dikw-core v0.6.5

0.6.5 — Default scaffold ships Gitee embed + rerank; eval cache keys retrieval config and surfaces absolute relevance scores

Added

• Eval rows surface absolute relevance scores for OOD calibration.
  (#249) Each retrieval-eval per-query and negative row now carries
  top1_score (the top hit's score) and top1_vec_cosine — the
  reranker/fusion-independent raw top-1 vector cosine, captured by an
  eval-internal HybridSearcher.top_vector_cosine probe (the production
  search() path and its ranking are untouched). Fusion scores (RRF is
  rank-based; combsum/combmnz are per-leg min-max normalized) can't carry an
  absolute magnitude, so expect_none / out-of-distribution robustness was
  previously immeasurable from rank order alone; the absolute cosine makes it
  observable (covered query high, OOD query low). The vector probe is skipped
  for pure-bm25 ablations so they stay embedding-free. A score-based OOD
  metric is deferred — this release only surfaces the signals.

Changed

• Default scaffold ships Gitee embed + rerank; unified rerank/embed
  degrade-logging. dikw init (config.default_config) now defaults the
  embedder to Gitee bge-m3 (dim 1024) and ships a Gitee
  bge-reranker-v2-m3 reranker, both keyed by one GITEE_API_KEY, so a
  fresh base reranks out of the box (OpenAI has no /rerank endpoint, so the
  prior OpenAI embedding default couldn't pair a matching reranker). The LLM
  default stays Anthropic. Read-path resilience is now uniformly observable: a
  transient query-embedding failure degrades the hybrid query to FTS-only
  (vec leg dropped) instead of 500-ing — hybrid mode only, single-leg
  vector/bm25 ablations re-raise for eval purity; transient rerank /
  embed-batch-skip degrades now log at ERROR (a configured leg that failed,
  was WARNING); an enabled-but-unconfigured reranker and a write path that
  defers embedding (no embedder wired / version drift) each log a WARNING so
  the silently-off leg is visible. Permanent provider errors (401/403/404,
  bad key/model) still fail fast on both the read path (→ 500) and the write
  path (ingest aborts) — the fail-fast-on-misconfig invariant is unchanged.

Fixed

• Eval snapshot cache keys the ingest-time tokenizer and reads query-time
  retrieval config live. (#250) The eval corpus-snapshot cache key omitted
  RetrievalConfig, so under the default --cache read_write changing any
  retrieval knob (rrf_k / weights / fusion / rerank_enabled / graph_*)
  and re-running silently hit the stale snapshot and reused the previous
  config — no error, wrong numbers, exactly on the retrieval-ablation workflow.
  The cache key now includes the only ingest-time retrieval field,
  cjk_tokenizer (a change forces re-ingest); every search-time knob is read
  from the live config on each _run_queries, so ablations sharing one
  cache_root are now both fast and correct. A defensive guard re-raises if a
  cache hit's baked tokenizer ever disagrees with the live one.