Matched observed timing trial

The packet-aware candidate must pass the complete, separately frozen 67-control qualification before this study starts. The candidate and the qualification result are hash-checked before preparation and every observed identity.

The target is fixed: fd7b81e48a4122b149c4925b, under profiled removal and the zero-background globe. Its native timing checks show the local prediction failing across a nearby packet boundary.

Four optimizer starts, no retries:

• One primary at the prior matched trial's failed primary endpoint, at most 200 native fitting calls, including the original fresh final checks.
• One predetermined repeat, seeded by shifting each normalized primary coordinate by 0.01 toward the opposite sign, clipped to existing bounds; at most 200 calls.
• The same two completed controls, each at most 40 fitting calls.
• Two native diagnostic calls per identity. Total ceiling: 488 native calls.

Use one worker with four total numerical threads. All neighbor-interval predictions and their native Jacobians count as fitting calls; no uncharged probes or restarts. At the diagnostic's measured speed, the ceiling is roughly two to three hours; actual completion may be earlier. Do not expand the allowance.

Reuse only the five hashed passing primary companions for this stretch from the previous matched trial. Preserve original measurements, nominal integration clock, initialization, receiver weights, calibration bounds and processing hypotheses. The original fresh stationarity, objective-change and physical-position gates stay unchanged. The only candidate change is how timing steps are proposed after a rejected full bounded step. Local interval boxes are numerical subproblems, not new physical restrictions on the allowed time offset.

Require both controls, the new primary, its repeat agreement and all five reused companions before reporting a complete comparison. Keep saved failures and earlier publications unchanged. A failed original gate remains a failure even at a genuine packet-boundary minimum. No calibrated significance or wider rollout follows.

Freeze all sources, inputs, environment and original hashes before launch. Check exclusive worker locks and completed dependencies. Pre-call journals retain failed and interrupted starts; atomic hashed outputs support continuation of unstarted identities only. Completed studies cannot restart. Publication performs no fits.