Redline retrieval snapshots, 2026-11-09

Fixed an actual retrieval integrity and resource-bound defect in tools/redline-publication-caller.cjs. Previously Node opened bounded regular files with NOFOLLOW/NONBLOCK and returned hashes of those bytes, then Python reopened the paths for preflight, artifact hashing, and validation. These reads could observe different files and a later gzip expansion escaped preflight's decoded-byte bound.

Retrieval now sends the exact captured buffers over stdin to redline-publication-snapshots.py. This narrow adapter supplies an in-memory path interface to the published launcher's unchanged validator. It expands custody gzip once with cap+1, checks overflow, emptiness and CRC/trailer, then supplies that same decoded buffer to both artifact inspection and custody validation. Drain metadata can select only the captured custody filename. There is no evidence-path reopen in Python. Empty command logs retain the existing explicit valid=false / empty evidence INVALID result.

Executable reproduction uses the actual adapter CLI and Git-loaded old adapter in memory (no source copies). Old retrieval returns a command-log artifact hash that differs from its validation hash after replacement. A second reproduction changes custody and its drain metadata after the bounded gzip preflight obtains its bytes: old validation accepts more than 64 MiB of decoded custody. Both are oldFAIL. Repaired retrieval uses the original snapshots through replacement, growth, FIFO, symlink, compressed replacement and second-read race controls. These small synthetic journals prove integrity behavior only, never custody of a real process. Original inert-child tests separately exercise actual custody.

Evidence: scratch/runs/20261109-redline-retrieval-snapshot-repair-1545/. Exact initial Git base is de80d8a8739e248c79aabc8fdc15c9ca3678d1b0; owned.patch includes only this adapter, memory adapter, test, and note. No source trees, media, or modules are archived.

Validation performed offline:

Watchdogs were observably prearmed independent siblings outside custody wait domains. Owned children and watchdogs were actually waited and streams drained; kernel reaps are not promoted to browser EOF. The original deadline plus 30s cleanup, sole ptrace/wait owner, readiness loop, explicit version barriers, default-OFF argv/spawn options, and generated consumer remain unchanged. Shared launcher/helper/collector and Q2/Q3/CPU/WAT/host providers were not edited.

Four inherited 1515 stdout/stderr gzip artifacts remain INVALID despite earlier PASS labels: the two *inherited*/original/chrome-stdout.gz files are raw-zero; the two stderr files are truncated 10-byte gzip headers with no decoded bytes. Exact paths, sizes and hashes are in inherited-invalid.json. They were neither modified nor promoted. Initial exploratory inventory counted only raw-zero files; the final inventory explicitly includes all four invalid files.

No native attempt, browser, local Chrome, Box, game, retry, worker, agent, or deployment was launched. Box ID/state/expiry/stop receipt: NOT_CREATED / N/A. Historical startup/PID8107 UNKNOWN and cancellation browser-wait UNKNOWN remain unchanged. Per-stop fsync cost is unmeasured; native readiness is false. Next work requires separately authorized current Redline preparation, including the new memory adapter in its staged closure; no native authorization follows from these offline controls.

Recovery 1565 used a fresh session after worker 1545 died without a final report. Inspection found its four files already committed as d7459f3f6; no unreported test result was accepted. Fresh evidence lives in scratch/runs/20261109-redline-snapshot-recovery-1565/, with its exact starting Git SHA and a small patch for the recovery changes. Added executable controls reject growth and truncation during fd capture, and check the explicit INVALID classification and matching snapshot hash for an empty command log.

Before further testing, nine inherited raw journals were losslessly compressed, independently decompressed and compared byte for byte, recovering 14,124,822 bytes. lossless-compression.json retains original and compressed SHA-256, sizes and the restoration recipe. No unique evidence was discarded. The four 1515 invalid gzip files were independently rehashed and remain INVALID. The recovery reran the snapshot CLI suite, all nine original caller scenarios, four launcher tests, four observer tests, 46 kernel tests, eight reaper tests, 14 readiness-boundary scenarios, and the required locked build; all passed. Fresh large journals were also compressed only after their tests completed, with byte-for-byte recovery verification and a separate manifest. Shared providers remain byte-identical to the initial repair baseline. No new runtime authorization or readiness conclusion follows from this recovery.