fix(#905): carry the BRepOffsetAPI_ThruSections capping-guard kernel patch

[gsdali]

What & why

ThruSectionsBuilder(isSolid: true) silently omits both end-cap faces for a closed section wire
with two or more periods of out-of-plane variation around the loop (a genuinely non-planar closed
curve) — build() returns true, but checkResult.isValid is false with no localized error,
and it's the root cause behind #702's demoted-solid symptom. Root-caused in
BRepOffsetAPI_ThruSections.cxx's MakeSolid(): it already tracks whether capping succeeded in a
local flag and discards it. This PR carries the fix as Scripts/patches/0026-*.patch, files it
upstream as OCCT#1462 (CI green), and
corrects a process error along the way — see "Notes for the reviewer".

Closes #905

CHANGELOG entry

ThruSectionsBuilder(isSolid: true) no longer silently reports success on an uncapped loft (#905)

ThruSectionsBuilder(isSolid: true) used to silently omit both end-cap faces for a closed section
wire with two or more periods of out-of-plane variation around the loop — a genuinely non-planar
closed curve, not just one with nonzero Z spread. build() returned true, but
shape.checkResult.isValid was false with errorCount == 0 and no localized error, and a
subsequent .healed() call silently demoted the result from a solid to a shell rather than
repairing it (root cause of #702). MakeSolid()'s capping helper (PerformPlan(), which only fits
a plane or a surface already attached to a wire's edges) already tracked whether capping actually
succeeded in a local flag and discarded it, unconditionally marking the result Closed(true)
regardless. Fixed in Scripts/patches/0026-*: throw instead, matching the pattern this same
function already uses for a null shell. Filed upstream as
OCCT#1462, validated via override-link.
Not yet shipped: the fix is carried, not baked into this release's xcframework —
ThruSectionsBuilder(isSolid: true).build() for this input still reports the old, wrong true
until the kernel is next rebuilt and re-pinned.

SemVer impact

NONE for this PR. It carries a kernel patch and documentation only; Package.swift's pin is
unchanged, so no consumer sees any behavior difference from this diff by itself. The fix takes
effect only once the kernel is rebuilt and re-pinned (a separate release-engineering step, same as
patches 0022-0025), at which point it will be a PATCH-level fix: no API signature change,
ThruSectionsBuilder(isSolid: true).build() correctly reports failure instead of a wrong success
for a non-planar closed section wire.

Checklist

• New or changed behavior is covered by a unit test in the same PR — N/A, see notes below:
  this diff has no OCCTSwift-observable behavior change (the kernel isn't rebuilt), so there is
  nothing for a Swift test to assert yet. The fix's test coverage is the new upstream GTests
  (NonPlanarClosedWireCappingFails, DegenerateVertexEndStillSucceeds) in
  OCCT#1462. A Swift-level regression test
  is deferred to whichever PR next rebuilds/re-pins the kernel, matching how 0022-0025 were
  each handled.
• Every new test and every new --self-test case was run once with its subject broken — done
  for the upstream GTest, not an OCCTSwift one: override-linked the unpatched .cxx and
  confirmed NonPlanarClosedWireCappingFails fails exactly as the issue describes (IsDone()
  reads true); override-linked the patched .cxx and confirmed it passes, along with
  DegenerateVertexEndStillSucceeds and all pre-existing tests in the file. Also compiled and
  ran the actual BOPAlgo_PaveFillerTest.FuseConeLoftWithBox_DegeneratedEdge, unmodified,
  against both: fails against the first (null-face) attempt's guard, passes against this one.
• The CHANGELOG entry above is complete, and docs/CHANGELOG.md is not in this diff.
• The SemVer impact above is stated, and docs/SEMVER.md is not in this diff.

Notes for the reviewer

Process correction, separate from the fix itself. Issue #905's comment 1 (an earlier session)
proposed the guard against a same-repo staging PR on the gsdali/OCCT fork —
gsdali/OCCT#1, that fork's first PR ever (verified via
gh pr list --repo gsdali/OCCT --state all, one result). That isn't how any other patch in
Scripts/patches/README.md was validated: every prior one went straight from local override-link
testing to the real PR against Open-Cascade-SAS/OCCT. Comment 2 on #905 is CI on that fork PR
catching a real regression in the first attempt's null-face guard
(BOPAlgo_PaveFillerTest.FuseConeLoftWithBox_DegeneratedEdge). This PR's fix is a different,
correct approach (see the CHANGELOG entry and the Scripts/patches/README.md write-up), and
gsdali/OCCT#1 has been closed as erroneous — its own last comment explains why and points here.

Why the fix isn't a null-face check. PerformPlan() leaves the cap face null in two different
cases: capping genuinely failed, or the wire is degenerate (a point section from AddVertex(),
e.g. a cone's apex) and needs no cap at all. A null-face check can't tell these apart; the local
bool B that MakeSolid() already computes can, since PerformPlan()'s own degenerate-wire
shortcut leaves it true. Full mechanism in Scripts/patches/README.md's 0026 entry.

Deliberately not done here, matching the precedent for 0022-0025: no OCCT.xcframework
rebuild, no Package.swift re-pin, no Swift-level regression test (would fail against the
currently-pinned kernel). All three are the release-engineering step comment 1 on #905 already
flagged as separate.

[gsdali]

Pushed a response to /code-review's findings (commit bd69e40):

• Fixed, and re-verified via override-link + re-pushed to the fork branch backing
  OCCT#1462 (CI re-running there): the new
  throw could leave face1/face2 (aliased to myFirst/myLast, exposed via
  FirstShape()/LastShape()) holding a real, never-added face. Now nullified before the throw.
• Fixed: Scripts/patches/README.md's "Pin consequence" paragraph was self-contradictory
  (said 0026 was outside the pin "same as 0022-0025" while also saying the pin includes
  0014-0025, which includes them — Package.swift's own census confirms all fifteen through
  0025 are baked in).
• Documented, not fixed: why GetStatus() stays _Done on this path, and why this patch
  throws rather than following sibling 0025's (GeomFill_Sweep.cxx overwrites a measured approximation error with the requested tolerance #597) surface-a-diagnostic-number pattern — both
  addressed directly in Scripts/patches/README.md's 0026 entry.
• Added: Scripts/repro/905-thrusections-capping-guard/, matching every other carried patch's
  convention (missing from the first commit).
• Filed separately, not folded in: ThruSectionsBuilder.generatedFace(from:) not checking
  IsDone() — a real, pre-existing bridge gap, but not introduced or worsened by this fix (the
  same partial-mutation-before-failure shape already existed for every other exception this OCCT
  class can throw). #910.