LexiconForge

Systematically investigate a queue of bug claims with verdict classification, spec-vs-code matrix (A/B/C), cross-cutting theme noticing, and a hard closing gate that prevents wrong-layer fixes. Produces durable evidence artifacts (per-issue READMEs, theme docs, regression tests). Includes archaeology tooling that maps Claude Code transcripts to git commits to identify the agent + prompt + tools that introduced a bug.