Solid offset and shelling with refusal on self-intersection

[GeneralPawz]

Problem

Axiolid has no solid offset, shelling, or Minkowski sum. minkowski and
offset_mesh match nothing in crates/, and there is no shell/thicken
producing a hollowed solid.

• OCCT: BRepOffsetAPI_MakeOffsetShape and MakeThickSolid are standard
  modelling operations.
• CGAL: minkowski_sum_3, plus polygon and polyhedron offsets.
• Manifold: offset and refine operations.

Note the planar half is already done: axiolid-overlay gained exact 2D
offset and stroke offset in v0.5 (3549c07). This issue is the solid half.

Why it matters

Shelling is how a wall becomes a wall assembly, a pipe becomes a pipe, and a
container becomes a container. In IFC terms it is the difference between a
solid extrusion and a hollow section — the RectangleProfile.thickness field
Axiolid already carries expresses exactly this intent at profile level, but
there is no operation to apply it to a finished solid.

Depends on

v0.9's general exact boolean (#77): an offset shell is naturally expressed as
the difference between an outer and an inner solid, and shelling a non-convex
solid requires it.

Scope

• Constant-distance offset of a planar-faced solid, outward and inward.
• Shell/thicken: hollow a solid to a stated wall thickness, optionally
  removing named faces to leave an opening.
• Self-intersection under offset must be detected and refused, not
  emitted. Offsetting a solid inward by more than its local half-thickness
  collapses it; that is a geometric fact the caller needs told, and it is
  precisely what the v0.7 detection work exists to catch.
• Concave edges under outward offset, and convex edges under inward offset,
  need explicit treatment — a naive face-plane push produces gaps or overlaps
  at exactly those edges.

Not in scope

• Variable-distance offsets.
• Offsets of curved-faced solids (same curved-surface prerequisite as General exact boolean over planar-faced solids #77).
• General Minkowski sum with an arbitrary structuring solid. Offset by a
  sphere is the special case worth having; the general operation is
  substantially heavier and has no named consumer yet. This is a deliberate
  narrowing — "CGAL has it" is not by itself a product requirement.

Verification

• bash scripts/gate.sh
• A cube offset outward by d has the volume of a cube with edge a + 2d,
  measured with the v0.7 provider — this fails immediately if edges or corners
  are mishandled.
• A shelled box has volume equal to outer minus inner, and its wall thickness
  measured across every face matches the request.
• Inward offset exceeding the local half-thickness returns a typed refusal
  rather than a collapsed or inside-out solid.
• Every result passes v0.7 diagnosis with no self-intersection.
• Mutation probe: skip the collapse check on inward offset; the over-offset
  refusal test must fail.

[GeneralPawz]

Landed on main as c42de3e (+ docs 8401f90).

offset_solid (outward and inward) and shell_solid in axiolid-construct.

Which offset this is

The miter offset, not the sphere offset. Offsetting by a sphere rounds every convex edge and introduces curved faces, leaving the planar family this crate works in exactly. Pushing each face plane along its normal and extending until the faces meet keeps everything planar — and gives (a + 2d)³ for a cube, which is the identity this issue verifies against.

Vertices are offset, not faces. Pushing faces and re-deriving the solid works for convex operands and silently fails for concave ones: at a reflex edge the pushed neighbours overlap instead of meeting. Moving each vertex to the common point of its own incident pushed planes gives the miter point in both cases and preserves topology, so the result needs no re-assembly.

Shelling is the difference between the solid and its inward offset — which is exactly why this depended on #77.

The collapse refusal took three wrong attempts

Each was rejected by a fixture, not by inspection:

1. Per-face normal preservation. Over-offsetting a convex solid inward turns it inside out as a whole, so every face reverses together and each still agrees with its own predecessor.
2. Per-face travel distance. A 2-cube offset inward by 1.5 lands its z=2 face at z=0.5 and its z=0 face at z=1.5 — both travelled exactly the requested 1.5, and they have swapped sides.
3. Half-space containment. This does catch the flip, but rejects the L-prism: a reflex corner correctly moves to (1.5, 1.5), outside the offset planes of the faces forming its notch — the operand class this exists to serve.

What holds for both is the slab between anti-parallel faces. Two opposing walls bound a slab an inward offset narrows by twice the distance; a width reaching zero means the walls have met. That is the "local half-thickness" the issue names, computed rather than estimated, and it says nothing about corners — which is where non-convexity lives.

Verification

bash scripts/gate.sh                    GATE PASSED
offset suite                            10 passed
architecture check                      44 packages

Closed-form oracles: cube outward (a+2d)³, cube inward (a−2d)³, shelled box outer-minus-inner, wall thickness measured off the result rather than assumed, and an L-prism offset inward checked against a hand-derived footprint (volume 5). Results pass v0.7 diagnosis with no self-intersection.

Mutation probe (mandated), anchor asserted before writing: skipping the collapse check fails both refusal tests.

Out of scope, as specified

Variable-distance offsets, curved-faced operands, general Minkowski sum. Also not implemented: removing named faces to leave an opening in a shell — that is in this issue's scope line and I did not build it. Worth a follow-up if you want it.