Next diagnostic: native two-sided timing checks

The software gate passed 43/43 controls. The saved-clock screen supports reliable side geometry for all eleven unfinished primaries and all eight unfinished repeats. Two unsupported endpoint geometries belong to already-passing smooth repeats; three passing smooth repeats overall remain in different timing basins. Their existing results must be retained rather than moved to a convenient knot.

The next work targets all six remaining stretches. It is a timing-location and stopping diagnostic, with every nuisance parameter held at its saved value during timing location, followed by the qualified full-nuisance stopping probes.

Prepare and qualify the adapter

Use the actual double-precision clock comparison, retaining separate receiver transitions. A nearest transition changes only one receiver epoch in this screen; it need not be the transition where the total timing gradient changes sign.

Enumerate the representable transitions within ±1 microsecond of each unfinished endpoint. Test the timing gradient outside this cluster and, when it brackets a negative-to-positive change, use a bounded search over the discrete transitions. Verify the chosen knot and both adjacent intervals with the clock oracle. This search estimates a time coordinate and must be recorded as an observed time-only optimization start, even though no joint nuisance fit is performed.

Software controls must cover multiple clustered transitions, nonmonotone interior gradients, absent brackets, derivative discontinuities, floating-point plateaus, finite search limits and failed nuisance stability. Retain all physical bounds and numerical tolerances. A failed bracket or an unchecked neighbor must abstain.

Proposed finite observed scope

• Nineteen time-only starts: all eleven unfinished primaries and their eight unfinished repeats. Reuse the three hashed passing repeats unchanged.
• Two existing smooth completed controls, checked without boundary relocation.
• Twenty-one identities / at most 26 native evaluations each / 546 total, four total numerical threads, no retries. No joint nuisance-fit starts.
• Limit timing location to at most twelve native gradient samples, followed by the ten-call two-sided check. Reserve the remaining calls for the original point and reproduction/relocation checks; never enlarge a started allowance.
• Expected ceiling is roughly two to three hours at measured native speeds.

Freeze source/input/environment hashes and originals, task identities, control and repeat rules, scalar-clock query limits and the native-call journal before launch. The runner must qualify before this prospective scope becomes executable; this preparation does not launch or charge observed work.

Report old versus new timing, objective, physical prediction changes, both side checks and all scheduled repeat agreements. A time relocation must reproduce the original prediction and cost within the preserved stability tolerances. The bounds, models, receiver weights, nominal clock and processing hypotheses are unchanged. Scientific decisions still abstain. A later combined numerical preference requires every model/case, control and repeat check; partial native success does not change a stretch's disposition.

The smooth repeats in different basins may subsequently require a small profile with nuisance reoptimization. That is a separate finite study, after the timing checks establish which endpoint cases remain problematic.