fix(#597): GeomFill_Sweep reports the achieved C1-conversion error, not the request

[gsdali]

Summary

Kernel half of #597. The bridge half is closed out (PR #741 fixed OCCTGeomFillSweep's own gate;
PR #751 investigated the two remaining bridge sites and found both obvious fixes broke real tests).
This addresses the kernel defect the issue's own comments re-scoped it to.

GeomFill_Sweep::BuildAll measures the swept surface's real error at GeomFill_Sweep.cxx:286
(SError = Approx.MaxErrorOnSurf();). When ForceApproxC1 forces a re-approximation through
GeomConvert_ApproxSurface, the branch finishes at line 325 with SError = theTol;, overwriting
the measured error with the requested tolerance (a literal 1.e-4) instead of reading
ConvertApprox.MaxError(), which reports what the conversion actually achieved and sits two lines
above, unread.

• Confirmed MaxError() is the right quantity, not another GeomPlate_MakeApprox drives the #522 approximator directly, and no census counts it #571-style trap. GeomConvert_ApproxSurface's Surf argument here is mySurface, the exact surface being replaced, unlike GeomPlate_MakeApprox::ApproxError() (which measures an intermediate object and broke 6/6 real tests when gated on in PR investigate(#597): the two remaining bridge-side sites cannot be fixed by rejecting on error #751). Verified by reconstructing the identical GeomConvert_ApproxSurface call from outside the kernel: its output matches the real forced build's degree/pole counts exactly, and independent geometric deviation measurements from the same baseline are bit-identical either way.
• Confirmed it moves. Patch 0019 (GeomConvert_ApproxSurface at GeomAbs_C0 returns a degree-1 collapse while reporting IsDone and a maxError five orders of magnitude too small #522) is what makes this possible; before it, every interior truncation error was structurally zero. Measured MaxError() = 2.54714 against the pinned kernel, matching Did the sweep / offset / unify / convert consumers move when #522 made truncation errors real? #572's own independent measurement of the identical fixture (2.547) to the printed digit, 25000x the stock 0.0001.
• Validated via override-link against the real, in-kernel object (BRepFill_PipeShell/GeomFill_Sweep, not just the externally-reconstructed prediction): ErrorOnSurface() goes from 0.0001 to 2.54714. Every other value the harness prints (returned surface geometry, both independent deviations) is byte-identical before and after, confirming this is a diagnostic-only fix that changes no caller's returned shape.
• No regression risk found. No bridge site gates on this number today: PipeShellBuilder.errorOnSurface is info-only, and OCCTGeomFillSweep's own error gate (PR fix(#725,#597,#689): reject rational Bezier joins, fix tolerance-blind sweep, fix networkSurface #741) never sets ForceApproxC1 so never reaches this branch.

CError's four literal 0. entries a few lines above are left untouched (no fabrication, per #726:
no 2D curve error is available from GeomConvert_ApproxSurface at this point).

What's included

• Scripts/patches/0025-GeomFill_Sweep-report-achieved-conversion-error-597.patch, one line, applies
  cleanly to both our V8_0_1 pin and upstream master (b8f597c6), clang-format-clean.
• Scripts/repro/597-geomfill-sweep-error-overwrite/ — runnable probe, before/after transcripts, and
  the fixture-derivation writeup (getting a repro to fire needs a spine with a C0 corner inside one
  edge, not at a vertex, since BRepFill_Sweep splits at vertices).
• Scripts/patches/README.md and CLAUDE.md entries following the established pattern for prior
  carried patches (0019-0024).
• Scripts/repro/597-geomfill-sweep-error-overwrite/draft-pr.md: a drafted, unsent upstream PR
  body, per okf/policies/upstream-occt-style.md. No write of any kind was made to
  Open-Cascade-SAS/OCCT.

Patch 0025 is override-link validated only (no full xcframework rebuild), same as 0022-0024.
As of this PR, Package.swift pins v2.0.0-kernel.1 (eleven patches); 0022-0025 are carried in
the tree but outside that binary. PR #754 (open) re-pins to v2.0.0-kernel.2 with all fourteen
(0010-0012, 0014-0024); once that merges, 0025 becomes the sole patch left outside the pin,
which docs/v2.0.0-plan.md's RESOLVED block already names by number ahead of time.

SemVer impact

None. This changes only a diagnostic value (GeomFill_Sweep::ErrorOnSurface(), surfaced as
PipeShellBuilder.errorOnSurface) from a hardcoded constant to a measured one; no public Swift API
signature changes, and no existing test asserts a specific numeric value for it.

Test plan

• python3 Scripts/check-bridge-index.py / check-null-handle-guards.py /
  check-docs-defaults.py / derive-bridge-header-split.py --verify /
  count-operations.py: all clean (no Sources/Tests touched).
• Override-link validation against the real BRepFill_PipeShell/GeomFill_Sweep object:
  ErrorOnSurface() moves from 0.0001 to 2.54714; returned geometry and independent
  deviation measurements byte-identical before/after.
• Patch applies cleanly (git apply --check -p1) to the pinned V8_0_1 tag and upstream
  master (b8f597c6); clang-format --dry-run --Werror clean.
• Full swift test / kernel rebuild deferred to the next pin cycle (consistent with 0022-0024;
  no Swift/bridge source was touched, so no regression is expected).

Closes #597