Shared-reference native qualification
Freeze 24 generated identities, using byte-identical inputs from the completed regrouped-native qualification: all three candidates, both processing cases and the same tangent, packet-boundary, broad-region and independent stationary-pose controls. Physical models and all 27 calibration limits remain unchanged.
The new candidate separates six bounded reference-error coordinates from the unshared remainder. Every probe must subtract its actual known-control shared vectors before checking that remainder. The common block must cover those vectors, preserve all 33 enclosure columns and carry its source/expansion binding. Missing blocks or changed input identities must fail explicitly.
Acceptance retains the existing limits: centers/known poses within 1 micrometre, all 162 finite-difference physical tangent-column checks within 20 micrometres per normalized parameter at the fixed 0.002 step, and probe residuals within the factored remainder plus the declared 10 micrometre arithmetic allowance. Normalized known-control shared vectors must stay inside their declared boxes, with 1e-12 comparison tolerance. None of these numerical checks establishes a whole-program floating-point enclosure. A failure stops the sequence; it cannot be repaired by dropping a physical direction or assigning errors independently at every epoch.
Limits: 78 native evaluations / 19,632 native packet visits, 348 bound evaluations / 7,536 reserved bound packet steps, and a new 180-second cooperative ceiling. Tangent identities reserve one native call and 55 bounds; other identities reserve four native calls and one bound. Charge before every operation. One worker reserves four numerical threads. Preserve all native values/Jacobians, common vectors, full bounds, coefficients and failures. No observed fits, original-file reads, protected ESA evaluation, retries, extensions, signals or reuse of a completed allowance.
Local component tests also check the exact radial constant-velocity/time product, static lever at a nonzero expansion center and source-binding rejection. This finite native study precedes a separate joint 33-coordinate support interface qualification and a separately frozen usefulness comparison. It does not by itself establish long-record usefulness, receiver-error adequacy or observational coverage. All scientific outputs abstain.