Warcraft III producer scheduling investigation, 20 November 2026

Run scratch/runs/20261120-warcraft3-producer-observation-scheduling-cause did not establish or repair the post-DOWN freshness cause. No production source changed. Single Player, Prologue, Thrall movement, objectives and normal Quit remain unproved. Do not use this run as scheduling qualification.

The preceding authentic-keyboard run established S from the original MainMenu.fdf and recorded applied DOWN followed by fresh=false at 253 ms, with fourteen other predicates true. Its status trace also held presentation sequence 113 throughout the pending interval. That observation alone cannot distinguish delayed callback delivery from a producer blocked elsewhere.

The bridge and adapter were read alongside the actual pinned guest-rpc, guest-worker and guest-thread-host paths. The adapter polls status after socket publication with a 5 ms delay between replies. The Worker sends check_input context synchronously; native-message-wait samples carry their producer clocks and endpoint identity asynchronously. Neither path justifies refreshing cached timestamps or weakening freshness. No proposed asynchronous status change was installed: callback starvation remained a hypothesis.

A first setup scope stopped before any browser or endpoint request. Its inherited board watcher incorrectly began at recipe preparation and included a shared disk-floor sample preceding launch. The run-local watcher was corrected to the actual process-launch timestamp for a separate scope, original2. The first scope stayed closed. Existing live filesystem checks were retained.

Original2 loaded the original with both CPU/render modules SHA-256 10bdae2c56c79eeb5d0b4ebbc602c8a233e36351380da274a70f7f5e25363564 and CRT 887eb5ce93edb7192ca3e9220f07f9ca0f94db02af5862ebcbdfcb852db99fd1, software GL GPS and a private 800 by 600 display. Whole host-consent pixels were personally reviewed. The prearmed independent trusted host UP returned. No guest gesture was prepared or dispatched.

The author's run-local passive diagnostic wrapper was faulty: after calling the original dequeued callback, it tried to JSON-serialize the renderer event, whose references include a circular wine/process object. At 16:07:41.163 UTC the real check_input import reported that diagnostic TypeError. The guest stopped. Native bootstrap then observed running=false/stopped=true and closed at .281. This is an author instrumentation failure, not evidence of the requested production scheduling defect. There were 73 ordinary requests and zero after closure. No replay, later endpoint inspection, or S/hold retry ran.

The retained startup-only scalar trace shows USER observations separated by 539, 950, 544, 799 and 1796 ms, while its page heartbeat continued. This is limited to startup before binding; it does not identify the previous post-DOWN cause. Console records were capped at 4000 characters, truncating timing tails; trace-prefix-analysis.json explicitly retains only complete scalar prefixes.

All 31 existing focused checks passed and 36 historical media/module/native proof references were hash checked before browser creation. No unchanged fixture calibration was repeated. Three run-local offline controls reproduce the diagnostic circular-event failure and verify its corrected scalar-only, exception-isolated wrapper, including logger failure, original exception identity and heartbeat cleanup. That corrected wrapper was not executed against the original and is not a production fix. The initial offline syntax and close-logger failures are retained with the final passing log.

Both scopes are permanently closed. All 19 supervised original2 processes and the setup process are nonlive; browser/display/host-helper actual waits and EOF are retained. Independent finite runtime/cleanup limits were 245/260 seconds. No test box was created. No Python, Wine, workers or deployment was used. The inherited shared bridge WIP and frozen Hex baseline remain unchanged. Only this note is committed; no production change warrants a new full build. Recovery, exact Git/diffs, source hashes, media references and final result.json record these limits. A future scheduling repair still needs a nonintrusive actual post-DOWN trace, a named production cause, old-fail/new-pass controls, and affected-original verification; this attempt supplies none of that credit.