Surface evaluation accepts a left-handed frame, and three call sites disagree on frame validity

[GeneralPawz]

Problem

Three call sites each carried their own frame validity rule, under three different tolerance policies, and only two of the three checked handedness.

Site	Tolerance used	Handedness checked
sampled/field/src/config.rs	tolerance.angular()	yes
execution/dispatch/src/section.rs	angular, clamped to 1e-6	yes
parametric/evaluate/src/surface.rs	hardcoded 1e-9	no

Two consequences.

A left-handed surface frame was accepted. to_local in evaluate/surface.rs checked unit length and pairwise perpendicularity but never the triple product. A mirrored basis satisfies both, so invert accepted it and returned reflected parameters. The reflection is self-consistent: those parameters round-trip through the same bad frame and look correct.

Surface evaluation ignored the caller's tolerance. invert already takes a Tolerance, but frame validation used a hardcoded 1e-9 regardless of the model unit.

Reproduction

Both added to crates/algorithms/parametric/evaluate/tests/invert.rs:

a_left_handed_surface_frame_is_refused ... FAILED

A Frame3 with z: -Vec3::Z and identity x/y is orthonormal, mirrored, and was accepted.

Resolution

SpaceFrame in axiolid-core: origin plus a right-handed orthonormal triad, private fields, to_local/to_world. It is the 3-axis counterpart to PlaneFrame and shares its error vocabulary, now renamed FrameError with a new NotRightHanded variant.

• Unit length, perpendicularity and handedness are dimensionless, so all three are judged against the ANGULAR tolerance.
• Handedness compares the triple product against +1. On an orthonormal triad that value is exactly +1 right-handed and -1 mirrored, so the two are separated by a margin of 2 rather than by a tolerance.
• Private fields make an invalid frame unrepresentable, not merely rejected.

All three call sites now delegate. dispatch/section.rs keeps its deliberate 1e-6 tolerance ceiling by capping before it calls the shared validator, so tighter caller tolerances still apply and looser ones still do not. evaluate/surface.rs threads its existing caller tolerance through instead of ignoring it.

Verification

• workspace 1113 passed / 0 failed, GATE PASSED
• mutation: deleting the handedness check fails a_left_handed_basis_is_refused in core AND a_left_handed_surface_frame_is_refused in the consumer, so the guard is proven at both layers
• zero hand-rolled orthonormality checks remain in production code

Landed in 6181dd2.