Locate propagation costs before changing the method
The complete prediction study has 16 time-limit failures. Profile two new, shortened generated-input calculations to locate costs before optimizing. This is diagnostic work, not a retry of any failed region.
Use the zero-field globe source under fixed-zero and profiled removal, each with its first 128 original packets and a query at 0.32 seconds. Construct new source references and physical-domain proofs for the shortened inputs; do not transfer full-history training certificates to changed sources. Keep sector 4, the original calibration caps, both source phases and all 27 derivative columns.
Finite allowance
- Two identities; one propagation bound each, 2,048 logical bound steps total.
- Two reference setups, two physical source setups and 256 reference appends.
- Two derivative setups, 256 derivative appends and 256 body-bound reservations.
- One worker, one numerical thread. No other numerical allocation at launch.
- 45 seconds per bound and 180 seconds overall, checked cooperatively.
- No retries, native trajectory predictions, LP proposals, observed fits, original-aircraft reads, protected ESA data or process signals.
Freeze source, inputs, compiler and environment. Preserve profiler output even if a bound does not finish. Report costs separately from qualification: profiler overhead and shortened histories do not establish full-history runtime, narrower prediction intervals or observational uncertainty. The completed comparison and all its failures stay unchanged.