Authoring rules for comments, commit messages and PR descriptions are missing or one line long #658

CLAUDE.md gives one line to comments and nothing at all to commit messages or PR descriptions, so agents pattern-match on what is already in the tree — which is exactly what should not be copied.

What the tree teaches today:

• 1,705 comment lines in src/ and test/ carry historical or coordinate markers. Live examples: /// <b>Round 2 Finding 2 — the selector object is the only source of truth:</b> (AcpVendorDescriptor.cs), /// §2.7 B6 arm-A: an upper bound on the WHOLE park report (ServerConnection.cs), // The handshake now runs on EVERY outcome... It used to be (CodexHookCommand.cs). None of these parse without the diff, the review thread or the spec.
• Commit subjects run to 108 characters and bodies carry review provenance (Qodo review findings on the first-run browser leg) and spec coordinates.
• PR descriptions run to several thousand characters of per-file walkthrough and per-suite test inventories — detail the diff already carries.
• There is no .github/PULL_REQUEST_TEMPLATE.md, so nothing constrains the description at authoring time.

kcap-server settled this: a Comments section with a two-part test, a Never list and three exceptions; a Commit messages section bounding subject and body and making messages survive squash-merge; and a PR template owning length, headings and its own Never list. The same rules should hold here, so an agent moving between the two repos writes the same way in both.

Repo-specific differences to reconcile rather than copy: this repo is public, requires both a Linear and a GitHub reference on every PR, and bans Linear ids in code comments.