Merge-to-main condition: adjudicate all 61 deprecations, and rescan Pass 1a/1b for duplication those passes missed

[gsdali]

A condition for merging refactor/381-pass1b to main, not a follow-up. Both halves are things
a major release is the only chance to get right.

Part 1: the deprecation surface

Sources/OCCTSwift carries 61 @available(*, deprecated) symbols across 13 files:

File		File
Continuity.swift	9	Curve3D.swift	9
Curve2D.swift	8	Surface.swift	8
BRepGraph.swift	5	Shape+Topology.swift	5
Shape+Modeling.swift	4	Conic2D.swift	3
OSDPath.swift	3	FilletBuilder.swift	2
Shape.swift	2	Shape+ShapeHealing.swift	2
ThreadFeatures.swift	1

Plus the first bridge-side one, OCCTFacesAreAdjacent (PR #779, #783).

Nothing checks any of this. No gate script, no CI step, and no release step. A deprecation is a
promise to remove the thing later, and "later" for a 0.x-to-2.0.0 line is this release. Carrying a
deprecated symbol past the major means carrying it until v3.

Each of the 61 needs one verdict:

1. Remove it. The migration has been available long enough. v2.0.0 is the moment.
2. Keep it deprecated, with a stated reason and the version it will be removed in.
3. Un-deprecate it. The replacement turned out worse, or the deprecation was a mistake. This
   happens: fix(#761): AAG's sharedEdgeCount is no longer capped at 10 #779 nearly deprecated OCCTFacesAreAdjacent into a replacement that costs a full walk
   of the face pair where the original stopped at the first shared edge.

Verdict 3 deserves real attention rather than being treated as unthinkable. A deprecation asserts a
replacement is better, and several of these assert it about code that has since changed.

Part 2: rescan Pass 1a and 1b for duplication

Both passes are closed (#380, #381, 60 sub-issues). On 2026-08-07 alone, six duplications were
found in code both passes had already audited, every one by accident while doing something else:

• isRadiallyInwardFillet re-implemented detectHoles()'s inline radial test, down to the 1e-9
  guard (PR fix(#762): a filleted or chamfered pocket junction is absorbed, not reclassified #778).
• OCCTFaceGetSharedEdgeCount duplicated OCCTFaceGetSharedEdges's comparison loop, which is the
  shape that let the 10-edge cap survive (AAG's hand-rolled face/edge adjacency duplicates BRepGraph's outerWire(of:)/sharedEdges(between:and:) #761).
• OCCTDocumentSetShapeColor duplicated its own RGBA sibling, one copy-paste from reintroducing the
  alpha loss being fixed in the same PR (PR chore(#763): triage Document.mm + IO.mm's 26 unmeasured-value candidates #768).
• The census's reachability check was duplicated across branches, and that duplication is how a
  bug existed: added to one branch in round 2, not mirrored until round 4 (PR chore(#771): teach the census a third sub-kind, a gate flag that never flips #774).
• HarnessRunner duplicated CensusRunner almost line for line (PR feat(#772): opt-in self-intersection check for analyze(tolerance:) #773).
• PocketFeature.isOpen's edge-to-face incidence duplicates BRepGraph's indexed version, 3-8x
  slower (PocketFeature.isOpen's enclosure test is 3-8x slower than the indexed BRepGraph equivalent #777).

Six in one day, in audited code, none found by an audit. That is not a reason to distrust the passes;
it is evidence that duplication found by reading a diff is not the same set as duplication found by
reading a subsystem, and the passes only did the second.

The rescan should be cheap and targeted rather than a re-run:

• Start from the six above and ask what shape each shares. The dominant one is two entry points
  onto the same OCCT call that drifted, which is exactly Pass 1a's stated subject.
• Prefer a derived, committed artifact over a grep, per docs/v2.0.0-plan.md's census-once rule.
  Scripts/derive-bridge-header-split.py is the pattern.
• Bridge first. Five of the six are bridge or bridge-adjacent, and the bridge is where "wrap the same
  OCCT class twice" happens.

Why both in one issue

They share a deadline and a reason. Removing a deprecated symbol and unifying a duplicated one are
both breaking changes, and okf/policies/semver-at-release.md puts that judgement at the release.
Doing them separately means assessing SemVer impact twice over overlapping sets.

Definition of done

• Every one of the 61 has a recorded verdict, and docs/SEMVER.md reflects the removals.
• The rescan artifact is committed and its findings are either fixed or filed.
• If a deprecation check is worth automating, it is a gate script with a --self-test; if it is not,
  say why rather than leaving the question open.

[gsdali]

Correcting this issue's own premise, from PR #797's measurement.

I wrote that the dominant shape among the six was "two entry points onto the same OCCT call that
drifted", and told the rescan to confirm that rather than assume it. It is half:

• A, two entry points with near-identical bodies: AAG's hand-rolled face/edge adjacency duplicates BRepGraph's outerWire(of:)/sharedEdges(between:and:) #761, PR chore(#763): triage Document.mm + IO.mm's 26 unmeasured-value candidates #768, PR fix(#762): a filleted or chamfered pocket junction is absorbed, not reclassified #778. Findable by a clone
  detector, and this is what the artifact targets.
• B, the same question answered by two different algorithms with no shared text: PocketFeature.isOpen's enclosure test is 3-8x slower than the indexed BRepGraph equivalent #777. A token
  detector can never find this. PocketFeature.isOpen's hand-rolled edge-to-face incidence and
  BRepGraph's indexed one have nothing in common textually.
• C, duplication inside one function's own branches, or between dev-tooling files outside the
  audited population: PR chore(#771): teach the census a third sub-kind, a gate flag that never flips #774, PR feat(#772): opt-in self-intersection check for analyze(tolerance:) #773. A different detection problem.

The consequence for the rescan's own claim: it can establish the tree is clean of shape A, and
should say so in those words. Shapes B and C remain findable only by reading, which is what found
all six in the first place.

Worth noting the shape-A detector nearly produced a useless answer: loose settings flagged 3219 of
4152 bridge units, almost all OCCT's own deliberate 3D/2D mirrors. Six measured tightenings brought
that to 38 plus 21, from which six issues were filed (#791 to #796).

[gsdali]

Both halves complete.

Deprecations (PR #798): 62 of 62 removed. The uniform verdict is defensible because of an axis the issue did not specify: 43 never appeared in a tagged release, having been added and deprecated inside this branch's own refactor work, so removing them breaks nobody and is finishing a mid-flight rename. Only 19 ever shipped, and v2.0.0 is their removal point.

Review found one real correctness issue: ThreadBuild is Codable, so removing .boolean broke decoding of persisted values, and the migration guide claimed every removal was a compile error. init(from:) now maps the retired key to .direct, which is the value the last release with .boolean already produced. count-operations.py then caught the knock-on: two new public entry points, 4,254 to 4,256.

Duplication rescan (PR #797): the premise was half right. The six share three shapes, not one, and only shape A is findable by a clone detector. Shape B, the same question answered by two different algorithms with no shared text, is #777 and is invisible to token comparison, which PR #800 then confirmed by finding an instance of it while fixing shape A.

Six issues filed, #791 to #796. #791 closed by PR #799, #795 by PR #800.

The detector's own review round is the more interesting record: four parser bugs made it under-report, and fixing them took the Swift corpus from 3,334 units to 3,431. Ninety-seven newly visible functions produced exactly one new duplicate pair, which strengthens the clean-of-shape-A conclusion rather than undermining it.