test(trace-adapter): exercise every refusal finalize_trace can reach

[lywinged]

Summary

One test file. Ten cases and six subtests covering every refusal finalize_trace can reach, and no change under src/.

The reason for it: trace_adapter.py has sixteen raise TraceFinalizationError sites, and removing each one on its own leaves ten of them with nothing to notice.

Measured by python -m unittest discover -s tests, which is what CI runs, by exit code rather than by a neighbouring runner's output. Neutralised means the raise statement replaced by pass, located by AST rather than by grep, so the two compound conditions and the two except handlers are counted alongside the twelve plain guards. Sixteen statements, none excluded.

each of the sixteen sites neutralised on its own	CI exit code
before	0 for ten of them
after	1 for all sixteen

The ten were: non-empty chained evidence, the six required configuration fields, duplicate classification_order, no policy decision present, bundle_digest absent, conflicting enforcement modes, no data-flow evidence, taxonomy conflict, the missing optional dependency, and the signing wrapper.

Each of the ten cases trips exactly one of the sixteen sites, traced at the line rather than inferred from the message, and the ten are disjoint from the six that tests/test_trace_adapter.py already holds up. Checked the other way round as well: for all ten, the string the test asserts on is a literal inside the raise at the site the tracer saw, read out of both files by AST rather than by eye. The one exception is hardware signer unavailable, which is the nested cause interpolated at runtime rather than a literal, and that is the point of that case. Nothing here duplicates a case you have, and no case passes on a refusal other than the one it names.

setUpClass is copied from tests/test_trace_adapter.py unchanged, including the Python 3.11 guard and its message, so the two files skip and configure identically.

Behaviour is unchanged and was already correct. Probing _policy_binding directly with six malformed event streams produced the right refusal every time before any of this was written. The gap was coverage, not conduct.

Two things the cases are built to show:

They are reachable rather than defensive. The eight cases that a bad event stream can cause go through EvidenceAccumulator, not by handing the adapter a dict, so a passing case is also evidence the refusal can be triggered from a schema-valid event stream. That matters most for bundle_digest: policy.required is ["engine", "engine_version"], so a conforming producer can emit a policy decision without a digest, and the guard is the only thing between that and a signed record. The case asserts self.validator.validate(event) first, so the file records that the schema permits what the adapter refuses.

Two of the sixteen are different in kind, and the file says so where they sit. No event stream can produce either. The missing optional dependency is reached by blocking the import, which is the only way in, so that case does not go through finalize_trace at all. The signing wrapper does go through finalize_trace, on an accumulator-built snapshot, but it needs a key that satisfies the shape check and then fails when used: object() is refused earlier for having no sign, so a case built that way passes on the wrong refusal. That was the first version of this test and it is called out in the file so nobody rebuilds it.

Contract impact

• No wire-contract change
• Backward-compatible contract change with fixtures
• Breaking contract change with migration notes

No schema, fixture, or src/ file is touched. The diff is one added file under tests/.

Evidence

• python tools/sync_schemas.py (exit 0, working tree unchanged after)
• python tools/check_schemas.py
• python conformance/runner/validate.py
• python -m unittest discover -s tests -v
• python -m build

Also run, since CI runs them: check_versions.py, check_otel_compatibility.py, check_typescript_schemas.py, compileall -q src, and tools/smoke_wheel.py on both the wheel and the sdist. All exit 0.

Not only on my machine. ci.yml unmodified, dispatched on GitHub's runners against the branch and against main, on the fork and before opening this, so the checks on this pull request are the authoritative ones and these are only what I had before asking you to look. Ten jobs across the two runs, every job success and no step failed. On 3.10, 3.11 and 3.13 the build and smoke_wheel steps report skipped, which is the workflow's own if: matrix.python-version == '3.12' doing its job rather than a step that did not run:

• branch test/trace-adapter-refusals at 7d81258: https://github.com/lywinged/agentrust-telemetry/actions/runs/32792667475
• main at 0737e54, for comparison: https://github.com/lywinged/agentrust-telemetry/actions/runs/32789449660

That branch run's own log, on 3.11, prints Ran 108 tests in 2.810s and OK (skipped=2).

That log is where the 3.11 row below comes from. The other three rows are from fresh clones here; on CI those legs are confirmed by exit code rather than by a line I quoted.

Python	before	after
3.10	Ran 89 tests, 4 skipped	Ran 89 tests, 5 skipped
3.11	Ran 98 tests, 2 skipped	Ran 108 tests, 2 skipped
3.12	Ran 98 tests, 2 skipped	Ran 108 tests, 2 skipped
3.13	Ran 98 tests, 2 skipped	Ran 108 tests, 2 skipped

Exit 0 in every cell. The two standing skips are PyYAML not installed and are on main too.

Also measured against the open #30, since it touches the same two files. On that tree the same ten sites survive being neutralised and the same six are caught, so nothing here overlaps it. The two merge in either order without a conflict, and the merged tree is Ran 108 tests at exit 0 with all sixteen sites red under the same sweep.

Privacy and security

• No real prompts, source, credentials, personal data, or production identifiers added
• New fields have capture semantics and privacy classification
• Security-sensitive changes received maintainer review

Every identifier in the file is under example.test or an example.* name, and the configuration block is copied verbatim from tests/test_trace_adapter.py. The run id run-governed-sdlc-001 and the event ids are the synthetic ones already in conformance/fixtures/valid/, varied in the last byte the same way your own adapter tests vary them. No new fields, so the second box does not apply. The third is yours to judge: nothing under src/ changes, but one case constructs a signing key that fails when used, so it is signing-adjacent by subject even though it cannot alter behaviour.

Limits

• On Python 3.10 this file adds nothing. The class skips whole, on the same guard and the same message as tests/test_trace_adapter.py, because agentrust-trace is pinned to python_version >= '3.11'. One of the four matrix legs gains no coverage from this.
• It proves coverage, not correctness. The refusals were already right. These tests pin the current behaviour; they do not argue it is the behaviour you want.
• The assertions match on message substrings, so rewording a refusal turns them red. That is deliberate for a diagnostic that a caller reads, but it is coupling, and it is worth knowing before you accept it.
• Two of the ten prove the refusal fires, not that a producer can cause it. No event stream reaches the missing optional dependency or the signing wrapper. For those two the case is evidence about the code path only.
• Nothing here touches validation.py, context.py, or the TypeScript package. The sweep covers raise TraceFinalizationError in trace_adapter.py and no other module.

One question rather than a change

classification_taxonomy is not in the required dict that _require_finalizable checks, so an empty one is caught later by data-flow taxonomy conflicts with TRACE configuration. It is refused either way. The diagnosis is the less accurate of the two, and I did not want to pin current behaviour in a test without asking whether the omission is deliberate.

[github-actions]

❔ Contributor Check: UNKNOWN

Check	Result
Profile	UNKNOWN
Credential	LOW
Overall	UNKNOWN

Automated check by AgenTrust Contributor Check.

[imran-siddique]

Verified the number this rests on rather than taking it. Parsed src/agentrust_telemetry/trace_adapter.py at pinned 0737e548 with ast: 16 raise TraceFinalizationError sites, at lines 96, 106, 108, 110, 116, 127, 129, 135, 137, 143, 151, 154, 164, 169, 174 and 260. Sixteen, none excluded, exactly as you say.

Counting by AST rather than by grep is the part that makes that number trustworthy. A text match would have missed or miscounted the two compound conditions and the two except handlers, and those four are precisely the ones a hand-written test suite tends to leave uncovered, because they are harder to reach than the twelve plain guards.

Reading the exit code rather than a neighbouring runner's output is the other half. A mutation harness that greps a test runner's summary text is measuring the harness, not the suite, and it fails silently when the output format changes. python -m unittest discover -s tests under the same invocation CI uses, judged by exit status, measures the thing that actually gates the merge.

Ten of sixteen returning exit 0 with their guard removed is a real finding about this file. Those ten guards existed and nothing would have noticed their removal, which is the same as not having them once someone refactors.

Two things I want to credit beyond the method.

Tracing each case to the line it trips rather than inferring from the error message. Two guards raising similar text is common, and message-matching produces a suite that looks disjoint and is not.

Checking disjointness in both directions against tests/test_trace_adapter.py. Adding ten cases that duplicate existing coverage would still have shown sixteen-of-sixteen afterwards while adding nothing, and only the reverse check rules that out.

No change under src/, all eight checks green. Approving and merging.

Separately and not for you to action: this repo was not on my triage list and its PRs have been going unseen. Fixing that at my end.