Draw dice with the rounded edges the solver collides them with

[drehtuer]

Step 4 on feature/realistic-rendering: rounded edges and corners on the dice.

The physics die was never sharp. JoltWorld gives Jolt a convex radius of 3 % of size_mm (0.48 mm on a 16 mm die). Jolt clamps that by its corner-error limit (the d4 gets 0.25 mm) and by the die's thinnest width. So the sharp picture stood up to 0.5 mm outside the colliding die: into the felt on a corner contact, and into each other where corners touched.

This draws exactly the solver's rounding (decision 91):

• The drawn radius repeats Jolt's limits over the same faces, so the drawn surface is the colliding surface: the gap is 0 by construction.
• Physics is untouched: the 0.03 share moved unchanged into simulation/api/HullMargin, which JoltWorld reads, so no golden, fairness or harness figure can move.
• RoundedEdges builds the mesh: flat faces shrunk on their own planes, edge cylinders (≤ 30° a segment), sphere-patch corners, smooth normals on the curves. All eight catalogue shapes are covered.
• Flat-part UVs equal the old mapping, so numbers and artwork sit exactly where they did; artwork now runs round the bends. The d4's corner numerals reach 0.19 mm onto the bend.
• DieMesh.of(shape, rounding = 0.0) reproduces the old mesh exactly.
• Triangles, sharp → rounded: d6 12 → 204, d20 20 → 260, coin 92 → 1,052. That is 100d6 at ~20k triangles, still to be measured with tools/harness.sh --rendered.

On the Pixel 10a:

• the full :render:filament device suite passes: 39 tests, 0 failures, including PrintedNumbersDeviceTest;
• in the gallery, edges and corners read softer, and dice on a face sit exactly where they did.

Tests: RoundedEdgesTest and HullMarginTest cover face planes kept, rounding 0 = old mesh, UVs, outward unit normals, a closed mesh, a zero gap and the counts. Function coverage is 93.2 %, branch 73.1 %.

Docs: docs/physics-and-rendering.md ("Rounded edges", with the "Dice bodies" bullet corrected), decision 91.

Based on #376.

🤖 Generated with Claude Code

[drehtuer]

Rendered harness on the Pixel 10a, 20 rolls of 100d6 with rounded dice (3,864 frames): GPU p50/p99 6.5 / 14.8 ms against 6.2 / 13.7 ms before rounding, work p99 25.5 ms against 24.3 ms, 57.4 fps against 57.3. That is about 1 ms of GPU at the capacity limit, and the GPU bar of 16.6 ms is still met. The work p99 was already over 16.6 ms at 100 dice before this change (the physics steps; see "Performance"). Recorded in the docs in 4c4… (latest commit).

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
100.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud