feat(absence): BM25-fed did-you-mean in absence_diagnosis (with MRR eval)

[HumanBean17]

Tracks deferred work from #431 (fork B). Design ref: docs/superpowers/specs/active/2026-07-12-hybrid-bm25-rrf-design.md.

Why

absence_diagnosis today uses vocabulary q-gram similarity + difflib did-you-mean. LadybugDB BM25 — now available on every search as a first-class signal after #431 — beats both for exact-identifier / typo cases (DistributionChunkService, processClientMessage).

Scope

• When BM25 returns strong identifier matches the vector path missed, surface them as "you may have meant X" in the absence / did-you-mean output.
• Wire the BM25 candidate list (already computed always-on after feat(search): hybrid BM25 + vector ranking via RRF (promote lexical to first-class) #431) into absence/absence_diagnosis.py:diagnose.
• Add an MRR-over-suggestions eval for did-you-mean quality (separate metric from the feat(search): hybrid BM25 + vector ranking via RRF (promote lexical to first-class) #431 ranking eval).

Why deferred from #431

Kept rank-only in #431 so the ranking eval (recall/precision @k) stays cleanly attributable. Did-you-mean quality is a different metric needing its own harness.

🤖 Generated with Claude Code