docs(types): refresh import-architecture references (closes #807)

[bokelley]

Summary

• Refresh CLAUDE.md's Import Architecture for Generated Types section against the actual tests/test_import_layering.py allowlist (drops nonexistent stable.py; adds capabilities.py, _forward_compat.py, _generated.py).
• Drop the Handling Missing Schema Types example — FormatId = str / PackageRequest = dict[str, Any] were both replaced with real generated types in earlier PRs; the section was misleading.
• Add a Codegen variant numbering is unstable callout so the next person who runs scripts/generate_types.py doesn't mistake the ~600-file ClassNameN → ClassNameM renumber churn for a real schema delta.
• Carry the stable.py doc fix through three other files that referenced it: src/adcp/types/capabilities.py docstring, src/adcp/decisioning/time_budget.py docstring, .github/ai-review/expert-adcp-reviewer.md layering-breach rule.

Why these are the only changes

#807's stated scope was (a) re-sync schemas, (b) regenerate types and inspect diff, (c) drop obsolete handrolled aliases, (d) update CLAUDE.md.

• (a) is already done — main is at 3.1.0-beta.3 via chore(schemas): resync 3.1.0-beta.3 cache + fix broken manifest patch (refs #807) #821.
• (b) was run as part of this work. The diff is 600+ files of pure variant-renumbering churn (Idempotency3 → Idempotency1, Assets162 → Assets161, etc.) — datamodel-code-generator numbers anonymous variant classes by traversal order, and even a latest-keyed re-run on the same pinned schema produces shifts. Committing the regen would break aliases.py's semantic re-exports of the numbered classes for zero spec value, so it's discarded. The new CLAUDE.md callout documents the trap.
• (c) the FormatId / PackageRequest stubs called out in the original CLAUDE.md no longer exist in code — earlier PRs replaced them. aliases.py does still re-export FormatId = FormatReferenceStructuredObject (a real generated type), which is the load-bearing form and stays.
• (d) is what this PR actually does.

Net: pure documentation hygiene. No public API or generated-type changes.

Test plan

• ruff check src/ — all checks passed
• mypy src/adcp/ — no issues in 892 source files
• pytest tests/ — 5010 passed, 40 skipped, 1 xfailed
• tests/test_import_layering.py specifically green (50 tests)

Closes #807.

[aao-ipr-bot]

LGTM. Pure docs hygiene catching CLAUDE.md up with reality — stable.py was already gone, the FormatId = str / PackageRequest = dict[str, Any] stubs were already replaced, and the actual allowlist is enforced by tests/test_import_layering.py:33-48. The doc was lying; this fixes it.

Things I checked

• Allowlist in tests/test_import_layering.py:33-48 matches the new CLAUDE.md list exactly: aliases.py, _ergonomic.py, _generated.py, __init__.py, capabilities.py, _forward_compat.py. No stable.py. Claim is sound.
• src/adcp/types/stable.py does not exist on disk — the old doc was pointing at a phantom.
• grep for FormatId = str and PackageRequest = dict[str, Any] in src/ returns nothing — confirmed the stubs the old section described are gone.
• aliases.py still re-exports FormatId = FormatReferenceStructuredObject (a real generated type), which is the load-bearing form the PR body calls out. Not regressed.
• src/adcp/types/capabilities.py:13-15 docstring whitelist now agrees with CLAUDE.md and the test.
• src/adcp/decisioning/time_budget.py:144 docstring now points at the test as the source of truth instead of naming an outdated subset. Right move — the doc no longer has to be updated when the allowlist grows.
• .github/ai-review/expert-adcp-reviewer.md:82 MUST FIX rule 6 now cites the test-enforced allowlist. The reviewer prompt was the third site carrying the stable.py ghost; good catch.
• Codegen-variant-numbering callout in CLAUDE.md is the right warning to bake in. datamodel-code-generator's traversal-order numbering producing 600+ files of ClassNameN → ClassNameM churn on a clean re-run is exactly the kind of trap that costs a day if you don't know about it.
• Public API surface: untouched. __init__.py not in the diff.
• Generated layer: untouched. generated_poc/** and _generated.py not in the diff.
• Test plan: all four checkboxes checked. pytest tests/ reports 5010 passed, including the 50 tests in test_import_layering.py.

Minor nits (non-blocking)

1. Commit message says "No code changes" but time_budget.py reformats two logger.warning strings from explicit multi-line continuation to single-line implicit-concatenation (time_budget.py:120, 130). The runtime string is identical — Python concatenates adjacent literals at compile time — so no behavior change, but the new form reads worse on a 120-col terminal and isn't part of the stated docs-hygiene scope. Probably an autoformat side effect that snuck in. Notable, not blocking.

LGTM.