012: Findings from retained repository-check logs
Dated design and evidence record. Original wording, line order and historical source identities are preserved; use the current guides for operation.
Record sections
- 012: Findings from retained repository-check logs
- Purpose and evidence
- Actual extraction and producer provenance
- Empty discovery already fails
- Build admission does not freeze consumed inputs
- Successful publication does not prove failure rollback
- Observed native output differs from undispatched deadline scenarios
012: Findings from retained repository-check logs ================================================ Purpose and evidence -------------------- These concrete examples refine the discussion in note 011. They preserve what the saved tests observed, rather than treating every red result as the same kind of failure. They do not authorize production fixes or waive assertions. All log line references below refer to this retained local file: .work/qualification/repository-check-port-20261014/obvious-repository.log SHA-256: 5f7da0fe5847e52fcc4c2382f19266efe837f7cc3bd30de1cef6690e3fae3570 Its adjacent source.sha256 identifies the tested Make/tools/tests/checker bytes: SHA-256: 29ce2cad1fa74261557ab8a458e7bf5f90f9fd1bfd0410fbe966d92d21740a89 That manifest was checked against the current files during this extraction. The checkout base is ad7048cbb8f0dd1b4d273bfdf47ac58568d34f74 with local changes; HEAD alone does not identify the executed source. Older initial/reviewed logs must not be retroactively bound to this final source manifest. This was a read-only extraction of saved evidence; no test or model was rerun. Actual extraction and producer provenance ---------------------------------------- Lines 324-332: test_actual_extractor_records_exact_producer_bytes errors because the actual extraction receipt lacks the producer_sources required by the imported verifier. This happens before any Git proof: installing Git would not resolve this result. It is a mismatch between the production receipt and the imported provenance contract, pending the policy discussion in note 011. repository_checks/test_fixture_provenance.py:64-74 separates this actual-output check from mutation fixtures that explicitly add producer facts. Those fixture facts exercise the detector; they do not establish that extraction emits them. Empty discovery already fails ----------------------------- Lines 1634-1656: the empty test and test/repository discovery probes return 5 and print NO TESTS RAN; the imported predicate requires exactly 2. This is an exit-status contract mismatch, not a silently successful empty lane. repository_checks/test_discovery_port.py:38 owns the exact assertion. Discussion must distinguish the required nonzero outcome from the specific status policy. Build admission does not freeze consumed inputs ---------------------------------------------- Lines 1514-1535: one probe observes the live checkout as the builder's context; another injects worktree movement and observes changed tools/wheelhouse.py bytes and mode 0755 instead of the originally selected bytes and mode 0644. repository_checks/test_build_inputs_port.py:119-130 owns the mutation probe. This demonstrates the window in the contained launcher/fake-engine fixture. It does not certify the behavior of a live BuildKit execution or real images. Successful publication does not prove failure rollback ----------------------------------------------------- Lines 1618-1623: failure injected at the second output installation leaves the candidate inference.lock instead of its predecessor. The fixture starts with an old lock/report pair and raises before replacing advisories.json, exposing a mixed generation. See repository_checks/test_build_inputs_port.py:299-313 and tools/package_lock.py:89-110 for the injected operation and publication order. The earlier contained lock audit recorded in note 001 establishes input binding and successful exact-byte application; it does not establish this rollback. Observed native output differs from undispatched deadline scenarios ------------------------------------------------------------------ Lines 1730-1735: an injected failing engine's synthetic private output marker appears in the launcher's captured output. This is a diagnostic-boundary probe, not evidence that real credentials were exposed. Lines 1749-1754: an injected successful engine emits 8,388,609 bytes, exceeding the imported 8,388,608-byte bound by one byte. See repository_checks/test_operator_inputs_port.py:380-398 for both probes. These probes executed. By contrast, the EOF/descendant probes at lines 1721-1727 and 1738-1746 fail before dispatch because no reviewed deadline adapter exists. An outer fixture timeout cannot substitute for an owned operator deadline. Preserve that distinction when discussing or provisioning the remaining checks.