Closed canonical worker rollouts: verified local gzip recovery

On 2026-11-13, 23 closed wine-worker canonical session files were losslessly encoded as local original.jsonl.gz siblings. Original bytes total 145,065,975; encoded bytes total 95,806,061. The file allocation difference is 49,254,400 bytes reclaimed, before the small retained recovery evidence overhead.

Actual /home available-byte samples were 1,476,890,624 before work, 1,519,116,288 after encoding, and 1,517,903,872 after final recovery verification. These are shared-filesystem observations, not attributed net savings. The final verification sample exceeds the 1,500,000,000-byte floor by 17,903,872 bytes; future capacity decisions require a fresh sample.

Evidence and recovery material are retained in the ignored directory scratch/runs/20261113-closed-worker-gzip-proofs/. Keep that directory and every encoded sibling. eligibility.json contains exact paths, original SHA-256, size, mode, nanosecond mtime, inode and allocation metadata. Each exclusive index-UUID.json additionally records the encoded SHA-256 and size and the closed-worker result/final provenance.

Selection used the coordinator's 24 proofs in scratch/runs/20261113-closed-worker-eligibility-proofs/proofs.json. Referenced closed result metadata and final bytes were read only to verify terminal exit 0, wine-worker profile, presence, sizes and hashes. The entire Diablo II session candidate was excluded. The live registry /tmp/wine-coordinator-current-launches.json, its task/output paths, explicitly supplied current thread IDs, and both coordinator IDs were excluded. Registry identity was rechecked before every removal. Exact eligible paths were owned on the append-only messageboard before encoding.

Every selected file passed first-record session_meta UUID/cwd identity, complete newline-terminated JSONL syntax and stable metadata checks. Other log payloads were not interpreted. Privileged complete /proc FD and mapping censuses found no openers and no inspection errors. Each gzip file and parent directory, exclusive recovery index and recovery recipe were fsynced before raw removal. Whole independent gzip -dc decodes matched original size/SHA. Every file was actually restored in RAM with exact bytes/mode/mtime and overwrite refusal. Final verification also restored every original canonical path from its durable index, checked bytes/mode/mtime and overwrite refusal, rechecked registry/no-openers/stability, then removed the raw files again with the encoded bytes and recovery provenance retained. Helpers were explicitly fsynced before these final removals.

Restore raw bytes before any future history consumption or resume. For one session, run from this checkout:

python3 scratch/runs/20261113-closed-worker-gzip-proofs/restore.py \
  scratch/runs/20261113-closed-worker-gzip-proofs/index-UUID.json

The recovery helper verifies both encoded and decoded size/SHA in RAM, opens the original path exclusively without following symlinks, restores original mode/mtime, and fsyncs the file and directory. Existing destinations are refused. Keep the gzip and index after restoration. RECOVERY.txt is the durable local recipe; encoding-results.json and final-verification.json record per-file checks and observed free space.

No production source, game media, Diablo II dependency, cache, binary, reserve, active/root session, authentication/configuration, SQLite, thread_history or usage-watch event was modified. No offload, new box, worker or child agent was used. Box ID: none; state NOT_CREATED; expiry null; stop receipt not applicable. Focused recovery checks passed; this docs-only publication requires no production build. Next: retain all recovery material, use fresh free-space samples, and restore raw history before consumption.