Does calculation precision explain the ESA stopping failures?
Question
The completed stopping diagnostic reproduces every saved calculation exactly. Seven field-globe fits still fail cost stability despite local position changes below 4 micrometres. Test two possible numerical effects separately: rounding when longitude differences are wrapped, and rounding accumulated during integration. This is a diagnostic, not a revised fit or an excuse to accept failed fits.
Fixed comparison
Use all 24 previously checked source identities. For each, reuse precisely the saved fitted point and its saved bounded correction, without recomputing that correction or selecting a new direction. Use the same earlier 800 positions, training-support increments, initialization, covariance and height-gravity anchor. No later observations, original aircraft files or protected flights are inputs.
Calculate each point in binary64 and extended precision. Reuse the binary64 output to compare the legacy longitude wrap with a wrap that leaves already in-range differences untouched; do not charge or run an extra native sweep for that algebraic comparison. Subtract receiver coordinates before converting extended results to binary64. Preserve binary64 packet-clock accumulation, query dispatch, query fractions, parameter products and processing-fraction mapping. Derive the scalar equations from the exact hash-bound original C++ source; do not copy or revise the physical mechanization. Require genuinely wider arithmetic on the host.
The three reported modes are binary64/legacy wrap, binary64/stable wrap and extended/stable wrap. The binary64 legacy result must reproduce both saved residual vectors exactly. Save a mismatch and stop that identity if it does not; do not continue to obtain a more favorable result. Report costs, cost changes, physical position changes and numerical provenance for every returned mode.
Qualification and allowance
Before preparation, qualify exact source/layout rejection, preservation of small longitude differences, a known-answer sub-binary64 coordinate displacement, all three physical candidates and both processing cases, full calibration vectors, fractional timing, height-dependent gravity, charge-before-execution, reference-mismatch rejection and unchanged labels. Bind the passing test report to source and environment hashes. Preserve the earlier precision implementation and its historical controls unchanged.
The first setup failed before starting any identity because the standalone worker loader could not resolve package-relative imports. Preserve that zero-calculation freeze separately. The corrected version requires an explicit standalone-loader test and a new source-bound qualification, identity and finite allowance; the completed setup must not restart.
Freeze 24 identities, 96 value sweeps, 18,421,785,600 packet/query work units, four workers with one numerical thread each, and a cooperative one-hour ceiling. There are zero optimizer starts and retries. Charge each native call before execution; save each returned array before judging reference agreement. No unused allowance transfers from a completed run.
No mode changes an original convergence label, threshold or scientific decision. An improvement would motivate separately qualified solver work. A negative result is retained and redirects diagnosis. Seven confirmation groups remain protected.