Older encoded transcripts: exact RAM trial, insufficient material capacity
All 39 older local gzip transcripts outside the 1800 session pool reconstruct exactly, but the combined raw-XZ trial saves only 20,217,856 allocated bytes before recovery metadata and evidence costs. No original was removed or persistently encoded. Original NFS actor/Pause/driving remains unproved.
Evidence: scratch/runs/20261113-older-closed-2030/. The 1730 inventory was only
a candidate index. A fresh filename/metadata census found 72 local
.codex/sessions/**/*.jsonl.gz files, all already inventoried: the excluded
33-file 1800 pool and this disjoint 39-file superset. The active root thread
01a1ba23-0e03-7ef3-9b07-c08de15542d3 is explicitly excluded. No raw active
session, live SQLite/authentication, source/media/save/ISO, remote-only backup,
1830 application log or protected recovery payload was changed. Payloads were
opaque bytes throughout.
Older proof, without a modern-schema shortcut
closure-audit.json records the actual older records and original append-only
board closure statements. 25 paths have original recovery or closure
statements; 14 remain unresolved in this bounded audit. Sixteen original
recovery records name retained gzip/raw identities and recipes: 14 in
20261109-coordinator-retired-event-compression/session-log-manifest.json, one
in third-session-manifest.json, and one adjacent .jsonl.recovery.json.
Every recorded full encoded hash, raw hash and raw size present in these
records matches the current original and the complete opaque decoded bytes.
All these older metadata files remain unchanged.
The third-session manifest and original November 9 board retirement explicitly
cover the older 01a1ba23-f5c1-7c83-8c7d-024ead7f758e session. Its modern ledger
entry is absent; that absence did not reject its original evidence. Likewise,
original November 7 board entries describe the closed 01a1b4ec... archive
and an earlier gzip-to-XZ attempt. That old archive attempt did not establish
the raw-XZ result: this trial obtains 22,861,972 raw-XZ bytes from its 32,643,661
gzip bytes and reproduces its original gzip exactly.
Actual current/pending registry and task metadata have zero direct candidate
identity/path matches. Two independent privileged /proc censuses find no
candidate-inode FDs/maps, producer-token matches or census errors; the second
runs after the codec process exits. The complete original metadata signatures
and encoded hashes are stable after the trial. These checks do not upgrade
the superset to removal eligibility: full transitive consumer and foreign-claim
qualification is deliberately unasserted, and unresolved closure stays
unresolved. Exclusions can only reduce the gross removable allocation.
Complete RAM-only codec experiment
The unchanged 1830 probe algorithm was reused, pinned by SHA in source.json.
It parses codec framing only, preserving optional header bytes and trailers,
checking header CRC where present, complete single-member deflate consumption,
raw CRC and length. Pinned zlib 1.3, raw window -15, memLevel 8, default
strategy and a whole-stream compression followed by Z_FINISH reproduce every
original encoded byte. Per-path proofs record the proven level, original
metadata, full encoded/raw hashes and exact header/trailer. No flush schedule
or payload meaning was guessed.
All 39 individual raw-XZ round trips pass complete raw identity and byte-for-byte
original-gzip reconstruction. The combined 109,346,621-byte raw stream produces
41,421,392 XZ bytes, SHA-256
686e799aa98ddc6c87c291de810acc1490482e4c87c11437aa2abbb8756cacb2.
Complete decoding and each original gzip reconstruction pass again. Independent
external gzip -dc verifies all 39 full raw hashes and sizes. The actual probe
parent wait returns zero. XZ existed only in process RAM and was discarded at
successful exit; no new durable recovery layer is needed or claimed.
Original allocation is 61,640,704 bytes, encoded size 61,553,034. Optimistic 4 KiB rounding of the combined XZ gives the 20,217,856-byte saving, before any index/helper/recipe or evidence charge, below the established 30 MB material criterion. This measured full-pool result is not a mathematical bound on every possible subset or codec. The stronger capacity obstruction needs no codec assumption: even zero-cost removal of the entire 61.6 MB superset cannot bridge the roughly 594 MB gap from 905,678,848 bytes available to 1,500,000,000 bytes. No partial small cleanup, filesystem encoding, unlink, RAM-file restoration or native floor-only retry followed.
All five parent manifest entries plus the parent manifest itself retain their protected modes and identities, and 675 older dependencies match their recorded full hashes and sizes. No older helper or index changed. The 1830 original gzip files must still be hydrated byte for byte using its protected reconstructor before any older helper, history/resume or consumer needs them. This experiment does not relax that rule or claim metadata/refusal restoration tests for a removal that did not occur.
Actor fix bcd413390, unchanged chooser/media, fresh current freeze, sustained
/home capacity above 1.5 GB, heavy lock/actual-child census and independently
prearmed observable watchdog remain prerequisites for native NFS. Live actor
binding, reviewed Pause/Continue and finite held driving with independent UP
remain open. No guest/config/clock/AI write, blind chain, UNKNOWN replay,
forced Present, reserve-2518 setting change, denied transport/credential retry,
worker or box action occurred.
Only this note is committed; production code did not change, so no production build was needed. Full codec, external decoder, privileged census, original identity, prior-proof identity and protected-dependency checks passed. Git SHA, scoped binary diff and owned new-file patches preserve source provenance; there is no source/media archive. Box ID NONE; state NOT_CREATED; expiry null; stop receipt N/A. Next requires another authorized material capacity source or explicit owner action; this worker stops with the concrete capacity blocker.