLY when STAGE 1 empty · conservative 0.80 thresh

Self-contained: no embedding server / LAN / VPN — fastembed runs in-process via ~/.drive-rag-venv. drive.hexa calls rag_retrieve.py, degrading to its inline keyword scan if the venv is absent. fix_recipes.txt is the runtime-editable KB.

Honest finding

Embedding alone underperforms keyword on this KB — MiniLM had an "area" attractor out-ranking correct matches; e5-large fixes it but still misses ~1/4 of in-vocabulary queries. Hence keyword-primary, embedding-fallback (not embedding-only).

Validation — _drivesim 100-scenario harness (70 code + 30 NL-git)

	result
final	100/100 (96 first-pass + 4 model-flake recoveries on retry)
git NL (commit/push/branch)	30/30 first-pass
harness fix	input race root-caused as CR-vs-LF (drive PTY input() needs \n) — drive itself was correct

Files

• product: DRIVE/{drive.hexa, DRIVE.md, rag_retrieve.py, fix_recipes.txt}
• harness: _drivesim/{gen.py, run.sh, run100.sh, retry.sh, drive_one.exp, manifest.json}
• excluded (generated/cache): scenarios/, build/, fix_recipes.emb.json

🤖 Generated with Claude Code