GeomPlate_MakeApprox drives the #522 approximator directly, and no census counts it

[gsdali]

Follow-up to #522. GeomPlate_MakeApprox is the one consumer of the defective code that does not
go through GeomConvert_ApproxSurface — it constructs AdvApp2Var_ApproxAFunc2Var directly
(GeomPlate_MakeApprox.cxx:302, :428, :464, :503). So it took #522's always-zero interior
error without appearing in any census built by grepping for GeomConvert_ApproxSurface, including
the PrecisCode census in Sources/OCCTBridge/src/OCCTBridge_Surface.mm.

Three bridge call sites

GeomPlate_MakeApprox approx(plateSurface, tolerance, 1, 8, tolerance * 10, 0);

• Sources/OCCTBridge/src/OCCTBridge_Healing.mm:950
• Sources/OCCTBridge/src/OCCTBridge_Healing.mm:997
• Sources/OCCTBridge/src/OCCTBridge_ProjLib_NLPlate.mm:232

All three use the six-argument form, so CritOrder = 0 and Continuity defaults to
GeomAbs_C1 (GeomPlate_MakeApprox.hxx:56-63). C1 means the degree collapse could not reach
them — the NDMINU floor is high there — but the reported error was zero-inflated exactly as
everywhere else, and these sites pass dmax = tolerance * 10 into a CritOrder = 0 G0 criterion
that the approximator evaluates against its own error estimate.

What to measure

1. Whether the plate/filling surfaces these three produce moved at all between a stock kernel and a
   0019 kernel. Same fingerprint-and-diff method as GeomConvert_ApproxSurface at GeomAbs_C0 returns a degree-1 collapse while reporting IsDone and a maxError five orders of magnitude too small #522's own 98-case sweep; the
   incremental slice rebuild in docs/guides/building-occt.md makes a real before/after cheap.
2. Whether the CritOrder = 0 criterion was doing anything before. If its threshold was being
   compared against a number that could not exceed it, the criterion was inert and dmax was a
   dead argument — the same shape of finding as Three parallel BRepAlgoAPI_Defeaturing bridge wrappers, the two shape-based ones split their feature (fuzzy tolerance) between overloads, and the header's cross-reference index only documents the third #497, where BRepAlgoAPI_Defeaturing ignored
   SetFuzzyValue entirely.
3. Whether Continuity should be defaulted or passed explicitly at these three sites. It is
   currently implicit, and BRepOffsetAPI_MakeFilling SIGSEGVs (uncaught) on non-C0 continuity against a curved boundary #430/FillingSurface.add(edge:continuity:) mis-maps both non-default cases: .c1 requests curvature, .c2 fails the whole build() #433/N-side filling is wrapped twice at two different OCCT layers — converge Shape.fill and FillingSurface onto one implementation #434 established that this family's continuity handling is
   where its bugs live.

Related: #430 (BRepFill_Filling/GeomPlate_BuildPlateSurface SIGSEGV), #434 (the two filling entry
points converged onto one implementation). Note BRepFill_Filling.cxx:712 mentions
GeomConvert_ApproxSurface but the line is commented out, so filling reaches the approximator only
through GeomPlate_MakeApprox.

Refs #522

[gsdali]

Fixed by #591 (merged into refactor/381-pass1b).

Six bridge functions build a surface with GeomPlate_MakeApprox — the issue's original census named three; measurement found six, since this class bypasses GeomConvert_ApproxSurface entirely and so sat outside every census built around that class. Five of the six hard-coded Nbmax = 1 with dmax = tolerance * 10, and together those two arguments made tolerance unenforceable rather than merely coarse.

Nbmax caps the Bezier patch count, and 1 is the one value that disarms the algorithm: AdvApp2Var_ApproxAFunc2Var::ComputePatches derives its cut decision from that cap, and at 1 every branch leaves it at "do not cut" — so AdvApp2Var_Patch::CutSense returns the same instruction whether or not the G0 criterion was satisfied. The criterion is still computed and still reported through CriterionError(); it just cannot act. dmax sets that criterion's own threshold (seuil = max(Tol3d, 10*dmax)), so tolerance * 10 asked it to accept 100x the requested tolerance — measured on a 25-point wavy plate: 0.098 reported against a 0.01 threshold, violated, surface returned unchanged.

All six sites now share one helper with one contract: Nbmax floors at 2, dmax is fixed at tolerance * 0.1 (matching the one site that already got it right), and continuity is passed explicitly (GeomAbs_C1, same as the implicit default all six already relied on — only C0/C1/C2 are accepted at all). Confirmed via kernel-source fingerprinting that this is unrelated to #522/#570: 54 plate fits compared stock vs patched, every control net identical — the implicit C1 default's degree floor was already above where #522's collapse could reach. Also confirmed the two .disabled("...segfault...") plate suites (18 tests) were stale — they pass cleanly, including against the pre-fix arguments, so re-enabled as an unrelated finding from the same investigation.

Note: this closing comment doesn't auto-link on the project board since Closes only populates closingIssuesReferences for PRs based on the default branch — this PR targeted refactor/381-pass1b.