Keep failed structured-component evidence schema-valid

Product outcome

A failed CI job must still publish truthful, schema-valid evidence for every selected component it actually ran. A primary quality-contract failure must remain the job/report verdict, while later structured-content evidence recording must neither inherit a false component result nor crash with a secondary validation traceback.

Observed evidence

• Push run 33295699282, quality job 99214937712 first failed in terminology-check because frozen digests differed for podcasts.json, wiki.json, and wiki_search.json.
• The evidence step then recorded quality-evidence.json, regenerated a valid 69,982-byte content-invariants.json, passed the overall job.status=failure to that component, and raised ci.evidence.EvidenceError: machine output counts are invalid.
• The preserved artifact has quality-evidence.json and content_invariants-result.json, but no content_invariants-evidence.json.
• Current machine_output_claim assigns the failed non-test component the content-invariants-v1 format with only the five base zero-count fields. _validate_machine_output_claim requires that format to also carry records and structured_files, so the recorder rejects its own generated claim.

Lifecycle disposition

This is a new issue, not a reopening of #195. #195 specifically fixed failed test-log parsing and the zero-count fallback for unparseable test output; its accepted regressions remain present. This failure is a structured content-invariants-v1 claim and a multi-component job-status attribution problem. It is also distinct from #202, which covers missing or irregular artifact collection; this run retained a regular structured artifact and failed later on its count shape.

Normative authority

• Development process, especially truthful failure attribution and evidence retention.
• Risk-based verification and evidence reuse, especially component-specific normalized results/counts, digest-bound machine output, exhaustive buckets, and aggregate-report creation on failure.
• Playwright timeout contract, which states the established cross-component principle: a non-success result with no parseable counts retains digest-bound failure evidence and does not block report creation.

Scope

• Make the quality job record each selected component from that component's actual execution result, rather than copying one earlier sibling failure to work that subsequently succeeded.
• Make every non-success content_invariants machine-output claim valid for the declared content-invariants-v1 schema, including the no-parseable-counts fallback.
• Continue recording remaining selected components after one component fails, then emit the final aggregate failure report with the primary failure preserved and each component classified exactly once.
• Add focused CI-contract tests for the exact observed sequence and for a genuine failed/truncated content-invariants execution.

Non-goals

• Do not change the three frozen projection digests or the approved public headings Events and Community knowledgebase; their source/projection correction remains separate work.
• Do not weaken successful structured-content proof: a success still requires the complete, positive, digest-bound invariant payload.
• Do not make any failure evidence reusable or change expiry, trust, input, environment, bucket, or verdict rules.
• No product-page, content, route, template, or visual change.
• No workflow rerun as implementation evidence until this issue has passed engineer, independent tester, and PM gates.

Acceptance criteria

• Reproduce the run-33295699282 sequence in a deterministic test: the quality component fails first; a selected content-invariants execution then succeeds; recording completes without traceback; quality is failure, content invariants is success with its real records/structured_files counts and bound artifact.
• A genuinely failed, cancelled, timed-out, malformed, or truncated content-invariants execution emits a schema-valid non-success envelope with format-compatible zero derived counts when no trustworthy counts exist; recording and final report creation do not raise.
• The final report preserves verdict failure, classifies every planned component exactly once, and never converts the primary failure or any absent proof into success, reuse, or an undocumented skip.
• Success-path strictness remains fail-closed: missing, malformed, empty, digest-mismatched, or incomplete content-invariants-v1 success output is rejected.
• Regression coverage proves #195 remains intact for failed test logs and #202 remains intact for missing/irregular failure artifacts.
• make test-ci, make lint, make format-check, and make typecheck pass; the versioned verification plan/report records all component dispositions.
• Screenshots are not_applicable: this is internal CI/process behavior with no product render impact.

Dependencies

None. The test can use synthetic failure output and structured invariant artifacts; it does not depend on repairing the current projection-digest mismatch.

Related: #195, #202, #253, #261.

Start by locating machine_output_claim and _validate_machine_output_claim, then trace the quality-job recording flow for selected components. Build deterministic tests for the observed sibling-failure sequence and failed or truncated content-invariants output. Done means schema-valid non-success evidence, truthful per-component results, preserved primary failure, exactly-once classification, and passing make test-ci, make lint, make format-check, and make typecheck.