[visual-qa] Multi-arc CircularString offset/buffer densifies chords, not arcs; no round joins

[grootstebozewolf]

Observed

UX on TestBuilder RC2 (7cd60011).

CIRCULARSTRING (160 580, 270 790, 510 720, 430 390, 810 430)   -- 2 arcs
CIRCULARSTRING (137 675, 300 560, 395 711, 340 870, 530 900, 730 740, 680 580, 470 410, 230 360, 220 460, 110 480)  -- 5 arcs

Call	Result
Buffer.buffer(CS[3], 10)	CurvePolygon[9] (laser) + CurveOps.linearise log
Buffer.bufferAndInverse(CS[3], 10)	MultiPolygon[1576] (chainsaw)
Buffer.buffer(CS[5], 10)	Polygon[2575] (chainsaw)
OffsetCurve.offsetCurve(CS[3], 10)	CircularString[3] (laser)
OffsetCurve.offsetCurve(CS[11], 10)	LineString[44] (wrong)

Diagnosis (not self-intersection)

Clearance of the 5-arc case is 86.8 vs 20 required (curve_ring_simple_of_clearance). The offset ring should be simple. The bug is order of operations: densify chords then offset, so LineString[44] (~8.8 pts/arc) sits ~10 from the chords, not from the arcs (sagitta ~2 at r≈130 → sample distance 8–10 instead of constant 10).

Every CircularString join is a corner (shared endpoint, not G1). Samples: 42.9°–117.8° kinks, S-curve sign flips, major arcs (one −278.7°). Preservation dies at two arcs because OFF only handles a single 3-pt triple. Round-join edges are required (CurveRingOffset / Bible §7).

Expected

• Multi-arc offset: COMPOUNDCURVE of concentric CIRCULARSTRING pieces plus round-join CIRCULARSTRINGs at corners. Not LINESTRING.
• Sampled offset points: distance to the arcs is |d| (10), not 8–10.
• Multi-arc buffer: CURVEPOLYGON corridor (BUF-N / bufferOpenCompound), not POLYGON[2575].
• No CurveOps.linearise flood on those calls.

Not

• Not 3-pt OFF (already concentric).
• Not Info tooltip linearise ([visual-qa] Info click on CIRCULARSTRING floods Log with linearise warnings #84).