Expose raw semantic cosine scores in search results

[yasyf]

Hybrid search currently returns only the fused RRF score, which reflects rank position rather than absolute semantic relevance — a top-ranked result over weak matches is indistinguishable from a genuinely strong hit (RRF caps near 1/61 per leg before boosts).

This adds a semantic_score field to SearchResult carrying the dense leg's raw cosine similarity, threaded through fusion and both reranking paths without changing ranking behavior. BM25-only results expose semantic_score: None. The field is included in CLI and MCP output.

Motivation: agent consumers (we drive semble from the ccx/cc-context MCP facade) need an absolute relevance signal to detect all-weak result sets instead of mistaking top-ranked junk for evidence.

• Additive only; no ranking or API behavior changes
• Exact known-distance tests cover the cosine conversion (a sign inversion now fails), plus BM25-only None coverage
• pytest green, 100% coverage maintained, pre-commit hooks pass

[greptile-apps]

Confidence Score: 5/5

This looks safe to merge.

• No blocking issues found in the changed code.

_{Reviews (1): Last reviewed commit: "Expose semantic cosine scores in search ..." | Re-trigger Greptile}

[stephantul]

Hey, please read the CONTRIBUTING.md and open an issue first. Also, we use PR templates for a reason, please don't ignore them. Closing for now.