Stop leftover covering as LoCoMo 180 benchmax

[sdntsng]

Summary

Isolated leftover covering on the skip-ingest LoCoMo 180 is benchmaxxing-adjacent and must stop. Product hybrid /recall on tenant diag-mh-135 is 137/180; the industry lane on the same tenant is still 62/180; full product /recall is still 175/1540 = 11.4%. 137/180 is not 90%, not n=1540, and not a Mem0 same-pin.

Each covering detector is generic English, but the next detector was queued from this 180’s remaining ledger. That is Goodhart. internal/memory/recall.go is now ~50% leftover covering (24 query-shape detectors, 58 covering-line helpers). P28→P53 was +24, almost all one-item recoveries. Cycle-closeout already called covering saturating; we kept going anyway.

No P54. Do not add a peaceful/nature/miss-about covering for conv-44-q62. Remaining 43 misses are WRITE, count dumps, incomplete lists, OD hypotheticals, relative dates, and steal-slots. Optimistic covering ceiling on this 180 is ~141/180, still 21 short of 162.

Beating public benchmarks still means: compiler WRITE (S1), entity-scoped enumerate (S2), industry lane top-k 200 (S5), fair Mem0 same-pin after quota 2026-09-01, n=1540 only at S6.

Existing leftover covering stays. This PR does not rip it out and does not change recall behavior beyond a stop comment.

Changes

• Add docs/research/competitive/benchmax-audit-2026-08-25.md (verdict, pin mixing table, remaining-43 classes, stop/continue).
• Cycle-closeout 2026-08-25 honesty stop with required competitor compare (no new Mem0 score; do not refresh lead/trail from 137 vs 11).
• Kill-list updates: AGENTS.md, sota-execution-plan.md, public-bench-ladder.md, competitive/research READMEs, benchmarks README honesty-stop row.
• Comment on leftoverCoveringRareForQuery so the next agent does not add another query-shape detector from this 180.

Test plan

• P53 ledger still 43 misses (RETRIEVAL 17 / PROOF 14 / READER 7 / WRITE 4 / HARNESS 1) — no new 180 run (no product recall change).
• gofmt -l internal/memory/recall.go clean (comment only).
• go vet ./... / go test ./... not required for docs + comment; run if CI wants them.
• Confirm README 11.4% / 70% 1×30 were not replaced by 137/180.

Linear / tracking

• ENG-168 (conversational long-memory epic)
• ENG-176 (MH) — leftover covering was not the MH lever; MH still 18/33 on this 180, 7.4% on n=1540.

Risk / follow-ups

• Kill list is the main control; a determined agent can still add a detector. Reviewers should reject any PR whose unique gain is one remaining 180 item via a new looks*Query.
• Next genuine product work: S2 entity-scoped how-many (Melanie 7 vs 3, pets 13, ankle 38) without a named-person rule; or S1 compiler with re-ingest; then current-SHA industry lane.
• Fair Mem0 180 still quota-blocked until 2026-09-01.
• Do not FF main with this unless explicitly asked. Staging dev should pick it up so later agents see the stop.
• Do not merge S1–S5 compiler/answer/alias/KU (split from mixed S0 PR) #133 / S0–S5 SOTA execution increments (stratified harness + compiler/answer/hop/LME) #131 / fix(recall): instrument-purpose leftover covering over help flood #143 / P29: demote single-token attended compiler echoes (ranking only) #145.