Native capture latency: measured boundary, no original retry

The authored SDL fixture did not reproduce the terminal capture delay. No production repair is justified by these measurements, so no fresh original was launched. Puzzle8 clear, tally, Puzzle9 and ordinary Quit remain OPEN. The previous endpoint remains permanently closed; it was never queried.

The retained operation began at 1794957923.10849 with its immutable three-second deadline. Post-release progress was saved at 1794957924.5131106. The controller then unconditionally sleeps 1.2 seconds before requesting its snapshot and PNG. That leaves at most 395.38 ms for both requests. The PNG file mtime is 1794957926.232329, after the deadline; mtime is not HTTP completion evidence. The fixed sleep demonstrably spends the operation budget, but neither capture request time nor ingress, queue, rendering, encoding or response times were recorded. Removing the sleep without an affected causal reproduction would claim more than this evidence supports. All existing limits remain unchanged.

The public live png: command is parsed and queued for scheduled batch work. The PNG action repaints, then canvasToPng invokes presentation and repaint, encodes the entire canvas and writes it synchronously. settleLiveInput resolves removed scheduled actions; the production control server then serializes and ends its response. The retained log establishes that the PNG action ran, but cannot partition request-to-dispatch or dispatch-to-response time. The potential duplicate repaint is source structure, not a measured original cause.

A small authored native consumer loaded the authentic SHA-pinned SDL.dll, called SDL_Init/SetVideoMode(640,480,16), FillRect and Flip, then captured the whole 1024x712 desktop using the production renderer, encoder, synchronous file write and HTTP control server. This is direct authored capture dispatch; it does not reproduce test/run.js's guest batch scheduling, dirty DirectDraw presentation hook, original scene, original module, or original load. It cannot exclude an original-only scheduling/presentation delay. No private game or SDL surface internals were read. The public SDL_Rect is an authored input.

Three initial successful observer-off/on/off runs took 44–86 ms end to end. The instrumented handler spent 9.7–15.6 ms repainting, 36.9–56.5 ms encoding, and 0.39–0.74 ms writing. A runnable relocated fixture repeated off/on at 40–150 ms end to end: repaint 8.2–22.8 ms, encode 28.2–90.4 ms and write 0.33–0.62 ms. The HTTP phase observer recorded response-finish for every instrumented request. These ranges do not establish zero observer overhead. All twenty captures decoded to identical full-screen pixels; only one selected whole image per successful process is retained, with no frame series. The selected image was personally reviewed as an SDL window of colored rectangles. Offline re-encoding of the retained original whole PNGs took 42–94 ms for the entry confirmation and 46–66 ms for the initial board, with identical decoded pixels. Offline timings do not measure the old runtime interval.

The first local helper failed because its copied VM setup lacked the current host's require binding. Adding createRequire(host.js) repaired the helper; this failure is retained separately and is not a game/runtime defect. The fixture remains a small run-specific input in evidence, not a production test. No tool/lib/source changes or canonical build claim are made by this task.

An independent observable watchdog was armed before each of six native child processes, including the failed helper. Every native and watchdog process was actually waited with stdout/stderr EOF; the enclosing helper and invoking tool also completed. Each successful SDL child called SDL_Quit and remained physically UP; no input was injected. Watchdogs were disarmed without firing. No box was created: ID null, state NOT_CREATED, expiry null, stop receipt N/A. Free /home remained above 1.5 GB. Foreign WIP and held investigations were preserved.

Evidence is scratch/runs/20261117-lmarbles-native-capture-latency-puzzle8/result.json. Git SHA, relevant working diffs, module/source/media hashes and selected named whole PNGs are recorded there; no source archive or media copies were made. The concrete blocker is missing original capture phase timing plus lack of an authored reproduction of that interval. Further work needs a causal authored reproduction before a tested repair can authorize the conditional original.