Improve search relevance scoring

The search index in crates/core/src/lib.rs (SearchDocument::search, lines ~360-376) is a naive substring matcher that assigns fixed bucket scores (100/75/25) and requires every term to appear via path.contains(term).

Problem

• Fixed bucket scoring (100/75/25) does not reflect actual relevance.
• Every term must appear via path.contains(term), so partial words and typos never match.
• split_whitespace over the raw query (lines ~317-322) produces crude tokens that won't match partial words.

Suggested direction

• Replace fixed buckets with term-frequency / proximity-based scoring.
• Support fuzzy and prefix matching.
• Add proper tokenization (stemming, word-boundary handling) instead of split_whitespace over the raw query.

References

• crates/core/src/lib.rs — SearchDocument::search (lines ~360-376)
• crates/core/src/lib.rs — query tokenization (lines ~317-322)

Start in crates/core/src/lib.rs by reading query tokenization around lines 317-322 and SearchDocument::search around lines 360-376. Trace the current substring and fixed-bucket scoring behavior, then define and validate relevance behavior that covers improved tokenization, prefix or fuzzy matching, and term frequency or proximity instead of requiring exact term presence.