mesh-compile: report warped faces' deviation as peak-to-peak (slab width), not corner-to-fit-plane

[GeneralPawz]

Follow-up to #254, from the consumer (axioval). #254 reports a warped authored polygon face as `Certified(w)`, where w is the largest corner distance from the fit plane used for triangulation. A warped face has no single true surface: either diagonal of a quad, a fit plane and a bilinear patch are all reasonable readings, and all lie within the corners' hull. Two such readings can differ by the face's whole spread along the normal.

Example: one corner of a quad lifted by h. The two diagonal triangulations differ by h/2 at the centre, while the fit-plane distance of every corner is h/4. So w understates how far another reader's surface can be from our mesh.

Proposal

Report the peak-to-peak normal spread of the corners: the slab width, max minus min signed distance from the fit plane, holes included, plus rounding allowance. Optionally keep w as a separate detail. Apply the same definition to warped B-rep plane faces (#257).

Done when

• The lifted-corner quad reports h (± rounding), not h/4.
• The ±h saddle reports 2h.
• Planar faces are unchanged (`Proven(0)`).
• The docs explain why peak-to-peak bounds the difference between any two readings within the corners' hull.

[GeneralPawz]

Landed on main in 411fb1d; it ships in the next release. One correction to the done-when: a square with one corner lifted by h reports h/2, not h. The fit plane through the centroid leaves the corners at ±h/4, so the slab is h/2 wide. That equals the gap between the two diagonal triangulations at the centre, and both readings, the flattened face and a bilinear patch all lie inside it. Measuring against the plane of the three unlifted corners would give h, but that bound is twice as loose for the same guarantee. A ±h saddle reports 2h, as specified.

[GeneralPawz]

Released in axiolid-construct 0.3.15, axiolid-evaluate 0.3.6, axiolid-reference 0.3.7, axiolid-mesh-compile 0.3.14 (9877cdc).