Darkstone: helmet restore, second opening and Neophyte’s Cape saved

Restored the actual helmet checkpoint through production HTTP/loadFiles/VFS and native Load a Game, crossed the second stone opening, broke a reviewed barrel, and collected the native-identified Neophyte’s Cape. It equipped automatically and raised AC 12 → 14. Early and final native New Save files were retrieved and independently hashed. No production code changed.

Evidence R: scratch/runs/20261108-darkstone-helmet-1621/. Predecessors: terminal retirement, 275dd018aad23b8e2988a78033cddaa2e25d8df3, and bounded browser startup, b3377657b. Preparation checkout: 83d6b743334c7b65eb3d5599f35ade21b92a73de.

Actual restore and reviewed progress

The selected source pair from Level One’s new-save-recovery.json, staged copies, and restored VFS bytes all matched:

Input Bytes SHA-256
characters.lpl 1,342,204 70fcfbe302a9ff1f1b7452223e67e800371086812d679ba04c03d4adffca036a
game641d_6af08808.sav 3,204,341 a4ecb215f5ffab7524cc0524c419c46498eb7ef14a0e3f404c880f72a769f45d

Before allocation, actual provision.bytes yielded a 1,234-byte manifest, SHA-256 ce610f155b3cee51bc7ff8b79ac8ee0c2343d30af0a5b5d76dfa452c48ac2b01. The local production HTTP handler and remote driver verified those exact bytes. Production loading mounted fifteen data files; read-only observation at 16:23:33.734 UTC independently hashed both checkpoint files in the actual VFS. The native Load preview showed the sole LEVEL 1 save and matching corridor. Ordinary OK restored it. The loading capture retains a pending release; no further input followed until restore-live-settled showed buttons zero and queue EMPTY, with inventory visibly closed.

Personally reviewed captures under R/review establish:

No hostile or human was attacked, no potion was used, and no private guest write, configuration, clock or seed change occurred. Pathfinding continued after some releases; subsequent captures were reviewed before the next target. Every DOWN had its independent UP armed first.

One review error is retained explicitly: final-native-save-menu still showed the world, although the input queue was empty. The next Save a Game click was sent before menu pixels were confirmed. The following capture shows the actual native New Save submenu, and final New Save was selected from that reviewed capture. This is not presented as perfect visual admission at every step.

Recovered incremental native saves

Driver execution began 16:21:59.237 UTC. Early read-only recovery completed 16:28:58.373; unique retrieval and independent hashing completed 16:29:06.230: 426.993 seconds after driver start, or 506.442 seconds after allocation start, before minute ten under either reference. R/early-save-recovery.json retains the reached second-room checkpoint:

Early file Bytes SHA-256
characters.lpl 1,342,212 ab6aeed5b7f22d2ab35908cb1e69c6594683b5e98bb2bdb3abf55b2239237614
game641d_6af0a33d.sav 3,204,341 373a1e43dc19e4235abd80f59a78bb648e8cb8b60285fe2810eecf4435ee09c3

Final native New Save preserved the broken barrel, equipped cape and room position. Read-only recovery completed 16:31:54.796; retrieval and independent hashing completed 16:31:59.447, well before the immutable 16:43:39.788 browser cap. R/new-save-recovery.json selects these continuationFiles under R/final-recovery:

Native path Bytes SHA-256
c:\\dsave\\characters\\characters.lpl 1,342,220 8dca71058d6e7ce4a1b5e148168e7cbd24043c55501b70f388535396eed4e309
c:\\dsave\\games\\game641d_6af0a3ec.sav 3,204,341 787f3c48f60fb174557fe386d7d36ce3a1aec6c33cbb6fc5035b99c41d5f34fb

The loaded SAV and original/staged input pair remained byte-identical. The final pair was not reloaded. Next restore must derive a new actual HTTP manifest for these continuation bytes, then use production loadFiles/VFS/native Load.

Verification and lifecycle

The immutable selected generation preserves 548 exact providers / 42 helpers, original runtime 90c1fcfee91c4680877083e80498afcb73c9384d, ignored module 6cf120d0d3b3db1823a8105eb26ef52a9093a7f871c37a46117475e213fda1c0, exact published kernel 275dd018a, and endpoint helper SHA-256 11683651fa033d56f861adeb64a3f46cb8422553b9651ba051f5200111614ecc. All providers passed cheap exact preflight and remote verification, and stayed unchanged. Joined uploads preceded the unchanged shared content publication; queue and parent both required its exact receipt. Five-file bootstrap and remote lock preceded provisioning. Seven controls, hover, eleven barrier tests and accepted kernel/startup/endpoint evidence were reused unchanged by independently rereading 68 references. No fixture churn, second original or runtime repair. The published terminal repair's 89 retained artifacts were also independently hash-checked; selected kernel and endpoint bytes match exact Git 275dd018a.

R/final-audit.json passes 28 exact command ACKs, zero UNKNOWN, 19 independent pulses, 52 trusted balanced native events, 28 PNG CRC/scanline decodes, exact restored VFS hashes and save recovery. No guest trap, exception or guest process exit was observed. This was docs/evidence work; no code build change is claimed. Exact Git SHA, empty relevant final binary diff, small owned helper patches, provider recovery mapping and existing media hashes are retained. No source archive, repeated media bundle or peer WIP was copied into the runtime.

Ordinary browser completion was requested. Chrome exceeded the existing 8s/5s close grace and exited by its driver’s SIGKILL fallback at 16:32:27.695. The initial evidence reader required exit code zero and refused; its log and remote result remain preserved. A separate read-only inventory recorded the actual signal without weakening process/port/profile/writer/input checks. Browser exit-zero is not claimed. The unchanged kernel supervisor completed normal drain: 359 owned tasks / 359 reaped, exit zero, no remaining tasks, all recorded PID/start identities independently absent, supervisor absent and remote lock free. The 58-identity parent census also drained. Both endpoints and ports closed; profile, writers and held inputs were absent. Kernel fork lineage was complete for this run; CLONE_UNTRACED flags were not audited, historical reconstruction and unsupported nonleader exec remain outside scope.

All 335 remote references / 265,799,084 bytes were uniquely retrieved and independently reread/hash-verified before normal API stop (278 unique contents, 265,784,546 bytes). The independently observable watchdog was prearmed before UUID-idempotent allocation; immutable deadlines were browser 16:43:39.788, independent stop 16:45:39.788, initial expiry 16:46:39.973 UTC. All three recorded local supervisor/normal/watchdog identities are absent, both roles exited zero, and supervisor children are empty. No local heavy runtime or local runtime lock. An initial local finalizer found its board watcher already absent; that harmless ProcessLookupError is retained, and finalization verified its absence.

Box bx_5mw95z27 is archived, final expiry null. Delegated normal stop stop_10273603198cf55a completed 16:34:12.334 UTC, forced false, error null; R/stop-receipt.json independently confirms it. No resume, deletion, deployment, old-box action or subworker occurred.

Next: restore the selected cape-bearing pair and continue from the second room. Full27, parity, audio and FPS remain open. RELEASE and stop after publication.