chore: pin the v3.0.0-rc1 kernel (V8_0_1 + seventeen carried patches)

[gsdali]

What & why

Package.swift pinned the v2.0.0 asset, which carries fifteen patches. Two more had landed in Scripts/patches/ since and were in no built kernel:

patch	issue	fix
0026	#905	BRepOffsetAPI_ThruSections::MakeSolid no longer marks a loft Closed(true) when it could not cap a non-planar extremity
0027	#913	CreateSmoothed's fixed-stride fill loop no longer overruns shapes when checkCompatibility(false) leaves sections with differing edge counts

Both were therefore exercised by no CI job at all, because build-and-test resolves the asset rather than building from source. That is #585's failure shape, and it is the exact thing CLAUDE.md's ten-second count check exists to catch.

Worth recording how it was found, because it is the failure mode rather than a description of it: the count check was run, found 16 against a prose "fifteen", and while the rebuild was being set up a git pull brought in 0027, taking the tree to 17. The drift happened during the session that was auditing for drift.

Rebuilt V8_0_1 from Libraries/occt-src with all seventeen applied. Three slices, zero errors. Published as v3.0.0-kernel.1.

Verification

Measured against the built artifact, not inferred from a clean build:

• Eight header-touching patches (0010, 0011, 0012, 0014, 0015, 0016, 0021, 0024): every line each adds to a shipped .hxx matched against the header inside the xcframework. 210 added lines, 0 missing.
• 0026: .cxx-only, but it adds a distinctive throw message. Present exactly once in each of libOCCT-macos.a, libOCCT-ios.a, libOCCT-sim.a.
• 0027: .cxx-only and adds no string literal, signalling through myStatus, so nothing in the binary can be grepped for it. Held behaviourally by StressBuilderLifecycleTests.mismatchedSectionEdgeCountWithoutCheckFailsCleanly, which is gated on OCCTSWIFT_LOCAL=1 (PR fix(#913): guard CreateSmoothed's fixed-stride shapes array against a mismatched section edge count #915 review, finding 1). Confirmed from the log that it started and passed rather than being skipped.
• Full suite, OCCTSWIFT_LOCAL=1: 5614 tests in 1460 suites, all passed, no failures, no crashes. That is what covers the five .cxx-only patches carrying their own regressions (0017, 0019, 0020, 0022, 0025).
• Published asset re-downloaded and hashed: 77df5a0a… matches the local zip exactly, so the pin resolves rather than 404ing. Package.swift's own sequencing note records that a correct checksum against a 404 fails identically, which is how a kernel.3 asset once passed a checksum check while resolving to nothing.

0018 and 0023 remain verifiable by no test of any kind, unchanged: the bridge stops both defects before OCCT sees them.

One thing this PR changes that is not the number

0027 is the first carried patch that cannot be verified by a green build-and-test, because its only test does not run in CI. Package.swift and CLAUDE.md now both say so explicitly, with the command to re-verify it locally. Previously the guidance was "a green build-and-test is behavioural proof", which is now true of five patches and false of one.

Three-way consistency

The v2.0.0 release check found a total of fifteen sitting above an enumeration of eleven, silently. So all three move together here:

ls Scripts/patches/*.patch | wc -l   →  17
Package.swift enumerated rows        →  17
CLAUDE.md prose                      →  seventeen

CHANGELOG entry

Kernel rebuilt: v3.0.0-kernel.1 carries seventeen patches, up from fifteen

Package.swift now pins the v3.0.0-kernel.1 kernel pre-release (V8_0_1 plus carried patches
0010-0012 and 0014-0027), replacing the v2.0.0 asset's fifteen. The two additions are
0026 (#905, BRepOffsetAPI_ThruSections refuses a loft whose non-planar extremity could not be
capped, instead of marking it Closed(true)) and 0027 (#913, CreateSmoothed refuses mismatched
section edge counts under checkCompatibility(false) instead of overrunning a fixed-stride array).
Both fixes previously existed only in the source tree, so a consumer installing the package received
neither, and neither was exercised by any CI job. No public API change.

Three verification facts are recorded in Package.swift rather than left to be rediscovered.
0027 is the first carried patch a green build-and-test cannot vouch for, because its only test
is OCCTSWIFT_LOCAL-gated and so never runs in CI. The suite total is blind to that: the same run
reports 5614 tests whether the test executes or is skipped, so only the per-test line distinguishes
them. And a kernel pre-release tag pointing at a commit that pins its predecessor is correct, not
a mistake to be fixed, because the alternative ordering is circular.

SemVer impact

NONE for the Swift API. No public entry point is added, removed, renamed, or changed in signature. The shipped kernel binary changes behaviour in two cases that were previously a silent wrong answer and a heap overrun respectively, both of which are the fixes named above.

Follow-up, tracked not forgotten

The v3.0.0-kernel.1 tag points at bc32a84e (main before this commit), so its tree pins the v2.0.0 asset rather than its own. Package.swift's sequencing note asks for the reverse order. I will re-point the tag at this PR's merge commit once it lands; nothing consumes the tag yet, so moving it is safe.