[drevendev]

Legacy carried_observation() replaced a malformed value-bearing predecessor series with [] before admission could inspect it. An AVAILABLE or PARTIAL predecessor containing null, false, 0, "", {} or a missing value could therefore appear to carry known-empty history. This repair preserves that malformed value so the existing parser/reconciliation/evaluator returns REVISION_HISTORY_CARRY_MALFORMED, UNKNOWN and no band.

Target: B1
Target: B3

Related to #83; partial repair only, keep #83 open. Related research/status work: #87.

Base and dependency

• Target: release/0.6.0, per CONTRIBUTING.md.
• Depends on fix(history): reject malformed carried arrays (partial #83) #88 at 9cf0c0cf678aeef620b5f6d6a1662533894c10b2, which supplies the three carried-array reader fixes and itself depends on 0.6: independently verified Git history and executable assurance #81.
• This repair's independently reviewed and fully tested head: 1a735f451f9fe6dae7948a014b660249501a1ffc.
• Legacy-only diff: one commit, four files, +184/-4.
• The release branch is currently 22a2894057d9b1bc003efe4ef932ed4de601e26b. GitHub's overall comparison consequently includes the unmerged prerequisites: 23 commits / 69 files. Review the linked incremental diff for this repair; it does not independently accept the inherited 0.6: independently verified Git history and executable assurance #81/fix(history): reject malformed carried arrays (partial #83) #88 scope. Draft until dependencies are reconciled; no merge requested.

Bounded behavior

• Remove the list-type coercion from the legacy AVAILABLE/PARTIAL value selection.
• Apply window filtering and replay projection only to actual lists; preserve other values for existing carried-record admission.
• Keep list projection, window filtering and valid observed failures unchanged.
• Add 30 regression instances: 10 invalid-value preservation cases, 10 public no-favorable-result cases, 2 actual-empty shape controls, 4 valid pass/fail projection/window controls, 2 missing-value cases, and 2 four-generation real-store failure/replay cases.
• Update Integrity specification and implementation-local conformance mapping.
• No Vital threshold, policy, history-lineage or runtime-dependency change.

The empty-list controls cover existing conversion shape compatibility; they do not establish complete acquisition or resolve valid PARTIAL inventory uncertainty.

Executed tests at this exact head

Windows 11 build 26200, Python 3.14.7, pytest 9.1.1, 2026-10-24. Source import confirmed from the fixing checkout, with PYTHONDONTWRITEBYTECODE=1 and PYTHONPATH=src.

Check	Result
New regressions on the real unpatched base package	22 failed / 8 passed, expected red
python -B -m pytest -p no:cacheprovider 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>	116 passed
Single full native Windows run outside sandbox: python -B -m pytest -p no:cacheprovider --basetemp=<dedicated scratch> --junitxml=<external evidence path>	2608 passed in 82.97s; 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, v2 packet and work verify/replay	6 checks passed
Four-generation v1 → v2 → v2 → v2 store chains	All bundles verify; retained failure and exact byte/ID replay
git diff --check <base>..HEAD, final worktree check	PASS; clean

An initial test fixture attempted replay after publishing the successor and correctly triggered predecessor validation. Its order was corrected before changing production code; the final unpatched red control above and all subsequent checks used the corrected fixture. Tests were not weakened.

Independent review

A separate reviewer inspected the exact four-file base-to-head diff and actual conversion, parser, reconciliation and public evaluator paths.

No actionable findings within this bounded scope. Independently executed:

• 93 focused tests passed in 2.28s;
• real-base red reproduction: 22 failed / 8 passed;
• 16 admissible non-null malformed AVAILABLE/PARTIAL store cases: predecessor and resulting bundle verify, UNKNOWN/no band with CARRY_MALFORMED, and exact saved-observation replay;
• 40 differential carrier/evaluator outputs identical for behavior that should remain unchanged: valid/empty lists, passing/failing/incomplete histories, window boundaries, malformed records and non-value-bearing statuses;
• diff whitespace check and source/evidence/patch verification.

The reviewer did not repeat the full suite; the 2608-pass full run above is a separate primary execution. AVAILABLE/null is rejected by the current Observation constructor before store creation, so null/missing values were checked through real conversion/evaluation tests; the store probe covers admissible non-null values. The review used a separate agent in the same owner/host environment, not separate credentials or release approval.

Public-safe evidence digests:

• bounded patch SHA-256: fc673a94e6aed4a8f3bb9e647dbb7036e6291e5b322aaeecaf2fa3cf7b3116c5;
• committed new test source and named commands provide the reproducible regression corpus.
  Private task state, absolute local paths, transcripts and raw private evidence manifests are not published.

Remaining #83 scope and limits

This PR does not close #83 or #87:

• v1 parent-kind conflict propagation when the same parent_id has different kind;
• preservation of structural conflicts through later complete-current reconciliation (conflict stickiness); propagation and reconciliation must be addressed together;
• non-value-bearing legacy status handling (UNKNOWN/UNAVAILABLE/FORBIDDEN/ERROR) without promoting missing evidence to an empty AVAILABLE carrier;
• complete-empty acquisition proof and inventory-level uncertainty for valid PARTIAL history across successor bundles, under Research: legacy Integrity predecessor UNKNOWN/UNAVAILABLE becomes AVAILABLE empty carrier #87's compatibility requirements;
• propagation through the other prepared release candidates.

Linux, Python 3.12/3.13 and hosted CI were not run for this repair. The broader release and prerequisite acceptance gates remain separate. No automatic issue closure, merge or release is requested.

[andy-zen-dev]

Independent exact-head review of the bounded legacy-conversion repair in #89: 9cf0c0cf678aeef620b5f6d6a1662533894c10b2..1a735f451f9fe6dae7948a014b660249501a1ffc (one commit; four files; +184/-4). This does not review the unmerged prerequisite stack included in the GitHub PR's overall diff.

I inspected the complete incremental patch in src/devostasis/revision_history.py, tests/test_legacy_carrier_admission.py, docs/spec/integrity-ci.md and docs/spec/conformance.md, with the actual surrounding _in_window, _replay_source, _carried_records, reconcile, integrity.evaluate and repository contribution rules.

No actionable merge-blocking finding in this bounded change. It correctly stops replacing malformed value-bearing 0.1.x AVAILABLE/PARTIAL series with []: non-lists now survive projection and are rejected by carried-record admission; actual lists keep the existing window/projection behavior. The 30 parameterized regression instances cover five malformed shapes across both statuses, downstream UNKNOWN/no-band behavior despite a passing current control, real empty-list compatibility, valid pass/fail preservation and four-generation bundle verification/replay. Documentation explicitly scopes the fix as admission, not accepted status/coverage semantics.

Exact-head hosted verification: Actions run 40097188476 is completed/success; its Python 3.12, 3.13 and 3.14 jobs and all reported steps succeeded. I did not independently execute a local pytest or historical replay in this review; CI is hosted evidence, not my own execution.

COMMENT, not merge/release approval: #89 remains draft and stacked on unmerged #88 and #81; the release branch is still 22a2894057d9b1bc003efe4ef932ed4de601e26b. Do not merge this cumulative PR before reconciling its prerequisites and independently reviewing the integrated exact head. Keep #83 open (parent-kind persistence is separately proposed in #90), and keep #87's legacy non-value-bearing statuses, completeness and valid PARTIAL semantics with the research authority. This review must not be interpreted as accepting those contracts or the inherited release scope.