Exact B-rep is the primary currency; tessellation strictly a requested output

[GeneralPawz]

The v1.0 compatibility promise depends on this being true at the API
boundary, not merely inside the compiler. It is the 1.0-level statement of
the same principle #3 and #15 establish inside the pipeline.

Depends on the v0.2 work landing first: this issue is the audit and the
public-API guarantee, not the implementation.

Done when

• no public entry point returns tessellated geometry unless the caller asked
  for it with an explicit tolerance
• the default path preserves exactness end to end
• a test asserts that an exact input survives a round trip through the
  public API without becoming a mesh

[GeneralPawz]

Audited and landed as 57ce0db.

The finding up front: the guarantee already held. No public entry
point converts exact geometry to a mesh without an explicit tolerance.
What was missing was any statement of it, and any test that would
notice if it stopped holding.

What the audit checked

Every public function returning a mesh, classified by INPUT type
rather than output, since only exact-in/mesh-out is a leak:

Entry point	Verdict
tessellate_primitive	takes Tolerance
compile::brep::tessellate	takes Tolerance
Tessellator::tessellate	takes TessellationOptions
profile_rings, profile_to_contour	take an explicit chord budget
triangulate(&Polyhedron)	exact decomposition, invents no tolerance
loft, sweep, convex_hull	take already-discrete input

ExactBRep has no consumer that turns it into a mesh implicitly.
TessellationOptions and TessellationRequest have no Default,
and there is no global chord-error or Tolerance::default() anywhere.

Landed

tests/exact_primary.rs, one test per acceptance criterion:

• no mesh unasked - ExactBRep carries no tolerance field, so a
  mesh cannot be spelled without selecting Tessellation and
  supplying one
• default path stays exact - a tessellated result carries the
  tolerance it was built to, so it cannot later pass as exact
• round trip - a square extrudes through the public facade and
  comes back with six analytic Plane surfaces

The round trip is end-to-end, not type-level: it runs a real
extrude_profile_exact and counts planes. Had the public path
tessellated anything, they would be triangles.

The facade module docs now state the guarantee, so it is a promise a
consumer can read rather than folklore.

Mutation evidence

Green tests prove nothing until they can fail:

• drop the expected plane count 6 -> 0: round-trip test fails
• weaken requires_exact_brep to false: criterion-1 test fails

Scope limit, stated plainly

Nothing in the workspace constructs GeneratedGeometry::Tessellation
yet - the enum has 6 references total and no producer. So the
guarantee is real at the API boundary and enforced by the type
system, but that particular variant is a declared contract rather
than a path under load. Worth knowing before treating this as
end-to-end proof for v1.0.

Gate passed, main at 57ce0db, tree clean.