docs: slim CLAUDE.md and split long-form rationale into .claude

[uipreliga]

Why

CLAUDE.md is auto-loaded into every request, so a 12,524-word changelog in it is a tax on every task. It carried a single 21,263-character bullet and a 14,025-character lint-rule catalogue.

The prose quality was never the problem — the location was.

What

Split by purpose rather than trimming uniformly:

File	Role
CLAUDE.md (12,524 → ~2,600 words)	Working reference: layout, criteria table, evaluation flow, dev commands, the lint rules that constrain routine edits, doc-SSOT rules
.claude/architecture-notes.md (new)	Long-form design rationale, moved verbatim under headings. Git-tracked but not auto-loaded, so the knowledge survives without the per-request cost

The CE rule catalogue is now a pointer. Every rule already carries its full rationale in its own module docstring, and a prose copy of a registry with no sensor decays silently — which is exactly what test_claude_md_names_every_registered_model already exists to catch.

Documentation links

Every one of the 15 criteria table rows now deep-links to its section in the Task Definition Guide, plus pointers to the report schema, dataset, dialog-mode, docker-isolation, CI-gate, plugin and per-agent docs.

• All 52 links and 25 anchors resolved against the tree (no link checker exists in the suite, so this was verified by parsing).
• Every anchored heading is punctuation-free, so GitHub, Starlight and python-markdown agree. Two targets with punctuation (stop_early, API Routing & Benchmarking) are deliberately linked at file level.
• No docs index table was added. CE028 does not scan CLAUDE.md, so a copy here would be an unguarded fourth index that decays.

Constraints preserved deliberately

• The CE030 model list stays on one physical line — that test is line-based.
• All six /coder-eval:* skill names (SKILL_DOC_SURFACES).
• The "six skills" count phrasing.
• setting_sources: [] guidance, which a line-by-line diff caught as landing in neither file on the first pass.

Verification

• make lint — 635 passed
• make test — 5,678 passed, 2 skipped (both environmental: needs a real ANTHROPIC_API_KEY; Windows-only)

Not in this PR

A proposal to replace the Design Principles values (DRY/YAGNI/KISS) with seven checkable "Never ship" rules, three PreToolUse/Stop hooks, and an Opus-in-tests lint rule. Discussed but not applied.

🤖 Generated with Claude Code

https://claude.ai/code/session_015qJBGVKjSeTucaw5h36Pt6

[github-actions]

Claude finished @uipreliga's task in 1m 36s —— View job

---

Review in Progress

Todo List:

• Read .github/code_review.md for review guidelines
• Read current CLAUDE.md for project conventions
• Run git diff origin/main...HEAD to see full changes
• Read both changed files completely
• Verify critical information preservation
• Check link integrity
• Perform cross-file consistency checks
• Complete "What's Missing" analysis
• Submit comprehensive review