[P2] Optional embedding similarity layer for semantic comparisons

[cursor]

Summary

Add an optional embedding-based meaning similarity layer between deterministic token overlap and optional LLM judge. Example: search_issues vs find_work_items → similarity 0.91.

Priority

priority:p2 — Milestone 9 Layer 2; must remain optional with no required SDK for default install.

Suggested labels

enhancement, priority:p2, area:diff, area:probes, research

Acceptance criteria

• Deterministic mode still works without embeddings
• Optional extra/provider for embeddings; thresholds configurable
• Model/provider metadata stored with results
• Feeds rename confidence (Multi-signal rename detection with confidence scoring #80) and collision scores (Detect tool collisions and emit confusability scores #79) when enabled
• Fake/stub tests in CI; live embeddings opt-in
• Labeled as HYBRID or MODEL-BASED appropriately in reports

Related

• Detect tool collisions and emit confusability scores #79 collisions, Multi-signal rename detection with confidence scoring #80 rename, Optional LLM semantic judge for ambiguous tool comparisons #81 LLM judge