feat(capture-attention): Phase A — recurrence-weighted preserve+relate+boost

[cipher813]

Summary

Adds an embedding-only capture attention path to mnemon that detects cross-session restatements at save time. When a new memory's content is semantically close (cosine ≥ 0.85) to ≥2 prior memories spanning distinct sessions, the new memory is preserved, 'restates' relations are inserted from new → each cluster member, and the canonical neighbor's confidence + new recurrence_count column accrete the load-bearing signal.

Default-off behind CAPTURE_ATTENTION_ENABLED for soak. Sibling track to the salience-tier work — both target the same failure mode (load-bearing facts under-weighted) from different angles: salience tier is recall-side weighting, capture attention is upstream curation.

Driver

Surfaced 2026-05-22 during salience-tier Phase 0 validation attempts: load-bearing facts (runway composition, career posture) stated "numerous times" in chat landed in mnemon as scattered fragments, never as a single canonical assertion. The operator was implicitly substituting for a missing mechanism — that's the gap this PR closes. Brian's framing: "I shouldn't have to compose memories for them to be accurately treated."

Plan doc (gitignored, private/mnemon-capture-attention-plan-260522.md) contains full invariants, phase table, and soak gates.

SOTA correction baked in

An earlier draft of this work considered "boost canonical + skip the new save" as the auto-apply path. Rejected during pre-implementation SOTA audit: each restatement carries different framing, and discarding it throws away the very signal the recurrence detector is honoring. The institutional pattern for dedup-with-importance in production retrieval systems is preserve+relate+boost — keep the data, link via relations, accrete importance on the canonical. MMR diversity at recall naturally suppresses near-dups without us dropping them at capture. Operator-reviewed merge stays in Phase C, not auto-applied here.

What ships

• documents.recurrence_count schema migration (additive, idempotent on re-open, harmless if flag stays off)
• Store.apply_capture_attention() + private helpers (_resolve_neighbor_docs, _pick_canonical, _boost_confidence, _increment_recurrence)
• Store.save() integration — gated by CAPTURE_ATTENTION_ENABLED; skipped when correction_of is set
• correction_of parameter on Store.save() (forward-compat for salience-tier Phase 2 promotion signals — operator gesture beats automated detection)
• CaptureAttentionUnavailableError named exception; save() catches + WARNs per acceptable-swallow category (secondary observability hung off primary save path)
• mnemon attention-status CLI — soak monitor showing boost-rate ratio, recurrence distribution, top canonicals, recent 'restates' audit trail
• scripts/calibrate_capture_threshold.py — data-tuned threshold selection from a vault snapshot (sample N pairs → tag → precision-recall at {0.70, 0.75, 0.80, 0.85, 0.90} → recommend precision-leaning sweet spot)
• tests/fixtures/capture_attention_pairs.json — regression-lock seed; replaced by operator-tagged pairs via the calibration script
• 13 new tests in tests/test_capture_attention.py

Composability

• Layer 0 (is_well_shaped) — runs BEFORE capture attention; scaffolding never reaches the path
• Layer 4 (HOOK_SOURCE_CONFIDENCE_CEILING) — boost respects the 0.5 ceiling for hook-sourced canonicals
• Existing relation types ('supersedes' / 'contradicts' / 'related') — new 'restates' type, no collision
• Embedding-only — no LLM dep, per the 2026-05-21 public-release decision (LLM-judge stays as roadmapped P2 opt-in)
• Salience tier Phase 1 (next track) — promotes from better-curated content once this ships

Soak gates (before flipping default-on)

1. boost_rate ≤ 0.25 over a 7-day window (measured via mnemon attention-status)
2. ≥80% precision on a 20-canonical manual review

Test plan

• All 13 new tests pass (pytest tests/test_capture_attention.py)
• Full suite 801 → 814 passing, no regression
• mnemon attention-status smoke-tested against a temp vault (flag-off path)
• Operator runs scripts/calibrate_capture_threshold.py against prod-snapshot vault to validate threshold = 0.85 is reasonable for their content distribution (post-merge, before flipping flag on)
• Operator flips CAPTURE_ATTENTION_ENABLED=True for ≥1 week soak; observes mnemon attention-status weekly
• Manual review of 20 boosted canonicals after soak; precision ≥ 0.80 → flip default-on

🤖 Generated with Claude Code