Exact 23 closed canonical log density, 2026-11-13
LOCAL lossless encoding reclaimed 26,509,312 allocated bytes against the existing gzip files. The original 23 gzip files occupied 95,891,456 allocated bytes (95,806,061 encoded bytes); the retained shared XZ occupies 69,382,144 allocated bytes (69,375,344 encoded bytes). All 145,065,975 original raw bytes and every historical expected gzip SHA remain recoverable. This is actual allocated saving, not a historical raw-size comparison.
The capacity was used for a newly watched NFS box and fresh original-media provisioning. All 609 remote whole-file checks and the actual provisioning parent wait/code0/both EOFs passed. Actual local capacity then fell below the floor: 1,486,725,120 bytes at the recorded refusal follow-up. The browser startup guard refused before remote submission. There was no browser/game launch, native input, Pause/Continue proof or driving transaction. The precise remaining blocker is sustainable actual local capacity after provisioning; this run does not establish a new NFS runtime defect or native progress.
Encoding evidence: scratch/runs/20261113-closed-log-dense-0340/.
The input scope is exactly the 23 canonical, closed, non-Diablo log archives from
scratch/runs/20261113-closed-worker-gzip-proofs/index-*.json. No additional
raw-session scan occurred. The six recent and worker0335's nine archives,
active/coordinator sessions, Diablo dependencies, source/media, databases,
auth/configuration, caches, vendors and reserves were untouched by this worker.
There was no external offload by this worker; shared free-space changes from
other workers are not attributed to this encoding change.
Every original index remains byte-identical. Exact original index, restore
helper and recovery recipe bytes are retained as *.prior, with prior hashes
in prior-provenance.json. The original restore.py command still works through
an explicit mapping.json containing raw offsets, full-stream and per-original
hashes, original gzip SHA and compressor metadata. Recovery requires gzip level 6,
mtime 0 and zlib 1.3, and refuses an expected encoded SHA mismatch. Before any
history consumption, restore the raw file through the original command:
python3 scratch/runs/20261113-closed-worker-gzip-proofs/restore.py INDEX [DESTINATION]
To reproduce the original historical gzip bytes exactly:
python3 scratch/runs/20261113-closed-log-dense-0340/dense_restore.py --prior-gzip INDEX [DESTINATION]
Keep both run directories, all original indexes, the mapping/helpers/recipes
and /home/user/.codex/sessions/2026/11/12/closed-canonical-exact23-0340.raw.xz.
The retained archive SHA-256 is
b8bfdfc465b0e7cb9971327f4a80d5dabc3582cd52c5bec3edcad7a7a89ea039.
Restoration uses O_EXCL/O_NOFOLLOW, preserves original mode/mtime and fsyncs
file/directory. Actual RAM restoration of all 23 originals and bit-exact gzip
reconstruction passed before removal. Independent full XZ decode and every raw
hash passed. Existing-file and symlink refusal, plus digest rejection before
creation, passed. A final post-removal legacy-command restore and independent
stream/segment reread also passed. LOCAL archive, directories, durable index,
codec-aware recipe and proofs were fsynced before retiring old encoded files.
The first per-file stage reclaimed 19,697,664 allocated bytes while retaining five gzip files with small individual benefit. A shared XZ stream over the same 23 originals reclaimed another 6,811,648 allocated bytes through cross-log repetition. No payload interpretation occurred. An initial privileged opener scan refused an unlink because a historical raw inode had been reused by an unrelated live session. The pending archive remained intact. The corrected scanner uses current lstat identities for exact owned paths; privileged FD/map checks and current-registry exclusions passed before removal. Both attempts and the metadata-only diagnosis are retained.
Native evidence is in scratch/runs/20261113-nfs3-dense-capacity-0340/ and
scratch/runs/20261113-nfs3-dense-capacity-0340-fresh/. The first attempt exposed
my missing RAM staging directory: strict local resolution failed before any
provider upload or remote command. Its failure remains terminal. A read-only
file listing returned 400 and is explicitly not absence proof. A corrected new
namespace created the RAM directory before its final freeze and new box.
Another read-only observer incorrectly waited for running; actual API
readiness was idle. Its local timeout wording is explicitly corrected by the
idle observation and actual successful provision receipt; it was not a game
command or an UNKNOWN input replay.
The corrected 607-file freeze preserved current04, all 133 original media files,
provider 5c2 and the corrected c34a7f99 adapter after directory relocation.
Identity/freshness/consumed-review/single-use and terminal no-replay guards were
unchanged. Actual provisioning explicitly used the owned /dev/shm range
folder. Parent f71449c03/remote609/49 outputs/stop0a022681feb64256 remain history
only. No old test matrices, 200ms retry, automatic route, Boids Pause proof,
guest memory/config/clock/AI writes, worker children or public deployment ran.
Both created boxes were promptly normally stopped, with expiry now null:
| Box | Final state | Normal stop receipt |
|---|---|---|
bx_kkyb394h |
archived | stop_a51f377cbcf0a08c |
bx_sb4q28vm |
archived | stop_93c66d321a3ef246 |
Both receipts are completed, forced:false, error:null. Both managers and independent watchdogs have actual code0 parent waits and both EOFs. First provision has actual code1/both EOFs; corrected provision has code0/both EOFs. The corrected box's one remaining receipt file was fetched twice and locally hash-reread before stop; no game/browser runtime was created. Owned RAM staging was removed.
Git commit SHA, relevant binary diffs, reconstructable owned-helper patches, media path/SHA references and sealed manifests preserve the work. No source or media tree was archived. Only this note is committed; no production code changed and no locked build is required. Next: restore additional sustainable authorized capacity, then use a fresh watched box and fresh provisioning for native Pause/Continue with two settled world/timer-stable full frames, followed by one reviewed held drive of at least 1500ms with a frame before independent UP/Escape and a fresh consequence review. There is no floor hold loop or input replay.