Compare full allowances, then connect prefix-specific source bounds

The streaming-reference qualification is complete; do not restart it. The cached transport layer is also qualified on its 24 native controls.

The six-source transport screen finishes with four compatible bounds and two time-limit failures. The runtime diagnosis is complete; follow its compiled arithmetic plan before another long comparison. Both east and height still miss the earlier complete-width target; a tighter vertical component alone cannot establish that every required axis passes.

The body-interval adapter has local tests for explicit reference rows and retained training ranges. Connect it at the actual packet midpoint using the streaming source state, with each native calculation retaining its original centers and all 51 physical/shared columns. Keep calibrated differences of later force samples. Include every reference center or coefficient discrepancy instead of assuming two maps are identical.

The adapter deliberately constructs new error bounds about the supplied body row. It does not silently reuse an old Taylor radius about an unspecified row; some other-coordinate allowances can therefore grow. Compare complete compatible outputs against the baseline, preserving valid per-axis minima only when all reference maps match. Report unsupported branches separately; a contradictory source interval must remain an error.

Before the complete candidate is frozen, include zero-width point domains, finite-difference tangent points and exact packet boundaries in the interval adapter checks. Account explicitly for reference-center rounding; a tiny numerical inconsistency must not be hidden by treating a contradictory interval as an unsupported branch.

New native/source qualification and complete prediction comparisons require their own finite freezes. No aircraft fitting or protected ESA confirmation follows directly from the source-reference result.