Closed application logs: material exact-gzip recovery; NFS below floor
73 closed application/API runtime gzip files have exact protected local XZ recovery. Net allocated savings are approximately 37 MB, including the entire new run and its evidence. The subsequent original NFS decision refused at 929,136,640 bytes available, below 1,500,000,000. No guest, input, live UI actor binding, Pause/Continue or held driving was started.
Evidence: scratch/runs/20261113-application-exact-xz-1830/. This is a different
pool from the 33 canonical session streams in the 1800 experiment. Those 33,
nonledger/coordinator sessions, the earlier raw-log/module pools, sources,
media, saves, ISOs, databases, authentication and remote-only backup were not
removed. No denied API/SSH/exec operation, credential change, transport, worker
or box was retried or created. Log payloads remained opaque bytes throughout.
Bound and exactness
The 1730 inventory of 1,157,441,767 encoded bytes was used only as a candidate index. Metadata matching examined inventoried scratch-run runtime log names, committed run notes, root completion-ledger reviews and recorded completed shutdown identities. The two disjoint bounded trials examined 258 gzip paths, 69,721,561 encoded bytes, rather than stopping at the first subset. This is not an eligibility claim for the inventory or every possible local log pool.
157 trials reproduced their original encoded stream exactly; 101 were refused for unmatched whole-deflate settings or framing/member forms. The final removal plan selected only 73 application/API runtime paths from 20 independently root-reviewed closed runs, excluding wrapper/preflight trial paths. Current and pending task metadata, foreign claims, actual producer cmdlines and inode FD/maps were checked. Current source references found during the audit name retained helpers/media rather than the selected log payloads. The remaining 185 original trial files are unchanged; all retain their full encoded SHA and inode metadata, and all 163 with previously established raw hashes passed their complete raw hash again.
Selected originals contain 68,450,507 gzip bytes, occupying 68,661,248 allocated bytes, and decode to 640,204,678 opaque raw bytes. 56 selected gzip files have optional header fields. Those fields were parsed only as codec framing and retained verbatim, including framing checksums where present. Every admitted member has verified CRC/length, completed deflate, and no unused or unconsumed input. Unknown or multimember/flush forms were refused.
zlib 1.3 reproduced every selected original using raw window -15, memLevel 8, default strategy, one whole-stream compress and Z_FINISH: 26 streams use level 6 and 47 level 9. Headers and trailers are restored verbatim. No flush schedule was guessed. Complete encoded and raw SHA-256 checks and actual byte-for-byte comparisons preceded persistent encoding and every unlink.
The protected XZ is 20,011,596 bytes, SHA-256
bf094b1d592e857fe35a58cb12183b1305acff8a8091a55d77f88fc0210c68bd.
Its complete decoded stream is 661,208,140 bytes, SHA-256
0ca7cda61e08da691efa6f324caccf0c1c86df90ec0e2f04a288d82987c25a09.
The stream also retains unused opaque raw bytes from exact but unselected
trials; their original gzip files remain intact. The index names offsets only
for the 73 selected paths. XZ preset 9 without extreme was used for this final
combined pass; this does not change the proven gzip reconstruction settings.
Both preceding subset passes also finished successfully. All three actual
parent waits returned zero.
Before persistent encoding, the complete recovery directory was staged in real tmpfs and its allocated blocks measured. Charging the existing entire run plus a further 3,000,000-byte evidence allowance still established 35,100,992 bytes net saving in RAM. Persistent recovery was fully written, fsynced, hashed and protected before the first unlink. All 73 removals completed; this was not a partial small cleanup. The net figure in the opening paragraph also charges the run's exploratory metadata, validation and source patches. Shared filesystem availability is reported separately; concurrent usage means its change is not attributed solely to this task.
Protected recovery: hydrate original gzip first
Retain this entire directory, its exact paths and every named older dependency:
scratch/runs/20261113-application-exact-xz-1830/protected/
It contains original-raw.xz, INDEX.json, reconstruct.py,
OLD-HELPER-DEPENDENCIES.json, RECOVERY.txt and MANIFEST.json. Files are
mode 0400 and the directory 0500. The dependency document names and hashes
675 retained older helpers, indexes, recovery metadata and codec/interpreter
dependencies. Every named file was rehashed unchanged after removal.
No existing old helper or index was edited.
Before any old raw restore helper or existing consumer needs a removed log,
hydrate its exact original .gz first:
python3 scratch/runs/20261113-application-exact-xz-1830/protected/reconstruct.py \
scratch/runs/20261113-application-exact-xz-1830/protected/INDEX.json \
ORIGINAL_ABSOLUTE_GZIP_PATH
Then invoke the unchanged old helper/index or consumer normally. Do not pass this XZ to an older gzip restorer, and do not substitute newly gzipped raw bytes. The reconstructor requires the pinned zlib version, verifies complete XZ/raw/original-gzip hashes, preserves original gzip mode and nanosecond mtime, and refuses overwrite and symlinks, including symlinked parents.
All 73 files passed real /dev/shm restores before and after unlink, with
mode/mtime assertions and overwrite/symlink refusals. After exact gzip
hydration, the unchanged 1640 opaque raw-restorer restored every raw byte
and passed its own metadata and refusal checks. Its fixture indexes existed
only in tmpfs; historical indexes were unchanged, and historical raw metadata
was not invented. This tests the opaque restoration chain without running an
analysis consumer or interpreting payloads. Independent external gzip and
xz decoders passed complete SHA checks in both phases.
Before each unlink, an independent privileged census rechecked actual current-inode FDs/maps, producer cmdlines, unchanged root closure records and current/pending exclusions. All 73 passed without hits/errors. The final privileged census found zero surviving original inodes or deleted-original FD/maps references. Protected data and all full hashes were checked again. Only redundant new tmpfs copies were cleaned after durable recovery and its post-unlink tests passed; the protected XZ/index data remains retained.
Original NFS decision
The 1550/1640 notes and actor fix bcd413390 were followed. The fix is an
ancestor of current main. A fresh 451-path freeze includes the current
runtime/source dependencies and unchanged existing media by path/hash, without
copying media. The native chooser export remains
e532c367bf58953a0d629ae0f3e5df93b8239daab0f713c1d8945e2dd92f087d.
The only changed pre-existing 1640 pin is the subsequently committed
test/run.js; three additional current dependency pins are included.
The one post-resource decision held the heavy-runtime lock and used a privileged actual-child census: zero heavy children, no census errors. It refused at 929,136,640 free bytes. The supervisor returned zero after recording that refusal and releasing the lock; no native process was created. No deadline/disk watchdog was started because runtime admission refused before allocation. An initial read-only ancestry check omitted its second Git argument; that preflight error was retained and corrected before this single capacity decision.
Live UI-actor binding, reviewed native Pause/Continue, two stable paused frames and one finite >=1500 ms held driving interval with independent UP remain UNPROVED. No guest/config/clock/AI writes, blind chain, UNKNOWN replay or forced presentation occurred. Next requires separately authorized sustained headroom above 1.5 GB, then a fresh freeze and heavy-child check, independently prearmed observable deadline/disk guard, live actor inspection/bind and manually reviewed native input. The resource/runtime outcome was posted on the board; NFS priority was released without another floor retry.
Only this note is committed. No production code changed, so a production build was not required. Meaningful checks are the full codec experiments, genuine external decoding, real before/after restoration chain, privileged per-unlink closure censuses, all-original/dependency rehashes, syntax and evidence seals. Exact Git SHA, relevant binary diff and small owned new-file patches preserve source provenance without source/media archives.
Box ID NONE; state NOT_CREATED; expiry null; stop receipt N/A. No existing runtime was stopped or box resumed/deleted. This single worker stops after publishing the concrete material result and capacity blocker.