Neverhood cold child scheduling trial, 2026-11-17

Evidence: scratch/runs/20261117-neverhood-cold-start-scheduler-remedy-room/result.json. Runtime unchanged. The earlier 9fc96756a original nursery/wake consequence remains valid; this finding does not replace it. Authored game-drawn Quit and native exit proofs from a853b016b remain valid and were reverified by hash. No fresh original was launched: the conditional remedy-proof gate was not met.

Five cold authored trials reconstruct the same sealed source pool 1cc971fb2d223e0bba6e20c21b094d70f219d107a2ed0d516a7c250d9a35b471 and retained module 5a4178950559d4f677d86aa75403908abee343493805f01a0b8aae8f1df2429a. The installed shortcut uses the populated 65-record/58-file state. Only the SETUP95 and NHC program records are replaced by authored programs; the other 56 installed files, registry and pending WININIT remain intact. Each child freshly hashes all 651,438,080 CD bytes and joins the hash worker before media adoption. Source, live provider, process/listener identity, one-use tickets, authenticated desktop hello and USER readiness semantics are unchanged.

The candidate restricted only newly spawned native fixture processes and their descendants to CPUs 1 and 2, selected once from a two-second per-CPU sample. Normal affinity was CPUs 0-3. The sample order was baseline1, affinity1, affinity2, baseline2, baseline3. Baselines 1 and 2 and affinity1 missed the unchanged 3,000 ms aggregate readiness/capture deadline. Affinity1 published ready at 3,050 ms; affinity2 published at 1,864 ms and completed actual native child action, independent UP and Quit with native exit. Baseline2 published ready at 3,510 ms; baseline3 at 3,021 ms. Baseline1 reached listen() after 3,111 ms of child bootstrap time, without an observed listening callback before cleanup. Missing callbacks are not credited as completed binding/readiness.

The main-thread bootstrap-to-bind/call observations separate CPU execution, runnable scheduler wait and off-runqueue time, in milliseconds:

Trial CPU Runnable wait Off-runqueue residual Elapsed
baseline1, through listen call 530 1479 1103 3112
affinity1, through binding 430 1048 612 2090
affinity2, through binding 382 406 251 1038
baseline2, through binding 498 1885 302 2685
baseline3, through binding 547 802 421 1770

The fresh child CD scan used 445-537 ms CPU and 131-1483 ms runnable wait. Its measured off-runqueue residual was approximately zero to 9 ms, including sampling skew. Thus scan wall time is not evidence of equivalent physical-disk blocking. The main thread's consumePreparation barrier lasted 193-1037 ms; 189-1007 ms was off-runqueue time while waiting for the worker. The trace separately records actual source requires, WebAssembly compile/instantiate, constructor validation, restore/adoption and listen callbacks. Require times are inclusive/nested and are not added together. Per-thread counters cover the native main, desktop transport and media workers. Residual time is monotonic elapsed time minus schedstat execution and runnable wait; it includes blocked asynchronous barriers and has small sampling skew. Passive overhead is retained.

All allowed CPUs were 97-99% busy in the selection sample. Recorded ancestor cgroups had cpu.max=max and no throttling increments; CPU pressure was high. No competing process or host policy is identified as the cause. Affinity restriction did not provide a reliable deadline result: only one of two restricted runs passed. The strong host-load variation and incomplete planned sample prevent a causal improvement claim. No priority change, privileged scheduling request, denied operation, competitor intervention or cgroup policy write was attempted. No new redundant serial full scan was found; the existing hash and desktop transport already overlap host construction. Removing their required joins would change validation/readiness semantics and was rejected.

The driver stopped on its disk-floor assertion after baseline3 had completed native cleanup and harness communicate(), before it persisted its own captured stdout/stderr and returncode. Native, child, controller, release and watchdog actual waits with both EOFs had already been written. The missing harness returncode is explicitly unknown. The planned sixth trial never started. Post-run space recovered above the floor; no foreign files were removed. This interrupted sample is reported as such, not as six successful trials.

The completed records include 45 actual waits, all with both EOFs; recorded owned PIDs are absent. All timed-out trials latch terminal UNKNOWN and attest no later requests. Prior delay/cancel/wrong-port/foreign-listener/late-ready and game-drawn Quit proofs are reused by exact hash; no broad replay or new canonical build is claimed. Two selected whole authored dialog images were personally reviewed. No original input, room action or original Quit is claimed here, and the conditional one-original allowance remains unused. No box, source archive, persistent media copy or cachelink change was made.

The remaining blocker is reliable completion under the fixed deadline with host scheduler contention. This specific affinity candidate is insufficient; a remedy needs repeatable matched success with the same full-byte validation and actual native readiness before an original run is justified.