CodeAbra/iai-mcp

Changed

• Recall ranks curated knowledge above raw chatter. Knowledge-grade sources
  get a soft multiplier at final rank (default 1.05, IAI_MCP_TIER_BOOST, 1.0
  disables), and it stands down for verbatim cues. Teach chunks (doc:*) boost
  unconditionally — they are literal curated content — while semantic summaries
  boost only when the literal_preservation profile knob is relaxed past its
  strong default, because that knob is precisely the raw-versus-summary
  preference. A default install still keeps raw originals on top.
• A query worded differently from the stored text can surface its target.
  The in-tree BM25 lexical index now fuses into recall: when the cue carries a
  genuinely rare token (IDF-gated, IAI_MCP_LEX_MIN_IDF), warm BM25 hits join
  the seed set and earn a bounded rank bonus. IAI_MCP_LEX_FUSION_OFF=true
  kills the lane. The recall path never pays the index rebuild — the index is
  built by scoped search or the nightly warm-up and kept current by a cheap
  per-insert feed, and a cold index contributes nothing.
• Exact-similarity authority no longer wipes a hit's graph rank. When a
  record surfaces through both lanes the richer pipeline score survives the
  merge, so authority keeps guaranteeing inclusion without flattening the final
  order back to pure cosine.

Fixed

• A freshly created store no longer inherits a dead store's unflushed rows.
  Write buffers key on the object id and Python reuses freed addresses, so a new
  store could silently take on another store's pending content. Buffers are now
  purged at store construction.
• A knowledge summary's consolidation edges survive a crash at mint. They
  flush to the edges table immediately instead of waiting in the write buffer
  for a threshold flush.

Both ranking levers came out of a recall-quality measurement on a real 18k-record
store by @Marsu6996, who also prototyped the
boost factor and the rare-token gate.