Half-Life: cold cached executable recovery and fresh combat attempt

Run scratch/runs/20261112-halflife-cold-cache-headroom-combat-2200/, with fresh lifecycle evidence in sibling -control/.

The explicitly authorized cached executable has been losslessly recovered. Actual /home free space changed from 1,462,583,296 to 1,659,904,000 bytes, an observed 197,320,704-byte increase under concurrent writers. This is the measured filesystem delta, not a logical-size estimate. No old source-archive inventory or exhausted HL/Q3 log compression was repeated.

The authorized ~/.codex/packages/app-server-daemon/releases/0.167.1-x86_64-unknown-linux-musl/codex is a relative symlink, bin/codex. The symlink and release directory/provenance/package metadata remain intact. Only that resolved regular executable, inode 2715739 with one link, was compressed and unlinked. Other binaries in that release remain untouched. The current selector remained exactly on 0.167.2-x86_64-unknown-linux-musl; active appserver, code host and other observed release executables resolved to 0.167.2.

Before compression and immediately before unlink, privileged /proc scans checked executable identities, every readable process mapping and file descriptor. Both scans completed without permission errors, candidate consumers or writable package/installer descriptors. The existing installer lock was held exclusively through recovery. Source device/inode/link count/size/mode/mtime/ctime and selector/selection metadata remained stable, with repeated complete source hashes.

The 320,186,856 original bytes have SHA256 dd5bfdf19d794dcc6fb482d2d5a2d9fb31d11f6caa44b303bb72890d461e8a02. The adjacent bin/codex.recovery-2200.gz contains 122,456,456 bytes, SHA256 1dd3dc00553dbff99512f892984c00b9d1e99e924147108f2c8efc62952a5d9f. It retains mode 0755 and mtime 1794334832906833000 ns. Gzip completion and its directory were fsynced. An independent gzip -dc process decoded the entire stream and matched raw size/SHA before unlink; a separate subsequent Python gzip read verified the complete raw stream again, encoded SHA, mode/mtime, absent raw path and preserved symlink.

Adjacent bin/codex.recovery-2200.json and bin/codex.restore-2200.py were exclusively created, fsynced and directory-fsynced before unlink. The journal pins the raw and encoded bytes, original metadata, selector state and restore recipe. Before any future selection of 0.167.1, run python3 /home/user/.codex/packages/app-server-daemon/releases/0.167.1-x86_64-unknown-linux-musl/bin/codex.restore-2200.py. It verifies encoded SHA, streams into a unique temporary file, restores mode/mtime, fsyncs, verifies full raw SHA/size, exclusively links the exact original target path without overwrite, fsyncs the directory and rechecks the published bytes/metadata. The retained original alias then resolves again. It never switches selectors, starts a process or deletes encoded recovery bytes. Do not select that release while its executable is compressed.

cold-preflight.json, cold-immediate-preunlink.json, preflight.json, recovery-journal.json, recovery-result.json and independent-recovery-verification.json preserve the checks. Unique gzip and recovery metadata remain local beside the original executable; the evidence run keeps journal/check records. No appserver restart, update or downgrade occurred. Current-release files, selectors, authentication/configuration, sessions, databases, cache/RAM/reserve settings and guest media were not modified.

The fresh attempt used the qualified 1 MiB caller, commit 0264ca075, with unchanged provider/custody, native Pause, finite input/cycle, semantic ACK and capture-selector helpers. The provider sent the large pak as 76 ordered ranges of at most 1 MiB, assembled and independently reread its whole SHA, and verified every remote file. All 956 frozen uploads passed; the actual provision child joined exit 0 with separate stdout/stderr EOFs. Final remote hash and staged-import checks passed before original launch. Relevant Git commit/diff, the one relevant untracked source patch, owned helper patches and media path/SHA references are retained. No source or media archive was created. The old 2055 whole-file PUT remains UNKNOWN and unreplayed.

The fresh browser mounted all thirteen earned1825 records and verified them before native Load. Normal UI selected the bottom 03:18 UPLINK row (a04-earned0318) and activated Load. The game then reached the living corridor at 35/0, using actual module SHA256 04c534192ee3fe39e388536c2ce710c12c47a5cd2ca4d95c19f9d00b1e8f0a17. Native Pause visibly settled at 30/0 in a11-paused-state. The actual CPU slice sequence advances from 34773 to 74719 across the run. The bound owning GL endpoint uses the software GL backend; the shell's WebGL selector does not establish a WebGL guest renderer. All 44 retained full PNGs pass their recorded SHA and CRC checks, and all 44 before/after bindings retain the complete 22-field identity. These checks establish identity and frame integrity, not successful aim or combat.

The camera/target sequence is a distinct original combat attempt, not a new capacity hold:

All 48 independent UPs in the retained pulse receipts are acknowledged and were armed before their DOWNs; recorded native events are trusted. No console, configuration, clock, AI, health, position, aim-memory or media write was used. No native Save or successor Load was issued. The normal-close read-only export independently decodes and hashes to the exact original thirteen path/size/SHA records, totaling 2,342,063 bytes. The original compressed checkpoint still has SHA256 df8d61b2fcdd285a6a28e2e180930e6ee60eff388cd9bf31c557d3409dd18e51; no unsuccessful state replaces it.

Normal native quit ACK31 succeeds. Chrome exits 0; the actual outer parent wait is 0 with separate stdout/stderr EOFs. The native process drain reports 361/361 owned processes reaped, complete command-log EOF and no survivors. All 453 remote evidence references are recovered locally and hash-verified before requesting box shutdown; retrieval-verified.json and final-verification.json record the independent reread. The closing capacity sample is 1,532,805,120 bytes free, still above the floor. No continuous free-space guarantee is inferred between samples.

One new private box, bx_h9s6q8va, was created only after its independent two-hour watchdog was armed and observed alive. Initial expiry was 2026-11-12T23:42:20.947Z. It is now archived, expiry null, stopped 2026-11-12T21:57:06.330Z. Normal stop stop_455031e54b1e930d completed, forced false, error null. The watchdog has actual parent wait0 and both EOFs. The local manager-parent tool session returned143 and did not publish its parent-wait file: that parent wait and its separate EOFs remain UNKNOWN, distinct from the verified watchdog wait, remote original wait and completed normal stop. No inherited box was resumed, stopped or deleted.

Documentation-only publication uses link/whitespace/diff checks; no production code changed and no build is claimed. Next combat work can use the preserved original checkpoint and the newly earned live target/motion evidence, but needs a settled centered target and supported range with fewer exposed aim corrections before the attack. This task stops after verified cold-binary recovery, the evidenced unsuccessful original combat attempt and normal box cleanup.