feat(verify): record a caller-supplied appraiser's per-layer report on runtime.measurement (#279, #431)

[revenue7-eng]

Implements the platform-measurement row of #279, in the shape agreed in #431.

What it does

verify_record has runtime.platform and one digest, runtime.measurement. For a measured-boot platform that digest is a composite over many layers, and a match says nothing about which layers were measured, which were appraised, or whether the evidence describes one boot. verify_record never sees the quote, log or reference values, so it cannot decide that itself.

Following citation.py, this adds agentrust_trace.platform_measurement and a platform_appraiser argument to verify_record. The appraiser is caller-supplied, called last with a copy of runtime, and returns the measurement it appraised and a status per layer. The result's new platform_measurement field carries the report:

• appraised, with a LayerCheck per layer: established, or not_established with layer_not_measured, measured_not_appraised or evidence_spans_multiple_boots;
• appraisal_rejected with appraiser_raised, appraiser_returned_invalid or measurement_mismatch (a report about another measurement is never attached);
• not_attempted with no_appraiser when none is supplied.

The module asserts nothing about the platform, derives no layer outcome, upgrades nothing, and produces no appraisal.status. No outcome moves revocation, the thumbprint, citations or whether verify_record raises. The names are not accepted normative text. No schema or wire-format change.

The two conditions from #431

1. Each cause has a fixture that produces it and a twin that differs only in that condition and does not. Sixteen signed vectors in examples/platform-measurement/: seven causes, each with a twin_of vector carrying the same signed record and the same context except the appraisal table. test_every_cause_has_a_twin_that_differs_only_in_its_condition fails if a cause loses its twin or a twin differs in anything else. Every vector without a source carries "synthetic": true; evidence_spans_multiple_boots exists only in the synthetic pair 013/014, because the board behind the real vectors resets its TPM in SPL and does not produce it.
2. not_attempted stays distinct from a pass wherever the result is summarised. Nothing in this repository summarises a VerificationResult today; the rule is stated in the module docstring and held by test_P4b_not_attempted_is_never_a_pass, so a future summary cannot collapse it.

Evidence

Vectors 015 and 016 carry measurements from published quotes on a physical board with a discrete TPM (Rock 5A, Infineon SLB9670), registered AK:

• 015: the v2.1.0-rc13 reference composite over PCR 0 to 9, where PCR 2, 3, 5 and 7 hold a single separator (layer_not_measured). Evidence and checker: https://github.com/revenue7-eng/tactiq-os/releases/tag/v2.1.0-rc13
• 016: a cold-start quote over PCR 0 to 12 from a development image, where four IMA entries in PCR 11 and 12 replay to the quoted values and the policy that leaves them unappraised is not in the log (measured_not_appraised). Evidence and offline checker: https://github.com/revenue7-eng/tactiq-os/blob/main/measurements/tpm-quotes-dev-20261002.md

Nothing in this repository reads that evidence; the vectors carry the appraiser's report as the input.

Checks

pytest (3406 passed, 48 skipped), ruff check src tests scripts, tools/check_dashes.py. The generator reproduces the set byte-for-byte and is picked up by test_generators_reproduce_fixtures; the set is registered in test_adequacy_all_sets and the new function and parameter in test_public_functions_raise_what_they_document.

[imran-siddique]

@revenue7-eng CI's Type check step fails on test (3.11), and test (3.12) was cancelled behind it:

src/agentrust_trace/platform_measurement.py:149: error: Argument 1 to "deepcopy" has incompatible type "Any | None"; expected "dict[str, Any]"

mypy does not narrow runtime from the measurement is None test. Changing line 145 to if not isinstance(runtime, dict) or measurement is None: clears it with the same behaviour; I checked locally that mypy src/agentrust_trace is clean and the 51 tests in test_platform_measurement.py pass. CI runs mypy alongside pytest and ruff. The review follows a green run.

[revenue7-eng]

The 3.11 job failed at the mypy step: runtime was not narrowed to a dict before copy.deepcopy. Fixed in the latest commit; ruff, check_dashes, mypy (2.3.1, as pinned) and pytest pass locally on it.

[revenue7-eng]

Rebased onto current main (after #462), per @lywinged's note there; the gate now reports "awaiting 1 independent maintainer approval" for 77c6562 instead of the HttpError. Both commits keep their DCO sign-off.

[imran-siddique]

@revenue7-eng CI is green on 77c6562 and both #431 conditions hold. One gap: an appraiser returning {"measurement": <the record's>, "layers": {}} comes back appraised with zero layers, so a report that established nothing reads as an appraisal. Refuse an empty layers as appraiser_returned_invalid with member="layers", and add it beside the "extra": 1 case in test_platform_measurement.py. This merges once that is green.

[revenue7-eng]

Done in the latest commit: an empty layers is now appraiser_returned_invalid with member="layers", and the case sits beside "extra": 1 in test_platform_measurement.py.

[imran-siddique]

The empty layers case is refused as appraiser_returned_invalid with the regression beside "extra": 1, which was the one remaining condition.