Neverhood separate native startup readiness

Evidence: scratch/runs/20261117-neverhood-visible-startup-normal-install-room/.

The wrapper now completes a passive native-startup-readiness preflight before starting the initial capture clock. It uses the published ordinary USER owner, actor, main/modal, generation and focus checks from 63d7b0d0e. A listener or native PID alone cannot qualify readiness. The preflight has its own bounded 30-second observation, also limited by the existing gameplay cutoff; it never extends the native invocation or input deadlines. Failure permanently closes UNKNOWN, including all later queries, captures and binds.

The first capture starts after readiness. Its native startedAt/completedAt values and three-second deadline remain unchanged. Review, bind, intent receipt and independent physical release code is byte-identical to 97670c226; preparation verifies equality after removing only the new atomic pending-presentation observation branch.

The focused sealed native run uses the retained core module SHA-256 5a4178950559d4f677d86aa75403908abee343493805f01a0b8aae8f1df2429a, published main core ef51a021, explicit owned patches, and the two published startup components. An authored x86 GetTickCount loop delays actual ShowWindow by 3.8 seconds. Wrapper readiness lasts 3.471 seconds with zero capture, bind or input before readiness; subsequent captures retain their original native timestamps and bounds. Real BUTTON consumption, independent physical UP and ordinary native exit pass. A never-visible window produces UNKNOWN with zero capture/bind/input and zero later HTTP requests. Controller units, participant diagnostics and the locked canonical build pass. Actual native and watchdog waits with both pipe EOFs are retained.

The first generated test incorrectly used Sleep, which the uncapped CLI intentionally charges to batch time. Its duration assertion correctly failed. That preparation error is retained and was replaced with the ordinary real GetTickCount loop; no runtime clock, original executable or media was edited. No original allowance was consumed by either generated test.

Existing lifecycle, participant, actor and input proofs are reused by exact component. The only native component differences are the passive command and published diagnostic helper. Removing the startup addition reproduces the published diagnostic helper exactly; removing the startup command and atomic presentation branch reproduces the prior runner exactly. Full unchanged protocol suites were not repeated. Fixture controllers retain two required Python pins and a hash of the complete inventory instead of duplicating thousands of source records.

The first genuine installer used published wrapper c1ec636f6. It completed passive startup readiness, produced a personally reviewed complete launcher image and consumed Install with a genuine native-button-dispatch-return receipt and independent physical UP. The subsequent public snapshot qualified presentation, but the following capture saw a newly created hidden modal and correctly refused modal.visible. UNKNOWN permanently closed the endpoint: seven requests, one bind, one DOWN and one physical release; no later HTTP. This is an actual snapshot-to-capture wrapper race, not a blank PNG inference.

The independent native watchdog subsequently observed disk-floor failure at 1,494,720,512 free bytes and delivered SIGTERM. Its deadline had not elapsed. The operator's cancellation request came later and was not the shutdown cause. Native, outer, persistent owner and all watchdogs have actual wait and both EOF receipts, with no remaining scope. Exit zero is host cleanup, not native Quit.

The transition fix adds native-presentation-capture: it tests public USER and ordinary actor readiness and invokes the existing capture synchronously when ready. Pending readiness allocates no capture, input owner or PNG. There is no await, pump or renderer operation between the readiness check and capture. The existing capture validator, native timestamps and three-second deadline remain unchanged. A refused capture is never retried. Python polls pending readiness only inside its existing initial capture or aggregate intent deadline.

The focused authored native fixture reproduces the same modal.visible refusal during real WM_INITDIALOG. The atomic path observes pending with no PNG until the real dialog is presented, then produces a valid native capture. A real MessageBox-to-delayed-DialogBox transition passes the full wrapper, genuine BUTTON receipt, independent UP and final ordinary exit: the two finite intents complete in 1,300 ms and 461 ms. Units, participant diagnostics and the locked build pass. The first transition test used a MessageBox-only export to locate the DialogBox button and correctly failed its test assertion; that authored lookup was replaced by the public USER snapshot. Neither preparation fault consumed an original invocation.

Published e16a4713b contains the faithfully tested atomic fix. The conditional second ordinary installer used that publication and the identical retained module. Complete unchanged inventories of all 257 CD members were verified before preparation, immediately before execution, and after actual shutdown. The passive startup preflight preceded its first capture. Three personally reviewed finite actions followed the normal route: Install, default directory OK, and Create Directory OK. Each received genuine native BUTTON consumption and independent physical UP within the unchanged deadlines. Their aggregate intent times were 546 ms, 776 ms, and 1,598 ms respectively.

The complete retained second/original/installation-next.png visibly says “Installation of The Neverhood was successful.” The authentic output has 65 overlay records and 58 files, including NHC.EXE, SMACKW32.DLL, HD.BLB, the installed SETUP95.EXE, readme and generated shortcuts. The product registry persisted Installed=1, Data Drive, Installed Directory and Run File. The normal shortcut targets installed SETUP95.EXE with no arguments or working directory; its target bytes equal the unchanged CD launcher. All 58 output files were hashed. Fifty-one byte-identical CD copies are recoverable by ISO path, member, size and SHA-256; seven unique generated outputs are retained. No guest files, registry values, scripts, pending renames or launch paths were synthesized.

Before the success OK intent could check native identity, the independent watchdog observed 1,494,151,168 free bytes, below the 1.5 GB floor, and sent SIGTERM. Its 920-second deadline had not elapsed. Actual native wait/EOF preceded the attempted intent. Missing /proc/843272/stat permanently latched UNKNOWN before bind or DOWN; no later HTTP occurred. One subsequent local capture command was an operator error: the permanent latch refused it before identity checking or HTTP, and no capture was created. It is not a replay or successful retry. Both original endpoints remain permanently closed.

Native, adapter, outer supervisor, persistent worker owner and all independent watchdogs have actual wait and both EOF receipts, with empty remaining scope. Native exit zero was host disk-floor cleanup, not ordinary native Quit. Disk space was subsequently restored by the separate storage worker; this does not revive either endpoint or authorize a third installer.

Both installer allowances are spent. Installation success is visibly proved; acknowledging its completion dialog, normal Play handoff, an installed room, and a visible object action remain unproved. The separate installed-game allowance was not used: no reviewed normal Play handoff was obtained before shutdown. The authentic installed files, registry, shortcut and pending WININIT renames are preserved for a separately qualified normal launch; no manual boot transition was applied. No boxes or workers were added. Box ID and expiry are null; stop receipt is not applicable.