LBreakout: retained startup state reaches finite theme grayscale

Evidence: scratch/runs/20261110-lbreakout-bounded-startup-state-0410/ (ignored). One fresh original startup ran from 6351efbdd with module SHA-256 4f2c69a1528e01a9ab5127201d5e8a222da2a35236377a1351f444f55fa5a7d9. It failed the unchanged two-second completion deadline before any first image. There was no game input, menu/gameplay acceptance, retry or second native run.

The retained near-deadline state identifies original pre-menu theme grayscale work. It does not demonstrate a missed partial-update/menu return. No runtime repair is supported by this run, and the startup gate remains unchanged.

Native observation and its limits

Startup armed at 2026-11-10T04:13:48.079Z; its deadline was 04:13:50.079Z. The last existing snapshot sampled at 04:13:50.075Z, four milliseconds before that deadline. The caller recorded its first failure at 04:13:50.086Z. The sample is 11 ms old at failure; its PC must not be reported as a synchronous failure-time PC.

That sample has PC 0x451ae0, the original EXE's _memcpy import jump. The stack's current return is 0x426c59, immediately after the first memcpy in _stk_surface_gray. EBP is 0x074ffb68, ESP 0x074ffb2c. The retained EBP frame also contains return 0x418687, immediately after _theme_load calls _stk_surface_gray at 0x418682. Original hash-pinned instruction bytes and disassembly establish both calls.

The frame describes surface 0x01385eec, region (0,0,640,480), shift zero, current column 240 and row 168. Row offset is 0x69000; column byte offset is 960. The current four-byte copy goes from 0x7e6093c4 into local 0x074ffb54 (EBP−0x14). These values consistently imply pitch 2560 and pixel base 0x7e5a0004. Those are inferences from original local arithmetic; the retained sample does not contain this grayscale surface's header or pixels.

There are 107,760 preceding pixels and 199,440 remaining pixels including the current one in this bounded loop. This is a coordinate calculation from one sample, not a measured native rate or proof of the eventual native return. The target is distinct from the retained screen surface 0x00c14abc, whose actual header reports 640×480, pitch 2560, pixels 0x7eed0004. No fillSurface/fillRect is present: this sample is outside the original SDL fill loop. There is no native remaining-fill counter to invent.

Menu pointer and dirty count are both zero. The only observer event is its arm; neither the watched SDL full-update return 0x426026 nor subsequent menu-manager return 0x410849 was recorded. Together with the original theme caller, this supports unfinished pre-menu initialization at sample time. The partial-update path's existence remains established by earlier original controls, but this attempt supplies no evidence that accepting its other return would repair startup. The intervening 11 ms and eventual startup remain unknown. Earlier readiness page/identity slices are separate observations and are not merged into this sample or relabeled as failure-time state.

Offline change and validation

classify_startup_failure in tools/lbreakout-case-inspection.py decodes only already-retained state. It checks the original EXE hash, PC, two caller returns, bounded frame layout, copy arguments and coordinate consistency. It reports sample age separately from the completion deadline and failure time, and returns unknown for unsupported or inconsistent frames. It issues no live protocol requests and changes no guest or observer behavior.

tools/lbreakout-bounded-startup-state.test.cjs checks the actual original instruction anchors and runs original grayscale on an isolated original SDL 640×480 surface in compiled Wasm. Successive original read copies advance from column zero to one, source advancing four bytes. Both per-pixel read and write copies execute. The original function returns with its cdecl stack restored; all 307,200 pixels match the original grayscale arithmetic. This control validates finite code and frame interpretation, without claiming native execution time or native completion. Eleven corrupted/missing-frame controls reject unsupported classification. The native sample decodes to column 240, row 168 with the correct 11 ms age.

Before allocation, the emitted attached-Worker handler, actual host init, original Worker installation, actual installed instance, real Wasm, original SDL fill timeout and complete original menu suffix caller controls passed. The generated caller was composed exactly as startup_workflow_source(pause_workflow_source(...)), with the established first-image presentation template. Duplicate composition was refused. The 6351efbdd preload retained PC/registers/24 stack words/surface/menu/dirty state. The accepted 4fc62b7d7 fixture evidence was hash-verified and reused, with no native fixture repeat. Explicit level start, separate P pause/P resume and Space controls, and the 080064edf unobserved/combined-play refusals persist.

Focused startup observer/full-caller, menu, paddle, pause, level-start, inspection, cached-failure/composition and installed-Worker tests passed, along with locked builds. Preparation failures are retained: a startup test correctly refused an already-composed input, another lacked its input helpers, and the first grayscale test compared a read copy with its writeback. Corrected offline controls pass; none caused a native retry. Shared CRT, CPU/JIT, renderer/startup providers and other titles were read-only. No native guest writes/artificial calls, clock/tier/budget/deadline changes, source archives, evidence media copies, local Chromium or nested workers were used.

Cancellation, evidence custody and stop

First browser failure at 04:13:50.086Z invoked owning Chromium SIGTERM. The independent remote monitor recorded cancellation at 04:13:50.102342Z, signalled Chromium 5889 and launcher 5880, and retained the actual launcher parent's wait() result −15 plus separate stdout and stderr EOF records. No subsequent game protocol, capture, input, quit or extension was issued. The Chromium exit JSON was not finalized before launcher cancellation; no separate Chromium-parent wait claim is made.

All 34 remote evidence references were recovered as 32 unique artifacts and independently reread for SHA-256 and size. All 824 runtime pins were unchanged. There were zero PNG and zero gzip artifacts; no image or gzip validation is claimed. The verification checks CRC/decompression for any such artifacts present. Runtime source identity is the recorded Git commit plus relevant binary diffs, including untracked inputs; media is referenced by existing path and SHA-256. Unique logs and the actual loaded module were retained.

Only one box was allocated: bx_cm4px2vw, UUID 7109223d-46a3-4ef6-8922-2bce9c3df081. Its immutable independent 25-minute watchdog was observable before allocation. Initial API expiry was 2026-11-10T04:38:08.768Z; the original watchdog deadline was 04:38:09.023143Z. After artifact verification, the sole delegate requested unforced stop. Receipt stop_3dee994975a17bfe completed at 04:14:21.648Z. Final API state is archived, expiry null. The independent watchdog's actual parent wait returned zero with both stream EOFs, and its deadline digest remained unchanged. No inherited or other box was touched.

Next useful work is a separately scoped cost investigation of original pre-menu theme grayscale and its copy path. The current evidence supports bounded initialization work near the deadline; it supplies neither a runtime fix nor visible gameplay acceptance. This single-task worker stops after publishing the offline decoder, regression control and report.