[ToikY]

Change and mathematical scope

Add a source-bound, opt-in VERIFY-only parser-policy candidate and a genuine-Git regression harness. Existing export already rejects duplicate object keys; copied registered validation/planning are patched to reject them before dispatch. Only research/early-duplicate-verify-rejection is added. Shared scripts and workflows are unchanged; adoption requires a separate maintainer decision.

This is tooling evidence, not scientific false acceptance, a security result or a documented metadataPASS-implies-export promise. Repeated strings in a unique-key inputs array still pass registered intake and fail export. TASK/SOURCES parsing, numeric extensions, catalogue/provenance readers and unrelated policies remain outside scope. No supplier, new kappa, selected-assumption discharge or frontier change.

Task and overlap

Canonical main 7c49fe3 and priority revision8 were refreshed at entry and before publication without drift. The complete portable-replay card is retained with explicit CHANGED/partial scope: a metadata/export interpretation boundary, not a completed production rollout.

Live hub/index, all756 open PR bodies/heads across eight pages, scoped helper and bounded all-state indexed searches were checked. Bodies/titles were unchanged on the publication refresh; hidden/unindexed work remains unknown. Advance claim. Ali Romman's 5728, head70441d90c78bdfc48aed772d2b45e05d5a38e4fb, already implements and tests duplicate committed-VERIFY export rejection. 5726, head20bed137e8807376f3124dd6d03fc076e595b016, covers separate strict coordination JSON. Those results are credited, not rediscovered.

Validation

CPython3.12.14, standard library plus local Git, POSIX. Four disclosed preclaim original-tooling calls gave healthy metadata/export0/0 and duplicate0/1 on the same committed397-byte manifest, with no inert dispatch. The new harness then passed its diagnostic, direct normal/outer-O runs and both final registered commands. The final registered complete receipts are identical.

Each harness run performs144 actual tooling CLI operations over9 committed cases, two source phases, two inner Python modes and four entrypoints:100accepted,44rejected,28inert marker dispatches,9temporary apply checks/applications. Four duplicate forms reject with exact exit/diagnostic and no marker: same-value, conflicting, escaped and nested keys. Positive TASK/SOURCES/numeric controls and the remaining input-array limitation are explicit. Outcome-table SHA256 bfdbc3acb6cc9e66ee8766654a9c196d9d2c8331039663516de610a0edff7562. These are tooling calls, not144manifests or scientific claims; no mathematical/native predecessor verifier ran.

Committed bytes/modes, canonical source ancestry and official export/input coverage pass. Both read-only exported normal/O runs outside the original checkout reproduce the complete registered receipts with Git still available; all14members remain unchanged. Eight source/patch corruption controls reject before any source CLI. All13public blobs match and the public tree equals the tested tree. Static source-aware policy/harness/metadata reviews found no blocker; role overlap is disclosed, not blind/human/kernel acceptance.

Coverage limits: planner positives use changed package files, not input-only or sparse invalidation; optional sidecars/GitHub output are untested. The known inert marker has no children and a five-second registered cap; process-group limits are not universal descendant containment. The approved platform matrix, historical suites and legacy duplicate inventory remain unrun. Initial isolated-mode official-export attempts failed on the script's sibling import before export; the supported invocation passed, without code changes. Original logs are retained.

Reproduce:

python3.12 -B scripts/research_package.py --package research/early-duplicate-verify-rejection

Expected: PASS_EARLY_DUPLICATE_VERIFY_REJECTION and PASS_REGISTERED_RESEARCH_PACKAGE. Both original hosted checks passed on attempt1 at 2081246 in run42249215161: plan (711helper tests,24skips) and both registered tooling commands. CPython3.12.16 hosted complete receipts equal local exactly. Raw logs were retained and compared character-for-character with the fetched logs. The checkout has the exact tested public tree and canonical/public parents; it is a synthetic PR test-ref, not a repository merge. This CI runs tooling fixtures, with zero mathematical/native predecessor scientific commands.

Credit

Prepared for ToikY with substantial OpenAI Codex assistance across patch/integration, source-aware prototype/harness, policy, overlap and metadata roles. Credit Douglas Colkitt's runner, Tekalign's exporter, Ali Romman's5728/5726strict-parser work and all retained headers, Apache-2.0 license and notices. The exporter guard is prior work.

Open next questions

Inventory every committed registered VERIFY before any shared adoption: the candidate planner rejects ambiguous manifests even in otherwise-unselected packages. Then obtain the maintainer policy decision, refresh exact source pins and run approved historical/platform/integration checks. Input-array policy is a separate question. All five selected assumptions remain open.

Package, patch and exact sources.