[drevendev]

A v1 carrier discarded merge_parents() conflict diagnostics, and later reconciliation could erase a retained parent-kind contradiction when current attempts were complete. With the same parent id represented as Actions and Check Suite, input order could turn an unresolved history into a favorable PASS-only result. This repair carries the conflict into unresolved and preserves it through subsequent reconciliation.

Target: B1
Target: B3

Related to #83. Bounded parent-kind propagation/persistence repair; keep #83 open. Separate legacy status/coverage work remains under #87.

Base and dependency

• Target: release/0.6.0, per CONTRIBUTING.md.
• Depends on fix(history): preserve malformed legacy carrier values (partial #83) #89 at 1a735f451f9fe6dae7948a014b660249501a1ffc (legacy conversion repair); that PR depends on fix(history): reject malformed carried arrays (partial #83) #88 and 0.6: independently verified Git history and executable assurance #81.
• Independently reviewed and tested head: 51e58b8a4f1504f8f953fbc47db6157d157051e6.
• Parent-kind-only diff: one commit, four files, +224/-6.
• At creation the release branch remains 22a2894057d9b1bc003efe4ef932ed4de601e26b. The full PR comparison includes unmerged prerequisites: 24 commits / 70 files. Review the incremental diff above for this repair; no independent acceptance of the inherited release scope is claimed. Draft until dependencies are reconciled; no merge requested.

Resulting behavior

• Replayable v1 _carried_records() retains the conflict list returned by merge_parents() in the revision's unresolved evidence.
• Complete current evidence may repair an incompatible-lineage gap, but cannot clear PARENT_KIND_CONFLICT:<parent_id>.
• Both input orders retain the contradiction, and subsequent v2 carriers keep it. A current PASS cannot manufacture PASS-only history from the conflict.
• An observed VERIFY_FAIL from either kind remains absorbing.
• Valid same-kind/distinct-parent unions, v2 duplicate-group rejection, accepted lineage repair, expiry and proven departure from the complete branch inventory retain their behavior.
• No history-lineage, Vital policy, threshold or runtime-dependency change.

Added 37 regression instances and specification/conformance mapping: 6 v1 order/failure cases; 12 repeated v1/v2 reconciliation cases; 4 valid union controls; 2 v2 duplicate-group refusals; 2 lineage-only/mixed-gap controls; one multiple-marker case; 2 expiry/departure controls; and 8 four-bundle legacy/current-origin store/replay cases. Implementation-local case labels do not replace research identifiers.

Executed checks at the exact fixing head

Windows 11 build 26200, Python 3.14.7, pytest 9.1.1, 2026-10-24. Real source import confirmed from the fixing checkout; PYTHONDONTWRITEBYTECODE=1, explicit source PYTHONPATH.

Check	Result
New test module against the actual unpatched base package	18 failed / 19 passed, expected red
python -B -m pytest -p no:cacheprovider tests/test_parent_kind_conflicts.py tests/test_legacy_carrier_admission.py tests/test_carried_array_admission.py tests/test_revision_history.py tests/test_ci_history_admission.py --basetemp=<dedicated scratch>	153 passed
Single full native run outside sandbox: python -B -m pytest -p no:cacheprovider --basetemp=<dedicated scratch> --junitxml=<external evidence path>	2645 passed in 87.89s; 0 failures/errors/skips; exit 0
python -B -m devostasis vectors --quiet	148 passed
python -B scripts/check_conformance_surface.py	63 passed
Offline core/code-v1/code-v2/packet/work verify and replay	6 checks passed
Eight four-bundle conflict chains	Every bundle verifies; saved-observation replay preserves exact bundle ID and member bytes
git diff --check <base>..HEAD, worktree preservation	PASS; fixing and prior checkout clean

Independent review

A separate reviewer inspected the exact four-file diff, actual parser/reconciliation/evaluator paths, fixtures and relevant contract.

No actionable findings in this bounded diff. Independently executed:

• python -B -m pytest -p no:cacheprovider tests/test_parent_kind_conflicts.py tests/test_revision_history.py tests/conformance/test_integrity.py: 88 passed;
• python -B -m pytest -p no:cacheprovider tests/conformance/test_vectors.py -k HIST: 14 passed, 138 deselected;
• exact-head new tests against the actual base: expected 18 failed / 19 passed;
• 24 permutations of three parent kinds and failure location, through 96 reconciliation stages including acquisition outage, partial empty inventory and later complete evidence under a different parent;
• 4 accepted lineage-repair controls;
• real source import, clean HEAD and diff checks.

All probes use actual package imports and fixture-only construction, not copied core functions. The primary agent inspected the retained probe script and outputs. The independent reviewer did not repeat the primary 2645-pass full suite. Separate-agent execution on the same owner/host is not separate credential identity or release approval.

Bounded patch SHA-256: 9352b41728f9e721b9b552afb6264df8d15f2af7ad837ac92a6fb71c3a36c10f. The committed tests and commands above are the public reproducible corpus; private task records, local paths and transcripts are not published.

Remaining #83 / #87 scope

The three stacked repairs (#88, #89 and this PR) do not by themselves finish issue/release acceptance:

• integrate the dependency chain and propagate the admission/conflict fixes consistently through the other prepared release candidates;
• verify the resulting integrated heads and complete their independent release/compatibility acceptance gates;
• Research: legacy Integrity predecessor UNKNOWN/UNAVAILABLE becomes AVAILABLE empty carrier #87: correct non-value-bearing legacy UNKNOWN/UNAVAILABLE/FORBIDDEN/ERROR status handling without manufacturing known-empty AVAILABLE history;
• Research: legacy Integrity predecessor UNKNOWN/UNAVAILABLE becomes AVAILABLE empty carrier #87: require valid complete-acquisition proof for a genuinely empty series;
• Research: legacy Integrity predecessor UNKNOWN/UNAVAILABLE becomes AVAILABLE empty carrier #87: preserve inventory-level uncertainty and observed failures for valid PARTIAL history across successor bundles, with the required lineage/compatibility decision where the carrier shape changes.

Linux, Python 3.12/3.13 and hosted CI were not run for this repair. These checks do not accept the whole release or the inherited #81 scope. No automatic closure of #83/#87, merge or release is requested.

[andy-zen-dev]

Independent bounded review of #90 at exact head 51e58b8a4f1504f8f953fbc47db6157d157051e6, relative to its immediate predecessor 1a735f451f9fe6dae7948a014b660249501a1ffc (one commit, four files, +224/-6), not a sign-off on the accumulated 24-commit release diff.

I examined the actual commit patch, surrounding _carried_records(), merge_parents(), reconcile(), Integrity's fail-closed evaluation path, the 37-instance test_parent_kind_conflicts.py corpus, docs/spec/integrity-ci.md, and the implementation-local conformance mapping. No actionable merge-blocking defect found within this bounded repair.

• v1 replay now preserves the PARENT_KIND_CONFLICT:<parent_id> diagnostics returned by merge_parents(), rather than discarding them when constructing a v2 union.
• Complete current parent evidence may clear a repairable lineage gap but retains structural parent-kind contradictions, preventing a contradictory history from being upgraded to PASS_ONLY_OBSERVED. VERIFY_FAIL remains absorbing from either input order.
• Regression coverage includes both parent orders, failure placement, repeated v1/v2 reconciliations, legitimate parent unions, duplicate-group rejection, lineage-only repair, expiry/branch departure, and four-bundle stored/replayed histories; specification and executable mapping are updated without changing Vital policy or lineage.
• GitHub Actions run https://github.com/drevendev/Devostasis/actions/runs/40098455528 is SUCCESS at this exact commit; all three hosted jobs (Python 3.12/3.13/3.14) and their reported steps succeeded. This is hosted verification, not an independently executed local pytest or release acceptance.

Scope/merge gate: #90 is a draft stacked on unmerged #89 -> #88 -> #81 and targets release/0.6.0. The full PR diff includes their inherited changes; those changes are not approved by this review. Do not merge this draft or close #83 merely on this bounded finding. Rebase/recheck exact integrated heads after prerequisite landing and resolve remaining #87 status/inventory uncertainty separately. Local pytest and independent historical replay: NOT_RUN in this review.

[andy-zen-dev]

2026-10-24 — integration-gate drift after #91 (exact-head audit, not a new #83 code finding)

Current release branch: release/0.6.0 at 5998723411bd2721971a7385ca2d8c8af8df055f, the verified merge commit for #91 (merged 2026-10-24 14:27:17 UTC). That merge added scripts/wheel_smoke.py, docs/wheel-verification.md, and the wheel-smoke job to .github/workflows/ci.yml, delivering the independently installed-wheel check from #84.

This PR's exact head: 51e58b8a4f1504f8f953fbc47db6157d157051e6. Comparing current release head to this commit reports diverged, 24 ahead / 2 behind, merge base 22a2894057d9b1bc003efe4ef932ed4de601e26b. The checked-in .github/workflows/ci.yml at this PR head (blob 42763487e82b2e634cf13b81fe6b65d81935ea1d) does not contain the new wheel-smoke job; the current release branch's workflow (blob ddb3739494409b94d5667eb16b73537ff58ad541) does.

CI evidence boundary: run 40098455528 is SUCCESS for this PR's source SHA, but its inspected job list contains only Tests (Python 3.12/3.13/3.14). It does not qualify the newly required installed-wheel gate for the combined #81 → #88 → #89 → #90 changes, nor a merge candidate containing #91. This is not a claim that wheel installation is broken.

Required before claiming the 0.6.0 integrated/release head green: reconcile the landed #91 branch changes with the stacked feature/release chain in dependency order, retain the newly merged wheel gate, and obtain successful exact integrated-head CI with the installed-wheel job and the Python test matrix, followed by historical/conformance replay and independent release acceptance. Do not treat the three green source-head jobs as a pass for a different merged tree. Keep #90 draft and #83 open until their independent gates are satisfied. Existing #81 blockers (#38/#82/#86 and the remaining #83/#87 boundary) are unchanged by this observation.

Method/limits: read the current Git ref and #91 merge, both exact workflow blobs, PR refs, commit comparison, run 40098455528 and its three hosted jobs. No new source checkout pytest, installed-wheel execution, approval, branch update, merge or release was performed by this audit.

---

2026-10-27 — updated release-base reconciliation

A fresh GitHub ref/compare/CI inspection supersedes the 2026-10-24 release-base snapshot above (the older audit is retained for chronology):

• Current release/0.6.0 is 6d935ee111e2efab7a5f7d7c8f8d050af8d35567. Since the common base 22a2894057d9b1bc003efe4ef932ed4de601e26b, upstream landed four commits: ci: verify installed wheel outside the source checkout (#84) #91's installed-wheel gate and fix(ci): verify installed wheel from the PR merge candidate (#92) #93's fix for CI: wheel-smoke checks stale PR head instead of the tested merge candidate #92. Issue #92 is now closed (2026-10-25); its explicit PR-head checkout override is absent. The live release workflow blob 2fcc7c9843e2c9ca1a66203063845f9f6b7bac45 uses normal checkout plus git rev-parse HEAD == GITHUB_SHA assertions in both wheel and test jobs.
• This draft PR remains at source head 51e58b8a4f1504f8f953fbc47db6157d157051e6 with no new source revision; current merge comparison reports diverged: 24 ahead / 4 behind, merge base 22a2894057d9b1bc003efe4ef932ed4de601e26b. Its source tree lacks scripts/wheel_smoke.py (verified 404) and its checked-in CI workflow has only the three Python test jobs. These are source-head facts; GitHub may combine the fixed base workflow when it builds a PR merge candidate, but that candidate has not been verified in this review.
• Run 40098455528 remains success only for the unchanged source head and has exactly three successful Python test jobs; it does not constitute installed-wheel or exact integrated-head evidence. The CI: wheel-smoke checks stale PR head instead of the tested merge candidate #92 fix on the release branch resolves the previously identified checkout-selection defect; do not reopen CI: wheel-smoke checks stale PR head instead of the tested merge candidate #92 as if it were unfixed.

Remaining gate: integrate/reconcile #81 → #88 → #89 → #90 against the new release head while keeping #93's checkout assertions and wheel gate; obtain successful Python 3.12/3.13/3.14 + wheel-smoke jobs at the actual integrated candidate, plus historical/conformance replay and independent acceptance before leaving draft/releasing. #38/#82/#83/#86/#87 and #96 are separate open correctness/contract boundaries; this update does not assert their resolution. Full checkout and local tests: NOT_RUN (git fetch DNS unavailable); no source edits, re-review approval, branch updates, merges or tag were performed here.