3D engine that AI agents drive
- Cost of the #751 flicker fix (flush ranking and cut, #758 on), measured 2026-11-12 on main 2732ced2 (#939 merged) against 4443e672, the commit before #758. This change touches no bundle; it records the cost the owner accepted: a fix that stops this flicker for good is worth its size, memory and frame time.
- Gzip bytes (
npm run sizebundles, 4443e672 to main 2732ced2):entry:core85,185 to 90,062 (+4,877),entry:coplanar-overlap4,032 to 11,576 (+7,544),entry:path-tracer23,104 to 28,601 (+5,497),entry:path-tracer-progressive33,722 to 37,035,entry:path-tracer-realtime38,209 to 41,504,scene:lit-box54,333 to 58,788. Only part of that is the fix, because other PRs landed in the same window. Measured at the PR's own commits: #758 alone isentry:core+494 andentry:coplanar-overlap+3,848 (17599eed against 4443e672); #765entry:core+45 andentry:path-tracer+21; #769entry:coplanar-overlap+59 (its branch tip first reached 16,676 and was trimmed to 11,464 before merge). #935's own note counts 680 bytes inentry:path-tracer; the bundle did not move between the commits on either side of its merge, so that cost is not separable from the other growth in the window. Before #758, #752 (reversed-Z float depth) had addedentry:core+1,901 (83,284 to 85,185). - Memory per frame, read from the renderer (render-target ledger and
geometryBytes, 1280x720, 4x MSAA, HDR): render targets 65.22 MiB with the flush pass on, off and before #758, so the pass adds no target. Geometry on the Melbourne drape: 13.52 MiB before #758 and with the pass off, 14.32 MiB with it on (+0.80 MiB; 4 parts cut, 241,560 to 254,212 triangles, 15,848 vertices added). Trapstack-30-exact: 60 to 466 triangles;kitbash-duplicates: 12 to 99. - Frame time, headless Chrome on the shared RTX 3060 through gpu-lock, 3 rounds of 9 batches of 20 frames, each batch closed by a 1 px readPixels (this waits for the work queued, not for display). Quiet rounds, ms per frame, pass off (identical to before #758 on the drape) / pass on: drape roofs 0.41-0.42 / 0.43-0.44, drape overview 0.45-0.47 / 0.46-0.47, trap
stack-30-exact0.33 / 0.19,kitbash-duplicates0.17-0.23 / 0.23-0.24. Draw calls per frame: drape 10 off, 40 on;stack-30-exact2 off, 17 on;kitbash-duplicates2 off, 5 on. Rounds in which another session was using the card read 8 to 33 ms for every setting alike, so only the quiet ones are quoted; the gap between settings is under 0.03 ms on the drape and within noise on the traps. - Load time: the flush pass takes 0.9 to 1.5 s once at load on the Melbourne drape and 3 to 11 ms on the traps.
- Why the bytes are worth it: the drape flickered at every roof, bridge and river before the fix and shows none by eye after, and the ground-truth probe counts 0 true visibility flips with the pass on (#933, #939). The pass costs about 0.8 MiB of geometry, one second at load on a 240k-triangle city and no frame time this harness can see. Trimming the addon is a later optimisation, as the decision on #751 says.