First-order tangent angular uncertainty validation

Pre-registered 2026-10-30. Replace the enclosing normal-space ball with the first-order covariance of the normalized normal: J=(I-nn^T)/||N||, covariance J C_N J^T. Bound its95% two-dimensional tangent ellipse by a circle with radius sqrt(chi2_2(.95)lambda_max), in radians, capped90 degrees. This is an approximate delta-method confidence circle, not a rigorous finite-noise angular guarantee. Same5x5 disparity fit and .5px noise assumptions, no new threshold/window tuning.

Validate on27 exact-plane coefficient scenarios: slopes (0,0),(.3,-.1),(.8,.4), center disparities6,12,24px; pixel principal point,(40,40),(280,150).20000 independent Gaussian coefficient samples per scenario, one fixed RNG20261031. Compare empirical angular coverage and bound medians with the previous enclosing-ball construction; report fraction whose bound<=15deg. Necessary plane coverage gate: every scenario

=94% (nominal95% with fixed1 percentage point development tolerance). Failing this gate blocks adoption even if some intervals shrink. Do not tune after results.

Then diagnose the same27 analytic curved/boundary frame conditions regardless of coverage-screen outcome, explicitly diagnostic only if coverage failed. Prior reliability gate plus radius<=15deg; keep previous precision/retention/spatial/ directional criteria. Compare identical depth hashes to prior experiments. Curved model bias is not described by Gaussian coefficient noise. No ICP/real transfer. No new calibration fits or changed recovery results. Tests verify analytic tangent variance at the principal point and missing-patch handling. Pin all code/protocol/ baseline hashes before execution. Report empty selections and all failed cases.