Files
Brett Lawson 7b31044a1b solidcore: retire ssi_split_curves_at_singular_points (WR-77) (#17886)
* solidcore: claim retiring ssi_split_curves_at_singular_points (WR-77)

Claim commit for the draft PR. Part of #14246.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BUqX4Moqx5bbkSocgEbsSV

* solidcore: always split SSI curves at singular points (retire WR-77 flag)

The split of crossing curves where the surface normals turn parallel
(the fork's sm_SplitCurvesAtSingularPoints) is now the only path; the
UpstreamFixes field ssi_split_curves_at_singular_points, SolidCraft's
preset line, its fixes.json and issue_2260 entries and the flag-off
tests are removed.

Turning it on everywhere made issue_1502's thickened HoodCorner come out
with an edge tolerance of 9.3e-5. Cause: the offset's
join_through_tangent_points rejoins the two pieces of the split crease
curve, and BSplineCurve::create_by_joining leaves a full-multiplicity
knot at the joint. The edge's trim ends 6e-8 before it, at the node's
offset vertex, and the UV trims dropped from that curve fit a span to
3.5e-5 instead of 7e-8. The joint of two pieces cut from one curve is
exactly removable, so remove_exact_joint_knot now takes the knot out
when Piegl-Tiller removal succeeds within EFF_ZERO (1 + max |coord|)
for every copy (BSplineCurve::remove_one_knot, all or nothing), and
the 3D deviation grows by the certified removal bound. Pieces traced
apart keep their knot (a true break).

Part of #14246

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BUqX4Moqx5bbkSocgEbsSV

* solidcore: WR-77 flag-removed addendum; the pipes case joins by default

Adds the WR-77 addendum recording the flag's removal and the joint-knot
fix it needed. fork_099cf96's flag-off check that the pipes' union fails
1001 becomes a check that it builds on the default kernel.

Part of #14246

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BUqX4Moqx5bbkSocgEbsSV

* solidcore: goldens for the singular-point split as the only path

Test change: 16 goldens rows move with the WR-77 split always on: the
11 surface.surface_p2 rows are certified exactly (each new curve is an
old curve or an exact piece of one, the pieces tile it, every cut lies
where the normals are parallel, sin < 1e-6); the 2
topology.brep_planar_section torus rows have the same base and new
digests as the pair certified exactly for #17570; the 3 boolean.api_brep
issue2110 rows are checked by issue_2110_tangent_cylinders_union
(Level 2 clean, manifold, closed-form volume within 1e-6).

Part of #14246

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BUqX4Moqx5bbkSocgEbsSV

* solidcore: the crossing-node review test sees each branch in two pieces

Test change: review_crossing_node_has_both_complete_branches expected the
two crossing branches of tc_parabolic_cylinders_crossing_node as two
curves. With the WR-77 split always on each branch comes out in two
pieces cut at the node, where the normals are parallel (the moved row's
exact certificate shows they are exact pieces of the old branches). The
test keeps every per-curve check (cubic, monotone x, on y = +-x within
the certified rounding, enclosed residual <= tol) and now asserts four
pieces, two per branch, that run from x = 1 to x = -1 and meet at the
node.

Part of #14246

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BUqX4Moqx5bbkSocgEbsSV

* solidcore: certify the joint-knot removal by reinsertion and keep partial removals

Review round 1 of the WR-77 retirement:

- `BSplineCurve::remove_knot_certified` replaces the all-or-nothing
  `remove_one_knot`. Piegl and Tiller's A5.8 removal is kept only when
  reinserting the removed copies (Boehm, in rounded `Ivl` arithmetic) gives back
  the old knot vector and control points within copies x tolerance. The largest
  control-point distance bounds the move at every parameter, since the two
  curves share their knots and basis. Rational curves are left alone.
- `remove_exact_joint_knot` counts the knot's multiplicity at the joint. A cut
  that lands on an existing knot loses only the copies that the joint can drop.
- Docs: the test is the curve's shape at the joint, not where the pieces came
  from.
- Tests:
  - the certificate bounds a real move and detects a moved control point;
  - a cut on an existing knot keeps that knot;
  - a removal beyond its bound is refused;
  - the saddle join asserts both its knots and its deviation;
  - `the_pipes_join` asserts a manifold solid, a clean Level 2 check and the
    closed-form volume.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BUqX4Moqx5bbkSocgEbsSV

* solidcore: the WR-77 addendum describes the certified, partial joint-knot removal

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BUqX4Moqx5bbkSocgEbsSV

* solidcore: certify each singular-split piece's deviation against its parent

Review round 2 of the WR-77 retirement (#17886):

- `split_curves_at_singular_points` no longer copies the parent's deviation
  onto every piece. The new `piece_deviation` bounds, at every parameter, how
  far each piece (the leading part and every tail) lies from its parent:
  - `piece_moves_at_most` restricts the parent to the piece's interval and
    refines both curves to one knot vector, all by Boehm insertion in rounded
    `Ivl` arithmetic, then takes the largest control-point distance;
  - a move within EFF_ZERO (1 + scale) is the split's rounding, so the
    deviation stays the parent's bit for bit;
  - a larger move, from `Trim`'s end-span cleanup, is added, rounded up;
  - if a piece dropped a parent knot and cannot be bounded (it is rational),
    that curve is left unsplit.
  In `ssi_adv_ssi_clasic_ruled_f0_f0` the two cleaned pieces grow by 4.44e-6
  and 1.59e-6. The other 182 pieces are unchanged.
- The Boehm steps of the removal certificate are now shared helpers
  (`insert_ivl`, `farthest_pair`).
- Docs: the deviation is the 3D curve's only, and rational curves are skipped
  in the joint-knot removal, with the reason.
- The addendum gives the issue1818 experiments in place of the guess.
- New tests: a cleaned piece's deviation covers its sampled move, an exact piece
  keeps its deviation, and a rational piece that dropped a knot gets none.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BUqX4Moqx5bbkSocgEbsSV

* solidcore: state what piece_deviation leaves unrecorded

The move threshold EFF_ZERO (1 + scale) hides the split's rounding, but
also a real cleanup of an end span below that size; say so in the
module and function docs and in the WR-77 addendum. The None branch's
"keeps every parent knot" case keeps the parent's deviation by an
argument from N_CrvRemoveDegenSegs, not a certificate; say that too.
Doc comments only.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BUqX4Moqx5bbkSocgEbsSV

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-16 22:33:10 +00:00
..