TR2 startup presentation repair — 2026-11-08

The production single-app maximizer could raise the blank main window over the startup property sheet. It now waits while this instance has a visible dialog. A small actual Chrome fixture reproduces the covering behavior with the selected renderer and shows the dialog remaining visible after the guard. No original, game, remote allocation or input ran in this preparation.

Evidence is scratch/runs/20261108-tr2-startup-presentation-1750/ (R). source-head.txt, relevant.diff, and the small owned new-file/helper patches record recovery inputs. Git is the source archive. Existing media remain at their selected paths/hashes; no source archive, media copy or peer WIP is used.

Concrete difference and causal path

Both originals requested the same startup route, apart from the HTTP port: index.html?debug&threads&app=tomb_raider_2_demo&gl-renderer=webgl&d3d-renderer=webgl. The normal drivers differ only in observer activation inside the later Control hold. That code was never reached in 1743. Selected CPU bytes and startup page/renderer/shell bytes are unchanged between the two originals.

1717's actual shell console reports single-app=false, viewport 945×889, screen 1920×1080. Its reviewed startup shows the full desktop and readable Graphics options/OK. 1743 reports single-app=true, viewport 780×437, screen 800×600, coarse pointer false. Its startup log additionally records Single-app mode: maximized hwnd=0x10001 400x300, absent from 1717. The headless browser version/launch environment differs as well. Neither viewport size nor its appended SwiftShader flags alone diagnoses the pixels.

The causal code is maximizeForSingleApp in lib/browser-shell.js. Its 50 ms timer already detects visible dialogs and avoids charging those ticks against its wait limit, but previously continued to select the blank main window and send WM_SYSCOMMAND/SC_MAXIMIZE. Host sys_command routes that through renderer.showWindow(hwnd, 3), which expands and raises the main window. Owner grouping cannot protect an unowned property sheet. Read-only disassembly of the existing exact TOMB2.EXE confirms the original constructs a zeroed 40-byte PROPSHEETHEADER at stack offset esp+0xc, leaves its hwndParent field at esp+0x14 zero, and passes that header to PropertySheetA at 0x452993 (IAT 0x463018). The relevant 0x452800–0x452998 instructions are retained in original-property-header-disassembly.txt. This is static inspection, not guest execution. 1743's inspection also records visible Graphics child 65545 under top-level root 65538. Its parentHwnd:0 alone would not establish the Win32 owner, which is why the executable's header construction matters. A dialog can therefore remain live and painted while the newly raised gray main covers it. A failure-free CPU at yield 7 is fully compatible with that presentation failure.

The production change returns from that timer tick while a dialog is visible, after the existing app-exit and exclusive-display cancellation checks. Once the dialog closes, ordinary maximization still occurs once. There are no new guest calls, timers, callback owners or restoration paths. The existing timer retains the same wine/renderer references and cancels on app exit or exclusive display takeover. The callback ownership census host-file hash was reviewed and updated; broader invocation ownership remains pending.

This is a tested causal stacking defect matching the measured startup path. The failed original did not retain numeric z-order or per-window backing pixels, so its final stack cannot be independently reconstructed. The authored counterfactual fixture is not a rerun or proof that the repaired original has started successfully. The next separately assigned original must review its actual dialog before ordinary Graphics Enter; no blind input is authorized.

Focused validation and selected closure

node tools/tr2-startup-presentation.test.cjs exercises the actual production timer body: 130 ticks with a live startup dialog send no maximize, then dialog closure permits one ordinary maximize. The old body fails that regression. Existing single-app mode, keepAspect and callback-census mutation tests pass. The regression also checks app-exit and exclusive-display cancellation while waiting, and that a different instance's dialog cannot block this instance.

The same test's optional --chrome R/fixture-final uses pinned production surface/renderer/input sources read directly from Git, the actual shell timer body, and authored completed GDI surfaces. It loads no PE or CPU module and creates no D3D context. Both cases keep identical canvas geometry and Chrome arguments. Before: root main 65537 becomes topmost, the live dialog is covered, and its OK pixel reads opaque gray [192,192,192,255]. After: dialog 65538 remains topmost and its OK pixel reads [0,128,0,255]. Personally reviewed fixture-final/before.png and after.png agree. This local Chrome viewport is 945×945, not the failed original's viewport; the counterfactual isolates stacking without requiring a viewport or GPU change. An initial fixture attempt named a nonexistent menu script, failed before painting, and closed Chrome with exit 0; that receipt is retained.

The locked full build result is retained in build-final-reviewed.log. The initial build stopped at the required callback-census review lock, before compilation.

R/runtime-pins.json is the explicit successor selection: 615 providers / 29 normal physical helpers, relocated with only owned run-path changes. Its predecessor is the immutable coverage-1733 manifest SHA-256 4cbcaf21b0ac26d969016af0ea8bf97a0e4a9df2465cafcc5ab5cb28f13505b0. Only lib/browser-shell.js changes among 586 non-helper providers; the other 585 remain exact. CPU remains d8facd0658b7f158c6e60bef2de0936471f438f8bf41e9240fd75635f1d319dc. Observer repair 8579c29a2, its 32 tests, independent release behavior and immutable stop/no-replay rules remain unchanged. The 85 unchanged coverage references / 446,299 bytes were independently rehashed and reused. Historical 614, predecessor 615 and selected coverage manifests were not edited.

The exact resolver checks every provider hash and the normal 29-helper closure. The actual driver preloader passes all 615 hashes and exits before Chrome or original launch. Full27, parity, audio, FPS, qualified original poll/Action and after-release consequence remain open. No backend, binding, input, clock, seed, private guest state or observer quota changed.

Closed local lifecycle and next step

Each small local Chrome fixture ran under /tmp/wine-runtime.lock with an observable independent 90-second timeout watchdog, armed before launch. Both browsers exited 0 with no signal, and their browser/watchdog PIDs are absent. No runtime remains. Free /home stayed above 1.5 GB; no RAM gate was used. Box ID/state/expiry/stop receipt: not applicable; no box created.

Next is one separately assigned original using the explicit successor selection and unchanged normal lifecycle/release rules. It must establish reviewed startup GDI pixels first. Owning D3D/WebGL2 authority and non-null initial raw/ARM are later, separate gates; prior Intel evidence cannot certify the fresh environment. This preparation worker releases its files and stops.