feat: Semantic cluster merging via pgvector

[leo-aa88]

Summary

Use pgvector embeddings to merge or relate clusters that are semantically similar but diverge in text (different templates, minor wording), improving explain/timeline/compare quality.

Motivation

• README Roadmap lists semantic cluster merging.
• Fingerprinting alone can split one incident across multiple clusters when normalization does not fully align.

Scope (proposal)

• Define merge policy: when to merge vs link (threshold, min cluster size, only post-ingest analysis?).
• Pipeline hook after clustering or as an optional analysis pass for explain / clusters.
• Respect determinism goals: document behavior when embeddings are disabled.

Acceptance criteria

• Measurable reduction in duplicate “near duplicate” clusters on sample data without breaking grounded counts (or clearly documented semantics).
• Tests that do not require live embedding API where possible (fixtures / mocks).

Risks

• Merging can hide distinct errors if thresholds are wrong; make thresholds configurable and conservative defaults.