P2.5.7 package-private curve DCEL (Option C, OV-P2 / Bar 2)

[grootstebozewolf]

Stacked on #37 (cursor/p256-mixed-overlay-17ab @ bb566aa). OV-P2 / Bar 2. Option C only. Do not stack on #27 (OverlayNG-for-circles / public SegmentKind is a dead-end draft). Do not merge this into #7. Does not merge #12 / #13 / #15 / #16 / #17 / #18 / #19 / #20 / #23 / #24 / #25 / #37 / #7. Does not push to feature/sfa-curve-rgr or master. Does not rebase #7 onto locationtech. Does not pick core HalfEdge.java / Quadrant.java.

Sentence

Curve DCEL for PLG/COV. Not a noder. Not OverlayNG-for-circles. Not public. Not a merge into #7.

PLG and COV do not need a noder (public or curve). They need an arrangement structure: half-edges, twins, next/prev, incident face, on CurveSegmentString members (an arc stays an arc). CurveSegmentNoder already names leftover edges for overlay kits. That is a different product. This PR is the DCEL.

SIGN retip (this tip)

CurveSegmentDcel.onString stays in R². Supersedes the HOLD on the hypot rewrite. Not a new rung. Not full curve noding. No public noder. Option C HOLDS (package-private in jts-curve).

• Arc: compare (dx²+dy²) to R² with tolerance eps², floored at 1 ulp of the two squares so a representable on-circle node is not rejected. Not hypot(d) − R. Not a length eps on d². Do not take sqrt(R). Do not divide by sagitta for the decision.
• Chord: the same squared residual (dx²+dy² ≤ eps²).
• CurveSegmentDcelTest pins extreme sagittas in R²: flat sagitta → 0, and a high-sagitta case.
• leaveAngle / compareLeave HOLD (endpoint quadrant + Orientation.index; atan2 only 1e-8 window / sweep sense). Untouched.

Tip: 2ef0bc6b.

What this is

Package-private CurveSegmentDcel in jts-curve. Built from CurveSegmentString members plus the P2.5.2 node set and already-named MIXED / shared-edge ends. Cycle order is the Faces left-most / next-outgoing walk, persisted as links (left face, CCW interiors). CurveSegmentFaces now assembles Geometry from the DCEL's bounded faces; it is a consumer, not this product.

Leave-angle robustness (locationtech locationtech#1224 / locationtech#1226)

This DCEL does not use core HalfEdge or Quadrant. It sorts its own Half leave direction and stamps TANGENT_LEAVE_ANGLE. Those locationtech fixes are ported into that walk only:

• Do not subtract to form a direction vector, then atan2 / == those deltas (Make HalfEdge.compareAngularDirection more robust locationtech/jts#1224).
• Quadrant of the leave ray comes from the endpoints (chord: origin→dest; arc tangent: origin vs centre), not FP dx/dy (Improve Quadrant computation robustness by using vector endpoints locationtech/jts#1226).
• Same-quadrant order is Orientation.index.
• TANGENT_LEAVE_ANGLE fires when that compare ties, or when an arc leave is inside the 1e-8 window (P2.5.4). Chord coincidence stays on the robust compare so Make HalfEdge.compareAngularDirection more robust locationtech/jts#1224-style distinct direction points do not collapse.

DEVELOPING.md and version history are untouched. Core HalfEdge.java / Quadrant.java are untouched.

Honesty lock

Generic walk-on-nodes==null is unsafe: a MIXED abort can hide a real crossing (HALF_DISC × HALF_CROSSING_UPPER). That pair stamps MIXED_HIDES_CROSSING and stops. A coincident leave-angle is snap-rounding (TANGENT_LEAVE_ANGLE). Pinch / kiss / holed Geometry-level stay null. No invented noder. No densify-then-snap flagged exact. No core SegmentString / HotPixel / noding.snapround.

Cells

Cell	Product
R1.5 two-disc crossing	DCEL: twins, next/prev, 3 bounded faces (lens + crescents)
H-SHELL-N-MIXED (named diameter)	DCEL: T-junctions at (±1, 0), inner + bite
HALF_DISC × HALF_HANGING × STADIUM_FOUR	DCEL: nine bounded faces, same cells Faces already named
P2.5.4 HALF_DISC × HALF_HANGING × STADIUM_ODD	named stamp TANGENT_LEAVE_ANGLE
locationtech#1224-style distinct direction points	leave compare ≠ 0; not a TANGENT stamp
HALF_DISC × HALF_CROSSING_UPPER	named stamp MIXED_HIDES_CROSSING
Pinch / TOUCH-ext / holed Geometry	null
Flat sagitta / high-sagitta onString	R² residual (eps²); hypot HOLD superseded

What this is not

• Not a noder. CurveSegmentNoder stays nodes + named leftover edges for overlay kits.
• Not OverlayNG-for-circles / P2.5.5 OverlayNG-for-circles (Option B contract, OV-P2 / Bar 2) #27 / Option B / public SegmentKind.
• Not a public API. Package-private in org.locationtech.jts.operation.overlayng.curve.
• Not a merge into feature/sfa-curve-rgr / PR SFA/SQL-MM curve awareness: Curve rename, arc-aware operations, OverlayNGCurve ratchet (#1195 RGR series) #7 / locationtech master.
• Not a cherry-pick of locationtech HalfEdge.java / Quadrant.java.
• Not PLG, not COV (they eat this DCEL later).
• Not leftover overlay kits (H-ANNULUS-TANGENT, TOUCH-ext, R-LL). Those stay kits.
• Not HP.4 / HP.5, not ML.2, not densify-as-laser.

Pins

• Twins and cycle links, not just "we got a Geometry back".
• Members stay CurveSegmentString (arc stays arc, chord stays chord).
• Slack 15%. OverlayNGCurve never Curved. Java 8, Checkstyle, PMD (AvoidBranchingStatementAsLastInLoop).
• No new public API. No circular noder in jts-core.
• onString extreme sagittas in R² (flat → 0, high-sagitta). leaveAngle / compareLeave unchanged.

Verify

mvn -pl modules/curve -am test -DfailIfNoTests=false -Dcheckstyle.skip=true -Dpmd.skip=true \
  -Dtest=CurveSegmentDcelTest,CurveSegmentStringTest,CompoundCurveShellOverlayTest,CircularArcOverlayTest,CircularDiscOverlayTest,CircularDiscPolygonOverlayTest,CircularLineOverlayTest,OverlayNGCurvePhase0Test,OverlayNGCurveRatchetTest,OverlayNGCurvePerfGateTest
bash dev/check-no-curved.sh
mvn -pl modules/curve checkstyle:check
mvn -pl modules/curve pmd:check

All green after this revision (2ef0bc6b):

• CurveSegmentDcelTest 10 / 0 (twins / next / face; P2.5.4 stamp; Make HalfEdge.compareAngularDirection more robust locationtech/jts#1224 leave-compare pin; extreme sagittas in R²)
• CurveSegmentStringTest 24 / 0
• CompoundCurveShellOverlayTest 28 / 0
• CircularArcOverlayTest 11 / 0
• CircularDiscOverlayTest 12 / 0
• CircularDiscPolygonOverlayTest 7 / 0
• CircularLineOverlayTest 7 / 0
• OverlayNGCurvePhase0Test 19 / 0
• OverlayNGCurveRatchetTest 50 / 0
• OverlayNGCurvePerfGateTest 84 / 0 at 15% slack
• bash dev/check-no-curved.sh OK
• Checkstyle 0 / PMD clean on jts-curve