test(vectors): pin which vector carries which discrimination

[lywinged]

What this changes

Takes the DEFECTS guard from #178 on its own, ahead of the 01-09 reissue, on @imran-siddique's ruling of 10 September: a guard that lands after the thing it guards protects nothing. The reissue keeps its original timing.

test_margins_have_not_thinned stores a count and _margin returns a set, so both are blind to identity. A vector rewritten into a copy of its partner leaves the rule's margin at two, and test_vectors_for_each_rule_are_independent still passes whenever some other declared defect happens to separate the pair. What is lost in that case is the specific property the vector was written to exercise, and nothing recorded which vector carried it.

tests/vector_roles.json records, per fixture, the rules it is load-bearing for and the declared defects it separates. test_no_vector_has_lost_its_role fails by name when a fixture drops one, on the same ratchet terms as the margins file: gaining a role is an ordinary PR, losing one is updated in the same commit with a reason.

Why it is this shape

Of the nine fixtures the reissue moves, exactly one carries a discrimination, which the corpus itself says rather than the argument: measured over all thirty, 04 is the only vector whose issuer key the verifier holds and whose 64-byte signature still fails to verify, so it separates checks_structure_only while 24, at 32 bytes, does not. 14 and 23 carry 64-byte signatures the verifier has no key for, 03 and 17 carry no receipt, and the other twenty-four verify.

What the guard holds is that shape, not the signer. A single-key reissue that collapses 04 to the wrong length, or renames it, fails by name. One that corrupts the signature at fixed length passes every guard, measured by flipping one bit of 04's 64 bytes: the shape survives even though the fixture no longer models a wrong signer. So the second published key that #178 asks for stays right, on the fixture's meaning, and it is a decision the reissue makes on purpose rather than one this test enforces. An earlier version of this paragraph, of the README paragraph, of the defect comment and of the docstring said the shape exists only under a second key; f9841a4 corrects the three files, and this paragraph is the fourth surface.

The remaining six of the nine record as load-bearing with no defect of their own, which is what they are: their partners in 17-30 carry the discrimination. 01 and 02 are absent from the file altogether, because they are the valid vectors and deleting a rule does not change the outcome of a fixture that passes.

The two-key requirement is written down in two places, the signature_or_key_mismatch defect comment and the examples/action-receipts/README.md paragraph that already explains why 01-09 pin an unpublished key.

Measured, not argued

Three mutations, each run against all three guards, and two more added after review:

Mutation	independence	margin ratchet	this guard
04 collapsed to a wrong-length signature	fails	passes	fails, names 04
06 and 25 swap which sits inside the grace window	passes	passes	fails, names 25
04 renamed	passes	passes	fails, names 04
04 with one bit of its 64-byte signature flipped	passes	passes	passes
tests/vector_roles.json deleted, on 62fea76	passes	passes	skipped and rewrote the file; fails by name since 154e7e6

The second touches only unsigned context, so both vectors stay stale and the fixture-correctness suite still passes 35 tests: the swap is a real role exchange rather than a broken fixture. The third is the likeliest thing a reissue actually does to a file. In both of those the existing instruments see nothing, because the count is unchanged and some declared defect still separates the pair. The fourth is the limit of the guard, stated above. The fifth is the reviewer's finding, below.

Each of the three ways to lose a role reports its own reason rather than sharing one closing sentence that fits only one of them.

vector_roles.json is byte-identical across PYTHONHASHSEED values 0, 1, 7, 42, 12345, 31337 and 99999, which matters because _margin returns a set and _roles iterates it.

Suite is 1403 passed and 1 skipped at f9841a4 on 3.11 and 3.12, with ruff check src tests scripts, mypy src/agentrust_trace and tools/check_dashes.py clean. The one skip is test_version's environment check.

Since opened

• 154e7e6: both ratchets in the module fail by name when their baseline file is missing, instead of writing one and skipping. Found by @rajnisht7: with the file deleted in the same change as a degrading reissue, the first run skipped and recorded the degraded state as the baseline, and the second run passed.
• f9841a4: the two-key sentence corrected in the README, the defect comment and the docstring, as above. No code path changes.

Type of change

• Editorial (typo, link fix, clarification: no normative effect)
• Non-breaking spec change (new optional field, new platform profile, informative addition)
• Breaking spec change (requires 14-day comment period and Project Lead sign-off)
• Schema change
• Example addition

None of the five. Tests and one README paragraph. No schema, wire format, signature semantics or verifier behaviour changes, and no fixture is modified.

Spec section

None.

Checklist

• DCO sign-off on all commits (git commit -s)
• CHANGELOG.md updated (for any normative change): not applicable, nothing normative changes
• Breaking changes marked with <!-- CHANGED: #NNN: description --> in spec text: not a breaking change
• Backward compatibility statement included (for breaking changes): not a breaking change

🤖 Generated with Claude Code

[github-actions]

❔ Contributor Check: UNKNOWN

Check	Result
Profile	UNKNOWN
Credential	LOW
Overall	UNKNOWN

Automated check by AgenTrust Contributor Check.

[rajnisht7]

the new test test_no_vector_has_lost_its_role checks the file vector_roles.json to confirm if all vectors are working properly, but if the file is missing, the code creates the file and makes the test green.
the issue is if someone deletes the file, the file will be created automatically during the test, and will make the test green but the deletion of file is never detected, that is guard can turn itself off

[lywinged]

the new test test_no_vector_has_lost_its_role checks the file vector_roles.json to confirm if all vectors are working properly, but if the file is missing, the code creates the file and makes the test green.

the issue is if someone deletes the file, the file will be created automatically during the test, and will make the test green but the deletion of file is never detected, that is guard can turn itself off

Thank you @rajnisht7 🙏, this is exactly the case the guard was supposed to make loud and it made it quiet instead. With tests/vector_roles.json deleted, the test wrote a fresh file from the current tree and reported skipped, and the next run passed. Measured on 62fea76 with 04 collapsed to a 32-byte signature and the file removed in the same change: first run skipped and rewrote the file with 04's discrimination gone, second run passed, nothing red.
Fixed in 154e7e6. A missing file now fails by name, and the message says to restore it from history, or to rebaseline on purpose in the same commit that changes the fixtures, with the reason. The margins ratchet above it had the same shape and takes the same fix, since your point applies to both. With either file deleted its test fails; with both present the module passes 37 and the suite 1403 with 1 skipped on 3.11 and 3.12, the one skip being test_version's environment check, and CI is green on this head.

[imran-siddique]

Approving. This is the guard from my 10 September ruling on #178, landed on its own ahead of the reissue, and it closes a hole the ruling did not name.

What the ruling asked for, and what each piece does. I said the DEFECTS assertion lands first because a guard that arrives after the thing it guards protects nothing, and that the reissue keeps its original timing. Both hold here: tests/vector_roles.json and test_no_vector_has_lost_its_role are the guard, the reissue is untouched.

The part beyond the ruling is the part worth having. test_margins_have_not_thinned stores a count and _margin returns a set, so both are blind to identity. A vector rewritten into a copy of its partner leaves the rule's margin at two and test_vectors_for_each_rule_are_independent still passes whenever some other declared defect separates the pair. The property that vector was written to exercise is gone and nothing records which vector carried it. That is a real gap and I had not seen it.

Ran it rather than read it. vector_roles.json records 28 fixtures. Suite is 37 passed. Then I mutated 04's signature from 64 bytes to 32, making it a structural copy of 24, which is the exact scenario described above. Two tests fail by name:

04-signature-key-mismatch.json: no longer separates ['checks_structure_only'] for
'signature_or_key_mismatch'. It still fails for the rule, so what it lost is the
margin rather than the coverage: it has become a copy of its partner.

The message distinguishes losing the margin from losing the coverage, which is the distinction that makes the failure actionable rather than just red.

Your measurement that 04 is the only vector of the thirty whose issuer key the verifier holds and whose 64-byte signature still fails to verify matches what I measured independently on 9 September, from the fixture bytes: 04 decodes to 64 and 24 to 32, and vector_margins.json records the rule at two.

The ratchet terms are right too: gaining a role is an ordinary PR, losing one is declared in the same commit with a reason. That is what keeps this from becoming a file people edit to make CI green.