We've added a deterministic, term-aware lexical scorer to improve repo_context focus queries without requiring an expensive model runtime. This update supports natural-language retrieval by splitting identifiers, applying conservative inflections, and rewarding concept coverage, while leaving literal repo_search exact-substring behavior untouched. To ensure long-term retrieval accuracy, we've gated 19 PHP retrieval judgments across multiple language versions as part of our standard blocking CI and test suite.