[Denny-Lin]

Longer frozen-model comparisons now retain 240 complete trials: four stable workloads, four new burst-pattern families, full/lazy shadow sessions, and online utility against fixed, EWMA and original-online controls. Add deterministic generators, explicit sample-buffer prefaulting, restoration guards, paired utility analysis and compact raw-counter evidence. Kernel/model source, parameters, watchdog thresholds and model defaults remain unchanged.

EWMA/carry reduces many MAE estimates but is not a universal replacement. Full alternating bursts regress 17.82% in every paired repetition; the original model already beats same-episode EWMA on that family. Shortburst online wake p99 worsens from medians 0.995 to 1.364 ms (paired +35.65% [18.63,39.36]); mixed wake p99 also worsens. There is no consistent throughput advantage. Preserve all controls, metrics and uncertainty; keep original bias/trunc defaults.

The initial lazy session completed 16 trials and failed its seventeenth with a real runnable-task watchdog stall (1.113 seconds). Linux scheduling and original preemption were restored. Its remaining 23 trials were not run. Preserve the full failed session and kernel dump; a complete unchanged 40-trial retry passed but does not resolve or erase the original reliability failure. Five traced diagnostic runs did not reproduce it. Proximity to host display wake is only a timing observation, not a proven cause.

The planned 40-trial observer ablation did not fit the 05:35 UTC research reserve and was not started. No model retuning or outlier removal occurred. Independently randomized environments in grouped time order do not establish a causal preemption effect; do not pool older binary/window/observer studies.

Validation: Linux build and eight new-profile verifier smoke trials; host profile boundary/sanitizer checks; restoration, compact-counter and paired-utility fixtures; exact reproduction of every published model score and all utility comparisons from compact data; documentation/local-link checks. Full raw archives remain durably outside Git with SHA256, and only verified owned guest copies were cleared. One ARM64 VM and custom workloads establish neither hardware nor general Linux superiority.