Half-Life 2 public installer lifecycle investigation

Evidence: scratch/runs/20261117-halflife2-public-installer-lifecycle-remedy/.

The latest original remains permanently closed. This task made no request to its endpoint, launched no original, and changed no runtime/controller code. Installation and gameplay remain unproved. The conditional fresh original allowance was not consumed because a causal remedy was not established.

The current published controller is 71d938888b6adeb8b0c31c4f2fc737786c8c133f. Its reviewed helper and scene bytes match the latest original's recorded source inventory. Its restored adapter matches that commit. The retained module is 928f8bd42c76b6435ec59d8b289b5a40a28da8b597a6e98e378b467311f11abb. The media remains at scratch/media/half-life-2-demo-original.exe, 787,701,212 bytes, SHA-256 24205fa8cb6bfab70a82dd713f0addc7cfb0b3c4df40441ebea0320dcb3b441c. No media copy or payload expansion was performed.

The retained public log establishes more than the passive poll summary: the hidden overlapped main 0x10001 was created, CreateWindowExA returned its HWND, and the application later dispatched WM_SIZE and WM_TIMER. All 140 passive startup observations refused readiness, with hidden main and null focus, over a 14.484-second observed span inside the fixed 15-second startup operation. The native process was subsequently host-cleaned with an actual wait and both EOFs. This does not establish native UI Quit or explain the missing visible focused scene. In particular, a blocked CreateWindow call and the earlier missing helper join are not supported explanations for this run.

The earlier successful original startup in the age-input-latency run used the same module and the same complete sealed source inventory except for tools/halflife2-reviewed-next.py. Its readiness completed in 3.360 seconds. The startup function's normalized AST is identical at 40fd97e98 and 71d938888. The newer pre-bind transport journal is a real helper difference; these comparisons do not establish or exclude a scheduling effect from it. The latest authored recovery/Cancel/Quit proof has an entirely identical sealed source map and module to the failed original. Its native wait/EOFs and selected whole-image hashes were checked. Both selected images were viewed in this task; they show the authored agreement/age controls, not original installer progress.

Existing authored modeless transition, changing-main/generation, native focus, independent release, and Cancel/Quit proofs are pinned by exact path and hash in inherited-proof-pins.json. They retain their authored scope. They do not prove the original's agreement-to-next-page lifecycle or eventual installation.

The shared src/09b-dispatch.wat creation-time WM_SIZE repair and test/test-create-window-failure.js belong to codex-dxsdk-rm; neither was edited or adopted here. An overlap inquiry was appended to the board. The retained original also dispatched a queued WM_SIZE after CreateWindow returned. Whether that notification affected its subsequent startup is unproved. The public CreateWindowEx documentation describes creation messages; the WM_SIZE documentation ties size/move messages to DefWindowProc handling WM_WINDOWPOSCHANGED. Those contracts alone do not determine this original callback's consequences.

The concrete remaining blocker is failure to acquire a visible, focused ordinary USER startup scene before the unchanged deadline, without public evidence identifying a causal API/runtime/controller error. Loosening readiness, assigning focus, extending the clock, or substituting the other owner's WM_SIZE change would not be an established remedy. No such change or unchanged original retry was made. Next work needs an authored reproduction of this hidden-main message lifecycle that identifies a documented API violation and a tested remedy before using the conditional affected-installer allowance. The original UNKNOWN cannot be queried, rebound, or replayed to obtain that evidence.

Fresh launch-order/terminal-original guards, receipt-arrival clock checks, and the seven controller supervision regressions were run under a prearmed independent deadline watchdog. Their actual waits, watchdog waits, both pipe EOFs, and final outcomes are recorded in the run. No new native runtime/module was built or tested in this task; this is a documentation-only publication. The exact source commit, relevant shared diff, proof pins, media hash, selected whole-image pins and final result are retained without a source archive. No workers, boxes, deployment, state/registry/media edits, private-memory inspection, proprietary disassembly, or held Sims artifacts were used. Box ID and expiry are null; state is NOT_CREATED and stop receipt is not applicable.