[aka-blackboots]

Summary

Three changes on top of the display bounds cache that is already on main (#45):

1. Faster subtract on small curved solids. A face trim is now built once per face pair instead of on every point probe; the body intersection graph validates both bodies once instead of once per face pair; and the non-intersecting handler stops at the first coincident or branching pair instead of finishing the whole graph.
2. Point-in-trim test skips far segments. The exact point-to-segment distance is skipped for a trim segment whose bounding box, widened by twice the tolerance plus a rounding margin, cannot reach the probe. The crossing count still visits every segment, and non-finite values take the exact path.
3. OpenGeometry.releaseIdleGeometry(). A new call that disposes every idle display record (records kept for a quick level-of-detail switch back) and returns { records, bytes }. A three.js renderer leaves a dispose listener on every geometry it draws. After an app disposes a renderer, that renderer's buffer state stays reachable through the idle records until they are evicted. This call frees it at once. Records used by displayed bodies are kept.

Results

Exact B-rep results do not change: no golden key moved, and the kernel snapshot is identical to main.

Subtract times, in ms, median of 5, through the SDK in node (inline kernel):

Case	Before (3.0.0)	After
Shelled D-shaped solid minus a six-arc cutter (16 − 8 faces)	419.6	16.5
Arc-edged seat minus its cut-out (8 − 4 faces)	135.8	30.6
Basin rim	97.8	24.2
Second D-shaped solid minus cutter	106.1	15.5
Second seat minus cut-out	136.3	30.3
Same first case, native release build	172.9	3.9

Orbit with the display bounds cache from #45 together with this branch: a four-storey architectural model in a three.js app, hardware WebGL (Apple M2 Max), median of 3 orbits of 3 s. Frame time goes from 116.7 ms (p90 125) to 16.6 ms (p90 17.2, at the refresh limit). The level-of-detail pass goes from 78 % to 1.2 % of CPU time.

Behaviour notes

• One error kind changes, in a rare double failure. Take a subtract where a later face pair would raise a real error (not a coverage gap) after a first branch, and the prismatic curved handler then declines. Before, the call returned that pair's error. Now it returns the generic coverage gap "generic analytic solid". No corpus case reaches this path. Success paths, handlers and attempted are unchanged.
• New API, additive. One added line in the public surface: static releaseIdleGeometry(): { records: number; bytes: number; }. The README "Display" row lists it.

Tests

• WC-style subtracts: The pan-like and seat-like subtracts are pinned to their handlers, attempted list, topology (22/78/60 and 10/24/16), volumes and result digest. These values were recorded before the change.
• Face trim grid: The grid test is pinned to answers recorded from the old trim query on a holed arc-edged solid and a bored tube. The tube has a cylinder face with one hole, so the test runs the band path.
• Boundary band: A boundary test checks probes at half a tolerance and at three tolerances from a trim edge and corner. It runs on every platform.
• Early stop: The until-branch body graph stops at the first branching pair. A face pair query still refuses an invalid body.
• releaseIdleGeometry: Node tests show that idle records are disposed and counted, installed records survive, a released bucket is requested again, and the call returns zero after reset(). A browser spec shows that every idle record keeps a disposed renderer's listener, and that after the call none is left.
• Mutant proof: The speed-only changes were proven with planted mutants in a scratch copy of the crate. Each mutant turns a named test red.

Checks

• npm run check: 18 of 18 steps passed.
• Browser steps on three 0.168 and 0.184: passed.
• Kernel snapshot: diff against main is empty.