fix(view): circle shape SIGABRT + CircleTo stack overflow

[willwade]

Fixes two crashes in CDasherViewSquare triggered when LP_SHAPE_TYPE == CIRCLE (the legacy circular/compass node shape), both reachable via the C-API render path that was never tested with it.

Bug 1 — IsSpaceAroundNode SIGABRT (debug builds)

DasherViewSquare.cpp:527 asserted CoversCrosshair(...) for any node spanning the visible region. That invariant only holds for rectangular shapes — a circle/ellipse (and triangles/quadrics) can span the full visible height yet legitimately miss the crosshair, so the assert was unsound.

Fix: guard the assert to DISJOINT_RECTANGLE, OVERLAPPING_RECTANGLE, and CUBE; non-rectangular shapes fall through to their correct per-shape computation.

Bug 2 — CircleTo stack overflow (SIGSEGV)

CircleTo() (DasherViewSquare.cpp:441) subdivided the arc recursively with no depth bound. The convergence test could be defeated two ways:

• NaN poisoning — DasherSpaceArc (the game-mode brachistochrone, GameModule.cpp:239) can pass arc endpoints with |y - cy| > r, making sq(r) - sq(cy-y2) negative → sqrt(negative) → NaN → UB on the myint cast. UBSan: "nan is outside the range of representable values of type 'long long'".
• Integer-division stall — once y1, y3 are within 1, the midpoint stops advancing.

Either way the recursion never terminated → stack exhaustion → SIGSEGV.

Fix:

• depth cap (kMaxCircleSubdivisionDepth = 20 — 2^20 segments is far more than any real arc needs; ~13 levels reaches pixel accuracy),
• y1 >= y3 no-progress guard,
• clamp the sqrt argument to >= 0.

The remaining arc at the cap is approximated by a straight line, which is visually fine at that granularity.

Audit for similar issues

Per the bug report's heads-up ("we may have other settings like this"), I checked every DASHER_ASSERT and recursive draw function in the file:

• DasherLine2Screen (only other recursive draw fn) — bounded by integer-coordinate halving (~log₂(MAX_Y) ≈ 22 levels) + convergence guards. Safe.
• TruncateTri's asserts are internal to the triangle clipper; hold for triangle geometry. In scope only for triangle shapes.
• All remaining LP_SHAPE_TYPE dispatch sites (NewRender, CoversCrosshair, IsSpaceAroundNode) already handle CIRCLE.

So CircleTo was the sole unbounded recursion and IsSpaceAroundNode's was the sole geometry-unsound assert.

Tests

• test_circle_geometry.cpp — C-API smoke (circle frames render, deep forward driving, parameter switching in/out of circle mode).
• test_circle_view_internal.cpp — deterministic internal test driving DasherSpaceArc/CircleTo with degenerate coords (both LeftToRight and TopToBottom). Verified it crashes on the unpatched code (UBSan NaN cast + ASan SEGV) and passes clean after the fix.

Verification

• clang-format clean
• Debug build (asserts active): circle tests pass
• Sanitize build (ASan+UBSan): internal test passes clean; crashes deterministically on pre-fix code
• Release build: circle tests + all view/draw tests + deterministic_tests pass