v0.8.4

The dense+sparse hybrid retrieval arm — and the correction of an earlier wrong "infeasible" call.

Added

• Hybrid dense+sparse retrieval (BM25 full-text + 3-way RRF). A lexical/sparse arm now joins the dense vector arm and the keyword/tag arm in the injection pipeline. schema.surql defines a fts_analyzer (blank/class/camel/punct tokenizers + snowball English stemming) and FULLTEXT BM25 indexes on concept.content, turn.text, memory.text, artifact.description, skill.description. SurrealStore.fulltextSearch() runs a per-term @n@ OR query (recall) ranked by summed search::score, unioned into the candidate pool and folded into the graph-seed Reciprocal Rank Fusion. This adds recall of exact-term / rare-token / code-identifier rows the dense embedding misses entirely. (BGE-M3's sparse heads aren't reachable via node-llama-cpp, so the sparse arm is SurrealDB-native BM25 — no model sparse head needed.)
• Offline retrieval benchmark (scripts/retrieval-benchmark.mjs) — NDCG/MRR/Recall over the labeled retrieval_outcome table, to gate ranker changes with a number instead of a guess.
• scripts/drop-test-namespaces.sh — durable cleanup for the vitest temp namespaces (kctest_*) that accumulate on the local SurrealDB instance.

Changed

• Managed SurrealDB binary pinned 3.0.5 → 3.1.4 (bin-manifest.json) — aligns fresh installs with the version everything is tested against. Deliberately not 3.1.5: issue #7383 is a 3.1.5-specific RocksDB OOM under full-text load that would regress the new hybrid. All 5 per-platform sha256 from the GitHub release digests; linux-x64 download-verified.

Notes

• SurrealDB 3.x full-text BM25 uses FULLTEXT ANALYZER (renamed from 2.x SEARCH ANALYZER in 3.0.0-beta). search::score returns 0 on a near-empty index (#7290 fresh-DB corpus-stats edge) but is correct at real scale — fulltextSearch() falls back to a term-frequency rank for the empty-index case. skill indexes description, not body (full skill bodies are large enough to stall the FULLTEXT build / schema-apply).