Roll-up grouping quality: cards that read like topics

CHANGELOG.md

@@ -2,6 +2,8 @@
 
 ## Unreleased
 
+- Roll-up proposal quality overhaul: structural label hygiene (pronoun/contraction/closed-class/light-verb heads never title cards), store-measured generic-anchor detection (frequency-derived per store, no hardcoded topic list), broken-subspan label repair ("Us Part II" → "The Last of Us Part II") applied to card titles and instance lines, a group-utility gate (groups must aggregate distinct values or sessions with a coherent, specific label — failures are dropped, not proposed), utility-ranked proposals under the cap, and topic-shaped card titles with dominant value units. Measured on aggregation-heavy stores: junk-label rate 34% → 0%, instance-line defects 13% → 0%, with review proposals now reading as human-recognizable topics.
+
 - Deterministic retrieval ordering: every equal-score tie (RRF fusion, graph and temporal stages, FTS/vector stage runs) now resolves through a content-stable cascade (event date, content length, text, capture fingerprint) instead of falling through to row ids — re-ingesting the same content yields byte-identical packs (two-seed divergence 7/40 → 0/40); disclosed as `fusion.tie_break: content_stable_v1` in retrieval traces.
 - Benchmark harness gains a disclosed `--accept-rollups` step (default off, fingerprint-stamped) that review-accepts consolidation roll-up proposals through the real acceptance path, modeling the product's human review workflow; offline measurement found current roll-up grouping quality too noisy to help the benchmark, so the flag stays off and grouping quality is queued as product work.
 - Grounding for products: the context-pack entity note now recognizes quantity-qualified and possessive compound entities ("30-gallon tank", "my snake plant") with conservative false-positive guards, and a new answer-verification library seam (`vnext_answer_verification`) lets integrators opt into post-generation grounding checks; nothing changes by default.