Match business questions to code with native concept search

[forhappy]

Business-language questions now use bounded native synonym recall and witnessed document links to find implementation code. ask, non-exact search, and MCP discovery disclose how each approximate match was reached; exact lookup and structural evidence keep their existing contracts.

An explicit --semantic-search / semantic_search option enables local TF-IDF/truncated-SVD fallback. It uses only graph names and bounded prose, requires no model, credentials or network service, and returns typed corpus/deadline limits. Valid disposable FTS indexes are reused on reopen instead of being rebuilt after a read-only integrity check.

The fixed 30-question Dify corpus reaches its expected module in the first three source paths for 26/30 questions (86.7%); 16/30 rank it first. A fresh AST-v15 graph reproduced the result. Four misses and all truncation flags are preserved in the checked-in qualification report. This is module retrieval evidence, not implementation equivalence or an independent human/competitor comparison.

Compatibility: additive conceptMatches and bounded resource content fields require strict decoders to update. AST cache v15 and JSON index v2 refresh prose facts; existing store indexes need republishing. Historical realizations remain immutable. See COMPATIBILITY.md, MIGRATION.md and SECURITY.md.

Validation:

• Native baseline: formatting, workspace Clippy with -D warnings, 1,117 tests passed (2 existing ignored).
• Query, JSON/store, Markdown publication, language, CLI/product and MCP contract regressions passed.
• 500-query relevance gate and CompassQL TCK/support gates passed.
• JavaScript typechecks, unit/browser suites (120 browser tests), UTF-8 boundary regression and deterministic viewer assets passed.
• Complete code-graph fixture qualification, including final React offline/activation/anchor checks, passed.
• Fresh pinned Dify extraction and all 30 native qualification queries passed the fixed 24/30 threshold.