docs(spec): record the enforcement mode vocabulary crosswalk in 6.2.1

[zohebk8s]

What

Adds an informative section 6.2.1 recording how the three enforcement mode vocabularies correspond, and points the alignment sentence in 3.2.2 at it.

Why

Closes #344. Section 6.2 requires the cMCP attested enforcement mode and artifacts.policy_bundle.enforcement_mode to match, and 3.2.2 says the two fields align, but neither says which name denotes which state. Three vocabularies are in play:

State	Agent Manifest	cMCP runtime	TRACE claim
Decisions applied	enforce	enforcing	enforce
Decisions surfaced, call proceeds	advisory	advisory	advisory
Decisions recorded, reporting suppressed	audit-only	silent	silent

Until now the only written statement of the correspondence was a mapping table inside a consumer, agentrust-io/cmcp#584, so a second implementer had to rediscover it or guess. On cmcp main at 5b63642, 346 of its 385 files decode as text and the string audit-only appears in exactly one of them.

The section records the correspondence without asserting the states are identical. silent suppresses operational reporting while still recording each decision, so silent satisfies audit-only and audit-only does not imply silent. Keeping that difference explicit is what the review comment asked for; the direction of the implication is my reading and is flagged below.

Spec impact

Section 6.2.1 is new and informative. Section 3.2.2 gains a pointer clause. No RFC 2119 keyword is added, no requirement changes, and no conformance test is affected, so no sponsor is required under GOVERNANCE. CHANGELOG.md updated under Unreleased.

One sentence needs an author's eye. The specification has never defined what its three modes mean; enforce, advisory and audit-only appear only in two schema blocks and the 3.2.2 sentence. To keep the suppressed-reporting difference explicit I had to write what audit-only means, which makes this the first written definition of it in the spec:

audit-only describes evaluating and recording without applying, and says nothing about whether the decision is reported.

That reading follows the argument in the issue, but it is an interpretation rather than a transcription. Happy to reword it.

Test plan

Run on this branch, not assumed from the fact that only documentation changed.

• pytest -v passes: 1481 passed, 6 skipped, 1 xfailed
• mypy src/agent_manifest passes: no issues found in 26 source files
• ruff check src/ tests/ --select E,F,W --ignore E501 passes (the CI invocation, ruff 0.16.6 as pinned in requirements/dev.txt): all checks passed
• New or updated tests cover the change: not applicable, no normative text changed and no code touched
• If spec change: CHANGELOG.md updated

DCO

All commits in this PR are signed off (git commit -s). By submitting this PR I certify the Developer Certificate of Origin.

[github-actions]

🟡 Contributor Check: MEDIUM

Check	Result
Profile	MEDIUM
Credential	LOW
Overall	MEDIUM

Automated check by AgenTrust Contributor Check.

[imran-siddique]

Reviewed df0a026.

This implements the informative mapping requested in #344. I checked the table against cMCP's centralized manifest mapping and TRACE claim emitter. The text preserves the distinction between recording without enforcement and suppressing operational reporting, explains the caller-side translation, and leaves a normative vocabulary extension open. Accepting the stated audit-only interpretation for this informative crosswalk. No executable behavior changes; current hosted tests, lint and type checks pass.