[Clear1nsight]

Change and scope

The frozen H25 runner reads complete retained stdout/stderr into memory when producing hashes and counts. This adds a separate, source-bound bounded-log runner package: chunked hashing/counting bounds retained-log memory, and a final output-size check preserves STOP behavior when a fast child exits between live samples. The PR6029 package remains unchanged.

This is tooling and a finite defect repair. It does not claim scientific replay speedup, a C implementation, a new bound, κ gain, descendant containment beyond the original process-group contract, or completion of the six-area program.

Task and overlap

Base: 2485805. Head: f99829c. Only research/h25-bounded-log-guard is added (21 files). TASK.json preserves the exact replay-runtime card and records changed scope and partial outcome: the original before/after throughput acceptance gate is not claimed. Nearest predecessor: #6029, whose existing runner and evidence are credited in PROVENANCE.json and historical/. The added value is a bounded retained-log implementation with portable finite replay, without rewriting frozen predecessor evidence.

Validation

Linux, CPython 3.12, standard library only. From repository root:

python3 -B research/h25-bounded-log-guard/verify.py
python3 -B -O research/h25-bounded-log-guard/verify.py

Independent source review accepted adapter/source conservation. Parent-owned serialized validation passed normal and optimized Python: eight process controls plus eight real-file controls per mode, including binary data, stderr, nonzero exit, live oversize and fast-exit oversize. The original 32 MiB file fixture observed at most 134,752 traced bytes. Final exported source also passed with files read-only, directories read-only, empty PATH and no Git checkout dependency. Package metadata/pins passed; exact local validated Git tree equals the published tree 71a5e8e5d7e82d8ce994eaa3af864bd2ab65f35b.

Each registered command is capped at 180 seconds; original process controls retain their bounds and raw expected-STOP evidence. VALIDATION.json records executed receipts and hashes; README distinguishes earlier preparation snapshots from completed release checks. Failed staging/adapter attempts and historical accounting gaps remain retained locally; physical administration/publication timing remains unknown and is not reported as zero. Hosted CI has not yet completed at submission.

Credit and next questions

Chris R / Clear1nsight; substantial OpenAI Codex assistance. Apache-2.0; predecessor/source and independent-control credits are retained. Source bytes of the stabilized runner and original finite controls are conserved.

Smallest follow-ups, requiring separate review and controls: explicit failure receipts for retained-log path drift/open errors, and a separately specified containment contract for detached descendants. These limitations remain explicit; no broader research or automatic integration is included.

[Clear1nsight]

Publication handoff — hosted checks complete

• Result: standalone bounded retained-log defect repair; upstream PR6029 remains unchanged. Companion archive PR: Add lossless evidence archive and exact restoration controls #6317.
• Source binding: exact head f99829cfc669c0aa61d7d505b0a58f7e697c6582, tree 71a5e8e5d7e82d8ce994eaa3af864bd2ab65f35b, parent 2485805a6d47f36ea83ddde6327d453ac82d44dc; all 21 committed files match the tested read-only export.
• Hosted validation: run 41136195929 passed both plan and package replay at this head. Shared suite: 710 tests, 24 platform skips. Selector chose only research/h25-bounded-log-guard. CPython 3.12.15 replay passed normal and optimized modes; committed inputs unchanged.
• Coverage: eight original process controls and eight real-file controls in each mode; expected STOP/raw evidence, byte hashes/counts, source custody and bounded-memory gates retained. Earlier local normal/optimized and sealed-export checks also passed.
• Cost boundary: no before/after verifier throughput measurement or scientific speedup claim. Historical accounting gap and untimed physical publication tail remain explicit; hosted execution is separate from the local validation allowance.
• Credit: Chris R / Clear1nsight, predecessor/source and independent-control authors; substantial OpenAI Codex assistance, Apache-2.0.
• Remaining: retained-log path-drift/error receipts and detached-descendant containment need separately scoped work. This PR is open, unmerged; the broader six-area program remains paused and incomplete.