docs: pin Wave 1 local LoCoMo 14/30 (honest remasure)

[sdntsng]

Summary

Wave 1 (#101–#105: context vs proof, temporal features, retrieval budgets, assistant-extract policy) was merge-committed onto dev (a7a5184). This PR records the local remasure. It does not change product code.

Landed on dev via merge-commit bd987fa (same local-merge path as #100–#105; GitHub merge API is 403). main untouched.

LoCoMo 1×30 on rebuilt local API+worker: 14/30 (MH 3/10, OD 2/4, temporal 9/16). Failure ledger 16/16 READER_MISS with evidence/semantic/retrieval/coverage oracles all supported. OpMem 13/13. Marketing 17/17.

This is not an improvement vs Gate 0 staging 18/30. Multi-hop did not improve (4/10 → 3/10). Do not read 14/30 as beats-Mem0: same-day Mem0 was 12/30 overall but 7/10 MH.

Confound vs local PR2 6/30: that run used harness BRAINY_USE_RECALL=1 but the API process did not have BRAINY_RECALL_LLM=1. This remasure enabled the hybrid reader so PR5 is actually exercised. Do not attribute 6→14 to Wave 1 code alone.

Changes

• Pin: docs/benchmarks/artifacts/locomo-wave1-dev-1x30-20260813.md plus JSONL ledger
• Non-reg pins: OpMem 13/13, marketing 17/17
• Status / gap matrix / assessment pack / research README updated
• PR6–PR8 remain deferred (coverage oracles still supported)

Test plan

• go test ./internal/memory/ ./internal/jobs/ ./internal/api/ after Wave 1 merge (on dev)
• Local ingest + /healthz after rebuild
• python3 evals/run_opmem.py --systems brainy → 13/13
• python3 evals/run_vertical_eval.py --fixture-dir fixtures/vertical/marketing → 17/17
• python3 -m public.locomo.run_smoke --conversations 1 --questions 30 --run-id locomo-wave1-dev-1x30-20260813 → 14/30
• Optional after this pin lands: staging 1×30 vs post-cutover 15/30 (local vs Render confound)

Linear / tracking

Competitive program Wave 1 remasure. No new Linear ID.

Risk / follow-ups

• Remaining driver is still reader/enumerate emitting chat continuation, not missing edges. Do not start PR6–PR8 from this pin.
• Next measured loop: /recall answer/enumerate quality (fact, not dialogue dump).
• main / production was not touched. Merge this pin to dev only unless explicitly asked.
• LME-20 quality remains 0/20; do not start LME-100 as a quality run.