docs: record AI-normalization review rules in CLAUDE.md

Rules from the post-normalization review: don't silently change what existing
models return, verify doc claims against every code path they cover, record
self-disclosed defects in the PR draft, finish with a fresh build + suites, and
check git stash before concluding an edit was lost. Plus the standing
typecheck-baseline no-go.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

CLAUDE.md

@@ -1 +1,19 @@
 FOLLOW ./AGENTS.md
+
+- Any change that alters the response shape of an existing (pre-cutoff) model — as the
+  Responses-API finish_reason/reasoning change in OpenAIUtil.js did — must sit behind the
+  normalize policy resolution or be explicitly approved first. Equalization work never
+  licenses silent changes to what current models already return.
+- Before writing a behavioral claim into chat.md/chatresponse.md, verify it against every
+  code path it covers: the Mistral remap and the Anthropic coercer handled unmapped
+  finish_reason differently, and the shipped sentence was true for one and false for the other.
+- When you self-disclose a defect mid-session, immediately record it in a Known Gaps section
+  of the PR draft — disclosures that live only in the transcript do not survive to review,
+  and a prepared PR silently ships them.
+- Make the final act before handoff a fresh `npm run build:workerLib` followed by the suites,
+  stating exit codes; never pipe a suite run through grep when the exit code is the evidence.
+- When a user's uncommitted edit "vanishes" from the worktree, check `git stash list` before
+  concluding it was lost — GitHub Desktop auto-stashes on branch switch as
+  "!!GitHub_Desktop<branch>" and `git stash pop` recovers it.
+- `tools/typecheck-baseline.json` is a permanent no-go zone: never edit it and never run
+  `typecheck:update`, subagents included.