[Clear1nsight]

Change and mathematical scope

Records the precise interpreter boundary behind the P07/N22 G12 allocation-failure roadblock. A bounded journal can remove explicit bookkeeping allocations, but capacity alone does not establish complete acquisition registration, raw-error capture, or outward primary identity.

The package adds a conditional argument with three finite witnesses, 19 unchanged source windows from 11 immutable external files, and a 13-case finite transition model. The decisive CPython 3.12.3 paths were checked directly against upstream source at f6650f9ad73359051f3e558c2431a109bc016664. Target-build correspondence remains unqualified. This is not a generic Python impossibility theorem, C execution, target fault injection, or original CP12/CP15 acceptance.

The acquisition witness does not refute an obligation limited to Python-known cleanup. The pre-capture witness expressly identifies its raw-error observer interpretation. Journal reachability and delivery through the primary channel remain distinct. Any additional runtime assurance or weaker outward-error contract must be explicit.

Task and overlap

Explicit user assignment: use the completed P07/G12 Pro consultation to address the roadblock and submit scoped upstream contributions.

• Canonical base: 626e936ce4d61f2584316914fc71baf96940c512.
• Public head: 59140439e7f791876159e2f3838f1dead2e54953.
• Reviewed/public tree: 97b1a57f9f1f4019f4069610f4df85a7dfb665ce.
• Canonical investigation: Archive H25 cache source components and hand off N22, P07 and P02 #6321. Exact proposed predecessor Cursor/h25 g12 identity cleanup 5d9d #6339: 608ad23adc6c01264c8195daf7f18e29c17464a7.
• Companion Repair bounded G12 custody cleanup and publication-size checks #6346 implements three independent narrow repairs. This package addresses the different missing implication in a proposed stronger journal; it neither copies nor requires Repair bounded G12 custody cleanup and publication-size checks #6346's executable sources.
• Scope is changed from the retained portable-replay card to the explicitly assigned fixed-source G12 feasibility question.
• All9 added files are under research/h25-g12-interpreter-boundary/. The three canonical contract inputs are live byte-bound dependencies. External whole-source hashes and exact windows are explicit; CI does not refetch external files.
• Fresh all-state PR search for G12 + MemoryError returned no additional result; earlier G12 search and the Handoffs and requests for independent checks #1166 H25 handoff identified Cursor/h25 g12 identity cleanup 5d9d #6339. Those bounded searches are not proof of novelty.

Validation

Executed on Python3.12.3:

python3 scripts/research_package.py --package research/h25-g12-interpreter-boundary --check-only
python3 scripts/research_package.py --package research/h25-g12-interpreter-boundary
python3 scripts/check_research_commit.py --package research/h25-g12-interpreter-boundary

Metadata/pins, registered normal/optimized replay and committed-byte preflight pass. Each mode reports transition_cases=13 source_windows=19 target_runtime_qualified=false original_acceptance=false. The byte preflight checked12 package/input files on the reviewed local commit. Remote readback verifies the exact reviewed tree, two linear commits from the canonical base and all9 added file blobs.

The final field name explicitly means a descriptor was returned to Python, without claiming subsequent ownership registration. The model is an explicit abstraction of cited transitions, not a C semantics proof or an interpreter simulator. It preserves failure witnesses and success controls, rejects impossible failure placements at nonallocating sites, and remains effective under -O. Complete external texts were fetched and inspected locally; CI binds retained excerpt bytes and canonical contracts only. Hosted checks are pending at submission.

No target import, real descriptor/process operation, native build, original H25 replay, provider acquisition, historical budget transfer or performance measurement occurred. Original CP01-CP15, both exports, the uncached producer, all18 branch obligations, 52 helper IDs and complete costs remain unchanged. Launch/handoff stay None. No N22/P07 body credit, source adoption, O03 completion, gain or kappa change. Unknown physical and final costs remain OPEN_NOT_ZERO.

Credit

Chris R / Clear1nsight, with substantial OpenAI Codex source analysis/verification and ChatGPT Pro consultation assistance. Credit unpaidmoon with substantial Cursor Grok assistance for #6339, PR6321/6029 source/review contributors, Douglas R. Colkitt and the retained circuit/network/native/verification lineage.

New analysis/model: Apache-2.0. Unchanged CPython excerpts retain the complete pinned CPython LICENSE and PSF/predecessor notices in LICENSE-CPYTHON.txt; they are not relicensed. AI agreement is not human or proof-kernel acceptance.

Open next questions

See CLAIM.md and TASK.json.

1. Qualify the target implementation and exact original raw-error observer boundary; enumerate capture, handler, call and traceback transitions through the unchanged caller channel. No runtime assurance is presumed.
2. Prove a faithful atomic-owner primitive or private-table acquisition recovery alternative. Failed return conversion and failed post-return registration are different gaps.
3. Then compose and independently review the bounded journal within the original exceptional contexts and full parent/consumer/authority/driver source graph. A premise amendment cannot be labelled ORIGINAL acceptance.

[Clear1nsight]

Final publication verification at 59140439e7f791876159e2f3838f1dead2e54953: hosted run 41365279177, attempt1, passed planning and the actual registered bounded package replay. Neither job was skipped. Earlier head ef4dcefd3c02f573a79796e7aeb55c9cae71af2b remains preserved; only this final-head run supplies the current hosted result.

The remote tree 97b1a57f9f1f4019f4069610f4df85a7dfb665ce, two linear commits and all9 intended file blobs match the reviewed bytes. The final naming correction distinguishes successful Python return from later ownership registration; all13 cases and the three witness claims are unchanged. Both modes pass locally and through the registered hook. All11 locally retrieved whole-file hashes and19 literal source windows were separately read back; hosted CI checks retained windows, not target-build correspondence.

Canonical evidence and next gate; hub pointer. Companion #6346 also has successful hosted plan/replay at its exact head.

The PR remains draft with maintainer modification disabled. The original source gate remains OPEN: no target execution/fault injection, generic Python impossibility, full N22/P07/O03 acceptance, adopted source selection, measured gain or kappa is claimed. Unknown physical/final costs remain OPEN_NOT_ZERO.