Darkstone: streamed FINAL1910 restore, one healing response, new save recovered

One fresh remote original restored only the canonical FINAL1910 pair. Current native pixels established LIFE 14/35, BAT and its red health bar. One brief, independently released right-click targeted the visible red belt vial. The first consequence capture showed 14 → 11/35 with floating damage; encounter actions stopped. The already queued response subsequently consumed the vial and showed 35/35. Native New Save created a new game file and updated character file, both recovered and independently hashed before the immutable deadline, without reloading.

Evidence R: scratch/runs/20261108-darkstone-current-bat-2025/. No production code changed. The run has a quit transport failure, described below; successful repaired-driver quit/profile cleanup is not claimed.

Restore and actual runtime closure

The run reuses the accepted streamed transaction, including canonical gzip source selection, absolute deadline, exclusive fresh remote destination, sticky quarantine and no replay. 525 providers passed the actual uploaded startup barrier. Local require_accepted and remote barrier require both preceded the single original launch. Read-only VFS observation then found exactly these two original files, with matching sizes and hashes:

Original file Bytes SHA-256
characters.lpl 1,342,316 4762ae9c0d7d048595433d121629a87d5a1aa24e7c3ebe84c7605006e3c10baa
game641d_6af0ca48.sav 3,204,341 5e898dc9d6f7a13350c939bf5b712eccfff95ef1072c93023103705999f84ffc

There was no local raw checkpoint stage. The physical runtime closure uses only its active helpers, replacing the inherited collection of unused helper scripts. Exact existing paths/hashes and immutable Git blobs supply selected bytes when the shared working files differ. The native Load a Game preview showed the sole LEVEL 1 continuation, and its reviewed OK loaded that save.

consumer.cjs first invokes the published browser adapter's exact 2535a4fe… source guard over the unchanged run-local browser.cjs. Its anchored changes use published logger 979b3a8b… for browser JSONL and Chrome stdout/stderr from their first bytes, preserve stream-drain-before- profile-cleanup ordering, avoid duplicate executing-module files, and use the owner's 1.5 GB disk floor. Kernel e36d3499…, published in 252888061, launches with both journal_compression='gzip' and command_compression='gzip'.

The original Darkstone module 6cf120d0… and its selected readiness helper remain unchanged. The accepted Q2 module d8facd06… and Q2 readiness 9ba9ae8b… were not substituted into this Darkstone selection. The Q2 logging/kernel proofs were reused by hash. Historical 1940 UNKNOWN was not replayed; historical 1910 SIGKILL / normalChrome UNPROVEN is unchanged.

Current observations and the single response

The initial loading captures showed advancing native loading progress. The inventory key and pointer were sometimes still queued at capture time; no input was repeated merely because its first image was early. A corrected pointer position used the observed canvas rectangle rather than assuming that the rendered game's left edge equaled the canvas rectangle's left edge.

numeric-panel-settled.png establishes LIFE 14/35, MANA 25/25, Food 35, Gold 1713, AC 14, victories 9+0. live-bat-identified.png shows the same numeric health together with explicit BAT and its red health bar. Historical 16/35 and Food 36 are not current readings.

At less than half life, one-survival-response issued one 100 ms right-click on the visible red belt vial. The ordinary input pulse armed its independent release before DOWN; trusted DOWN/UP and both ACKs are retained. Using an item by right-click follows the original manual. No attack or second response was issued.

survival-response-result.png shows 11/35, BAT and floating red 1. Its state still contains the queued right-button DOWN/UP, with physical buttons zero. This first verified 14 → 11 consequence ended encounter actions; the queued response was not credited with healing in that frame. The next save-navigation capture, consequence-panel-closed.png, shows 35/35 and the red vial removed, establishing that the single response subsequently completed. Only inventory/menu/save navigation followed. No kill is claimed. 35/35 is the last numeric reading, not a verified post-save health value.

Exact new continuation

One reviewed Save a Game → New Save produced this pair. Read-only recovery observed it at 20:41:18.174 UTC, before the browser cap 20:51:48.557 UTC. final-save-recovery.json names the retained gzip artifacts and independently verified compressed and original sizes/SHA-256. Neither file was reloaded.

New file Bytes SHA-256
characters.lpl 1,342,324 2c78a17064029385dee4206c9314673debb4de05205ae63d2ff144b9e6081546
game641d_6af0de53.sav 3,204,341 cb2ca8093b2ffe45133cda87e3eea9ceafad66fd82246684ec40dd97306e66fc

The new game path is unique relative to the two-file restore inventory. Recovery retained only this new pair, without another copy of the old game.

Shutdown failure, actual waits, and delegated box stop

I supplied 100,000 ms to a quit ACK client whose selected source permits at most 55,000 ms. It rejected before command publication, then wrote its sticky transport-closed record. The failed attempt and exact refusal remain; admission was not reopened and that command was not replayed. This is an operator error, not a repaired-driver close acceptance result.

One separate public CDP Browser.close was issued. Chrome PID 6691, start 15606, has actual kernel reap sequence 3783, status 0, and Node's chrome-exit.json independently reports code 0 / signal null. Only after that wait, the still-idle driver PID 6683, start 15229, received one identity-checked SIGTERM. Kernel drain records its exit -15. The kernel supervisor PID 6682, start 15221, has an actual parent wait return code 1, because its final success assertion rejects a nonzero root.

The kernel nevertheless records 355/355 owned tasks actually reaped, complete lineage, empty remaining/unreaped/unresolved sets, journal finalized, and kernel command log EOF, with producer original SHA/size matching independent retrieval. Ports were not listening and a fresh nonblocking flock proved the runtime lock free. The last native input state has buttons zero and an empty queue. The browser profile is retained: no successful quit ACK, shared-close stream-drain receipt or profile cleanup is claimed.

All 631 regular runtime files, including files without extensions, every unique log, the retained profile, and kernel command log, were inventoried and recovered with compressed/original SHA/size and gzip CRC validation. Only unchanged selected source inputs were exempted from duplicate recovery. No symlinks were present. No owned-path FD reference was visible, but 48 process FD directories were permission-denied, so this is not a global zero-writer proof. CLONE_UNTRACED flags are not audited; Boat exec nonleader and service-parent terminal waits are outside this kernel descendant scope. Process absence is not treated as an actual terminal wait.

The independent, observable watchdog was prearmed before allocation with an immutable 25-minute cap 20:54:18.557 UTC. The create UUID was durably recorded before the single create request; box identity and expiry were saved before runtime. After all unique evidence was recovered and hash-verified, the watchdog performed the delegated normal API stop. Final box: bx_sgm4f79t / ARCHIVED, provider expiry field archiveAfter: null. Actual receipt stop_e43205641f748219, completed, forced: false, ended 20:45:34.902 UTC, error null. The local watchdog also has an actual parent terminal wait of zero. No box was resumed or deleted.

Validation, source recovery, and remaining scope

The own controls pass full source preflight, changed-source refusal, exact browser guard / changed-browser refusal, and two actual kernel launch controls (empty and 262,173-byte command output) with both gzip settings and actual terminal waits. The locked full build passes. 295 predecessor proof refs were independently reread and hashed; unchanged fixtures were not rerun.

Two preparation checks initially failed (manifest extraction and a mistaken Q2 module expectation), then passed after corrections preserving the original Darkstone module. One early read-only retrieval returned the preceding capture before its new ACK; exclusive creation refused an overwrite. No input replay or evidence replacement resulted. These failures and the quit failure remain in evidence; they are not represented as all-green runtime closure acceptance.

R records the exact Git SHA, complete provider PATH/HASH or Git blob selection, small run-input references and owned-source.patch.gz, reverse-apply checked. No source tree archive, media/module copies, old-run deletion, peer WIP edits, local Chrome, private guest writes/handler calls, configuration/backend/ binding/clock/seed changes, public deployment or subworkers occurred. Disk samples remained above 1.5 GB; there is no continuous global-floor claim.

Next requires separate authorization: continue only from the exact new pair, freshly review health/BAT, and repair the host quit admission/timeout integration before claiming successful repaired-driver close/profile cleanup. No further gameplay is requested here. Full27, parity, audio and FPS remain open. RELEASE and STOP.