Neverhood sealed native preparation and remaining room attempt
Run: scratch/runs/20261117-neverhood-current-registry-controls-room-progress.
No original installer or installed-game invocation has been started by this task.
The latest VB full native proof at e03ea0973 applies to the tested f1dd81 module
and unchanged native source graph; GL's complete-browser-capture UNKNOWN is a
separate route and is not used as a CLI admission condition.
A real helper fault was reproduced during concurrent main integration. Node
resolved relative modules against the mutable checkout before the sealed source
extension hook ran. Removing lib/glide-host.js from the live checkout therefore
broke an intact, hash-verified older preparation that still imported it. The
indexed loader now resolves inventoried relative and absolute JS/CJS/JSON modules
before consulting disk, verifies their bodies, and retains the existing refusal
of unsealed executable inputs. An absent physical fixture fails under the old
published loader and succeeds with the repair. Main, actual Node Worker, and OS
child JS/JSON loads pass without creating fixture files in the repository.
The Neverhood helper now retains its tested module before execution, uses the published process-parent supervisor, and enforces the assigned 900-second native and 920-second independent limits after preparation. It pins a published main ancestor and checks later graph membership instead of imposing a global HEAD freeze. It retains all 257 recursive CD member hashes before the clock. Native BUTTON receipts are checked against the actual target owner; original normal exit requires the invocation-bound guest exit, independent physical UP, and real wait. Game intents capture the full frame after release even when window metadata does not change. Menu transitions also retain a bounded post-UP full capture. A combined explicit review/intent command avoids a second Python startup without changing review freshness or any input deadline.
The authored controller's native-wait fixture also needed atomic publication:
its former direct write could expose an empty JSON file to the exit observer.
All eight final native controller cases pass, including game-scope capture,
menu transition, real normal exit, repaint, and focus/owner/semantic/stale/caption
refusals. The final outer process and watchdogs exit zero with actual waits and
both output EOFs. The registry diagnostic again passes 92 native cases including
main, CPUWorker, and actual OS-child persistence. The canonical locked build
passes (canonical-resolver-wait.json). The selected authored screenshot shows
only the second generated dialog; it is not Neverhood gameplay.
Publication is pending the coordinator's explicitly announced shared Git
critical section. The owned repair is preserved as owned-repair.patch; no
foreign integration files, index entries, source archives, or media copies were
created or changed by this task. After publication, the authorized original
attempt must use a complete published pinned graph and its materially matching
native proof. An incoming changed native graph needs appropriate proof; the old
f1dd81 proof must not be claimed for a different module. No old UNKNOWN endpoint
may be queried, rebound, or replayed. Authentic prior installation recovery is
verified, but its missing path values remain absent and have not been fabricated.
Published-graph successor
Successor run: scratch/runs/20261117-neverhood-published-loader-normal-install-room.
The coordinator released the Git critical section after publishing ae70a5473.
The successor owns the six released paths and excludes the remaining foreign
working overlays. Its immutable graph includes the published graphics, registry,
control, process-supervisor, installed-loader and VFS fixes through 68b9bdac4.
The old f1dd81 module does not qualify this changed graph.
The helper now runs the actual native CLI USER ownership, modal, nested, exit, shutdown and pointer fixtures when an earlier proof's material sources/module no longer match. Matching completed checks can be reused by module and material source hashes. The borrowed VB fixture controller uses a map-shaped source manifest; it receives an isolated adapter while Neverhood keeps the original list-shaped authenticated inventory. Python fixture controllers and watchdogs inherit the same sealed source pool through both Node and Python child launches; Python imports bypass mutable bytecode caches. Missing physical Python modules, imports and child scripts are tested alongside JS/JSON main/Worker/OS-child resolution and unknown-source refusal. This adapter changes authored host fixtures only, without modifying original media or guest application state.
The fresh causal loader probe reproduces old-loader failure and repaired-loader success against identical absent-on-disk JS/JSON fixture bytes. The successor also verifies all 92 native registry cases against its prepared graph. Original installer/game invocation counts and final native qualification/publication are recorded in the successor result, which is written after execution.
The successor published the loader/fixture repair as 284c4836f after the
locked build, 92 native registry cases, actual main/CPUWorker/OS-child source and
module inheritance, complete native modal/owner/refusal/focus/nested/exit/shutdown
and pointer proof, and all eight Neverhood controller cases passed. The changed
native graph compiled to SHA-256
5a4178950559d4f677d86aa75403908abee343493805f01a0b8aae8f1df2429a.
The published preparation matched those material source hashes and repeated the
own helper checks before rehashing all 257 CD members and arming the fixed
900/920-second original deadlines.
Exactly one fresh ordinary CD SETUP95.EXE session then ran. The complete fresh
launcher was personally reviewed. Install was acknowledged at the actual native
BUTTON dispatch return in 46.9 ms; the independent physical UP acknowledged in
163.6 ms. The first post-UP capture was refused with modal.visible: the runtime
had published the newly created directory dialog before its ShowWindow finished.
The controller permanently latched UNKNOWN. There were no later HTTP requests,
rebinds, input replays or second originals. Native process, watchdog and outer
supervisor all actually waited zero with both EOFs and no remaining owned scope.
The native log explicitly records SIGTERM orderly shutdown; this is host cleanup,
not native Quit or a genuinely successful normal installer exit. Installation,
Play, room entry and object actions remain unproved, and the conditional separate
installed invocation is not admissible.
The remaining causal helper defect is repaired by waiting on public modal visibility after UP before requesting the immutable capture. A changed window list alone no longer counts as a presented scene. The existing native capture validator remains authoritative, and all observation stays inside the original 3000 ms bound. The first pending public snapshot is retained without capturing an incomplete frame. Actual authored fixtures hide the newly created native modal through its native style and the production renderer's ShowWindow path, then either show it after 650 ms or leave it hidden. The delayed positive retains a complete visible second dialog in 1007.8 ms; the permanent-hidden case keeps physical UP, performs no invalid capture, and reaches permanent UNKNOWN with zero later HTTP. Ordinary menu and game-scope captures also regress successfully. No original endpoint was reopened to test this repair. Final publication, exact waits, selected images and artifact hashes are in the successor result.