RFAI-07: Hybrid retrieval, duplicate calibration, and memory-assisted generation

[Chris0Jeky]

Context

Week 7 from taskdeck-12-week-roadmap-v4.md.

Parent: #972
Depends on: #978

This issue turns semantic memory into product behavior: better context, better duplicate suppression, and evidence-linked proposal generation.

Scope

• Implement Reciprocal Rank Fusion over FTS5 BM25 and vector cosine results.
• Build a hand-labeled holdout for duplicate detection and retrieval evaluation.
• Calibrate duplicate thresholds empirically. Starting values from the roadmap are search points, not fixed acceptance criteria.
• Add near-duplicate detection at ingest with a clear similar to existing chip or equivalent review cue.
• Expand Capture and Chat context with EvidenceLinks from retrieval.
• Add retrieval eval fixtures and recall/precision reporting.

Acceptance Criteria

• Retrieval recall@10 is measured against a labeled holdout.
• Near-duplicate suppression precision is calibrated with a precision-favoring tradeoff.
• False-positive duplicate behavior is safe and reviewable.
• Proposal generation can cite retrieved board/knowledge context via EvidenceLink reason chips.
• FTS-only fallback remains valid when vector search is unavailable.

Suggested Verification

• Retrieval integration tests over fixed fixtures
• Duplicate calibration report committed with methodology
• Capture/Chat proposal tests proving EvidenceLinks resolve