Half-Life: lossless closed-log recovery and full-frame combat planning

Run scratch/runs/20261112-halflife-lossless-headroom-combat-2130/. Lossless recovery passes, but actual free space remains below the required 1,500,000,000 bytes. No fresh box, watchdog, provisioning, native input or combat consequence occurs. The original play objective remains unfinished.

Before selecting files, inspected 432 existing session/event compression and closed-transcript archive records; inspected-journals.json records their paths. The eligible-worker inventory requires actual start-worker result.json exitCode 0, with launcher and worker absent. The unknown 1500 worker, no-result workers, current worker, root and other games are excluded. Existing gzip archives are retained and never recompressed. Exact claimed raw paths are in selected.json; closed run log ownership follows its terminal worker. No source, media, capture, save or module is packed or removed.

58 raw log/event files totaling 18,927,633 bytes are now local gzip files of 2,721,115 bytes. Each stream is fsynced, fully decoded and compared to the raw SHA256 and size; encoded SHA256/size, original mode/mtime, inode/link count and recovery recipe are journaled and fsynced before raw unlink. The containing directories are fsynced. All selected raw files had one link. A subsequent independent read verifies all 58 raw roundtrip hashes, encoded hashes, modes/mtimes and absent raw paths. Gzip files remain beside their former raw paths. recovery-journal.json is the recovery authority.

Actual filesystem free space changes from 1,472,921,600 to 1,489,154,048 bytes, an observed 16,232,448-byte increase. This is a filesystem sample delta under concurrent writers, not a promise based on logical sizes. The later sample is 1,486,802,944 bytes, still 13,197,057 bytes short of the strict greater-than floor. Remaining eligible immediate raw logs total at most 40,222 bytes, individually at most 4096; even deleting all their bytes would not close this gap. They remain raw. No cache, reserve, RAM, other-worker or old hardlink cleanup is attempted.

Distinct full-frame route and aim review

Personally reviewed the entire existing PNGs, without producing crops or copying media; paths, sizes and SHA256 are in reviewed-frame-references.json:

Planning consequence: do not aim at the earlier floor position after the creature leaps, and do not use the guard silhouette as the headcrab target. A future freshly loaded, genuinely PAUSED frame must choose the target anew. If it is clipped at the left boundary, first use one bounded ordinary turn cycle to bring it fully into view, then review its full consequence and native PAUSED state before selecting an attack. The old 300ms left turn cannot calibrate a new aim: creature motion and clipping confound its result. Only an independently reviewed centered target with visually supported range justifies a subsequent single bounded attack cycle. Preserve uncertainty about hit/defeat until the full consequence actually shows it.

The guard-room alternative also requires a new living paused view of the threshold and floor enemy before choosing one bounded movement cycle. The old follow-on forward2000ms ended in death; repeating that duration from a 5-health historical image supplies no new route evidence. This plan records conditional review decisions, not an adaptive chain for execution.

Preserved transfer boundary and next original run

Reuse the locally qualified bounded 1 MiB caller and frozen helper references from the 2125 handoff, commit 0264ca075. No transfer, 954-pin or whole13 matrix is repeated. Prior2055's 79,150,544-byte whole-pak PUT stays terminal UNKNOWN and unreplayed; caller mismatch proves no production defect. The original earned1825 whole13 source SHA remains df8d61b2fcdd285a6a28e2e180930e6ee60eff388cd9bf31c557d3409dd18e51; it is untouched.

Once actual headroom permits, observe the independent two-hour watchdog before one fresh private creation; freeze helpers and use the qualified 1 MiB caller, all remote chunk/whole hashes, actual provision wait and both EOFs plus staged import before original launch. Restore exact whole13 through normal UI 03:18 Load, prove actual CPU/owning GL/stable22 and native PAUSED pixels before extended review. Preserve corrected Pause/semantic ACK and finite cycles: ordinary resume, one reviewed bounded move or attack, full consequence, native pause, independent UPs. Useful progress must precede a new earned Save/export/Load; preserve original13.

Git SHA, relevant binary diff and the owned recovery-helper patch are retained in the run. Documentation-only publication uses link/diff checks, with no production-code change or locked-build claim. No box is created, so new box ID, state, expiry and stop receipt are not applicable. No old box is queried, resumed or stopped. Prior2055's archived receipt remains historical evidence in the linked handoff. Parent pending reviews are untouched.