Compare runtime without changing the calculation
Use the two original shortened sources from the completed runtime diagnostic: 128 packets, zero-field globe, both processing assumptions, sector 4 and the 0.32-second query. Rebuild physical-domain certificates for these sources.
In one single-thread worker, run the unchanged implementation then the integer implementation for fixed-zero removal; reverse order for profiled removal. Record propagation time without profiler instrumentation or compilation. Require identical predictions, coefficient matrices, physical bindings, source counts and every error-remainder array. Retain failures and slower outcomes.
Separate finite allowance
Four fixed identities; four propagation bounds; 4,096 logical bound steps. Four reference, source and derivative setups; reserve 512 source appends, 512 derivative appends and 512 body bounds. One worker with one numerical thread. Each bound retains a 45-second cooperative ceiling; the new overall ceiling is 240 seconds. No retries, native trajectory oracle, observed fits, LP calls, original-aircraft reads, protected data or process signals.
Freeze inputs, sources, compiler, execution order and environment. A small engineering screen passes only if all outputs agree, the new implementation is faster in both cases and its total measured propagation time is at least 5% lower. One observation per case is not a general speed guarantee. Only a passing screen justifies the separately frozen full-history performance test; it cannot qualify complete prediction widths or observational uncertainty.