ldr 0.33.10 → 0.33.12

package/README.md

@@ -143,6 +143,7 @@ MPD-embedded parts, such as Studio/OMR unofficial parts or stickers, become part
 | `lodPx` | `0.5` | LOD: maximum screen-space error in pixels. |
 | `lodEdgePx` | `0.5` | LOD: draw edge and conditional lines only where one stud (20 LDU) spans at least this many pixels. `0` = always. |
 | `lodEdgeMinPx` | `0.4` | LOD: draw only edge segments at least this many pixels long on screen. `0` = every segment. |
+| `lodEdgeCap` | `0.75` | LOD: a part draws at most `lodEdgeCap` × d² of its edge segments (its longest), d = its size on screen in pixels. `0` = no cap. |
 | `lodUnder` | `true` | LOD: skip undersides that can't be seen (camera behind every bottom face of a cluster's parts, or the body drawn as its closed bounding box). Lossless in the fidelity checks. |
 
 `LdrObject`: `object`, `stats` (`instances`, `drawn`, `missing[]`, `studsHidden/Total`, `undersHidden/Total`, `buildMs`, `condBatches`, `texBatches`, `textures`, `glossmaps`, `clusters`), `stepCount`, `stepViews`, `texturesReady`, `setStep(n | null)`, `setEdges(on)`, `setConditionalLines(on)`, `setCulling(studs, undersides)`, `setSide(side)`, `setLod(on)`, `setGhost(on, { opacity, color, whiten }?)`, `getBox(target?)`, `dispose()`.
@@ -201,7 +202,7 @@ How much each optimization contributed on Datsville 100k (Rounds 1–20, before
   - Full Datsville: **1.7 → 9.9 fps**.
   - Datsville 100k: **8 → 43.5 fps**.
 - **Conditional lines** (opt-in `cond` packs) cost **+12–14 %** frame time on Datsville 10k and 100k (M4 Pro, before Round 60), and add 7–14 % to the zstd pack size.
-- **Edges:** with LOD, edges take the Falcon fleet from 9.8 → 12.9 ms (19 ships) and 18.3 → 22.9 ms (75 ships) on the Mac Mini M6 (Round 67); before Round 62 they cost 12.8 → 33.3 and 19.3 → 45.1 ms there. A simplified level draws its own outline: square studs a square per stud loop (8 segments instead of 16-32), bodies at their bounding-box level the box's 12 edges, and a line batch whose edge LOD leaves no segment issues no draw. Each cluster draws only edge segments at least 0.4 px long on screen (`lodEdgeMinPx`; 0.15 px until Round 62, when 85% of the segments drawn at the fleet overview were 0.15–0.5 px long and each darkened a whole pixel, so far ships looked charcoal) and none where bricks are sub-pixel (`lodEdgePx`). Far views drawn with edges get lighter: 2.6% of pixels differ from every-edge rendering on the 19-ship overview, 0.8% on the SF city, up to 4.7% on Datsville overviews. `lodEdgeMinPx: 0, lodEdgePx: 0` keeps every edge.
+- **Edges:** with LOD, edges take the Falcon fleet from 9.8 → 12.9 ms (19 ships) and 18.3 → 22.9 ms (75 ships) on the Mac Mini M6 (Round 67); before Round 62 they cost 12.8 → 33.3 and 19.3 → 45.1 ms there. A simplified level draws its own outline: square studs a square per stud loop (8 segments instead of 16-32), bodies at their bounding-box level the box's 12 edges, and a line batch whose edge LOD leaves no segment issues no draw. Each cluster draws only edge segments at least 0.4 px long on screen (`lodEdgeMinPx`; 0.15 px until Round 62, when 85% of the segments drawn at the fleet overview were 0.15–0.5 px long and each darkened a whole pixel, so far ships looked charcoal) and none where bricks are sub-pixel (`lodEdgePx`). A part also draws at most 0.75 × d² of its segments, its longest, d = its size in pixels (`lodEdgeCap`, Round 72), so a 6 px technic beam no longer draws ~100 segments into ~20 pixels. Far views drawn with edges get lighter: 2.6% of pixels differ from every-edge rendering on the 19-ship overview, 0.8% on the SF city, up to 4.7% on Datsville overviews. `lodEdgeMinPx: 0, lodEdgePx: 0, lodEdgeCap: 0` keeps every edge.
 - **Build steps:** `setStep` takes 39 ms on the 19-ship fleet (26,619 steps); on Datsville 100k it went from a full rebuild (752 ms) to 66 ms.
 - **Visual fidelity:** pixel diffs show culling and single-sided faces are lossless (≤ 0.009 % of pixels differ). Low-detail studs only change stud silhouettes.
 
@@ -228,7 +229,7 @@ Reproduce: `cd bench && bun run setup && bun run perf` (regression gate) or `bun
 - **ROTSTEP** is reported (`stepViews`, `instance.view`), not applied; `planStepViews` takes a `fixed(step)` override if you want to honour it.
 - **Packs are static:** they hold only the parts they were built for. A model that uses parts missing from the pack lists them in `stats.missing`.
 - **Occlusion culling:** beyond hidden studs and undersides there is no occlusion culling (bricks enclosed by other bricks are still drawn). Frustum culling and LOD apply per spatial cluster.
-- **LOD and edges:** edge LOD drops sub-pixel edge segments and hides edges where bricks are a fraction of a pixel, which visibly lightens far-away views drawn with edges (2.6% of pixels on the 19-ship Falcon overview, up to 4.7% on Datsville overviews). Set `lodEdgeMinPx: 0, lodEdgePx: 0` to keep every edge (much slower on large models), or `lod: false` to turn LOD off entirely.
+- **LOD and edges:** edge LOD drops sub-pixel edge segments and hides edges where bricks are a fraction of a pixel, which visibly lightens far-away views drawn with edges (2.6% of pixels on the 19-ship Falcon overview, up to 4.7% on Datsville overviews). Set `lodEdgeMinPx: 0, lodEdgePx: 0, lodEdgeCap: 0` to keep every edge (much slower on large models), or `lod: false` to turn LOD off entirely.
 
 ## Attribution and license
 

package/dist/cli.js

@@ -1,18 +1,18 @@
 #!/usr/bin/env node
 
-// ../../../../../../../tmp/ldr-build-Fw2uhS/src/cli.ts
+// ../../../../../../../tmp/ldr-build-HYKc29/src/cli.ts
 import fs4 from "node:fs";
 import zlib from "node:zlib";
 
-// ../../../../../../../tmp/ldr-build-Fw2uhS/src/node/build-pack.ts
+// ../../../../../../../tmp/ldr-build-HYKc29/src/node/build-pack.ts
 import fs3 from "node:fs";
 import path3 from "node:path";
 
-// ../../../../../../../tmp/ldr-build-Fw2uhS/src/node/resolver.ts
+// ../../../../../../../tmp/ldr-build-HYKc29/src/node/resolver.ts
 import fs from "node:fs";
 import path from "node:path";
 
-// ../../../../../../../tmp/ldr-build-Fw2uhS/src/spec.ts
+// ../../../../../../../tmp/ldr-build-HYKc29/src/spec.ts
 var norm = (s) => s.trim().toLowerCase().replace(/\\/g, "/").replace(/\/+/g, "/");
 var COLOUR_ID_RE = /^[A-Za-z][A-Za-z0-9_]*$/;
 var namedCodes = new Map;
@@ -817,7 +817,7 @@ function fileStepViews(f) {
   return out;
 }
 
-// ../../../../../../../tmp/ldr-build-Fw2uhS/src/model.ts
+// ../../../../../../../tmp/ldr-build-HYKc29/src/model.ts
 class ModelLimitError extends Error {
   limit;
   code = "ldr-model-limit";
@@ -1259,7 +1259,7 @@ function blendedColourHex(code) {
   return "#" + [0, 1, 2].map((k) => (a[k] + b[k] >> 1).toString(16).padStart(2, "0")).join("").toUpperCase();
 }
 
-// ../../../../../../../tmp/ldr-build-Fw2uhS/src/node/resolver.ts
+// ../../../../../../../tmp/ldr-build-HYKc29/src/node/resolver.ts
 var TOP_STUD_RE = /^(stud|studa|stud2|stud2a|stud6|stud6a|stud10|stud15|stud2s|stud2s2|stud-logo\d*|stud2-logo\d*)\.dat$/i;
 var UNDER_STUD_RE = /^(stud3|stud3a|stud4.*|stud12|stud16.*|stud18a)\.dat$/i;
 var norm2 = (ref) => ref.trim().replace(/\\/g, "/").toLowerCase();
@@ -1586,11 +1586,11 @@ function collectDeps(lib, refs) {
   return out;
 }
 
-// ../../../../../../../tmp/ldr-build-Fw2uhS/src/node/compile.ts
+// ../../../../../../../tmp/ldr-build-HYKc29/src/node/compile.ts
 import fs2 from "node:fs";
 import path2 from "node:path";
 
-// ../../../../../../../tmp/ldr-build-Fw2uhS/src/node/lod.ts
+// ../../../../../../../tmp/ldr-build-HYKc29/src/node/lod.ts
 import { MeshoptSimplifier, MeshoptEncoder } from "meshoptimizer";
 await Promise.all([MeshoptSimplifier.ready, MeshoptEncoder.ready]);
 function reorderMesh(P, N, C, I) {
@@ -1896,7 +1896,7 @@ function levelLines(P, I, E, err, always, tol = 0.01) {
   return out;
 }
 
-// ../../../../../../../tmp/ldr-build-Fw2uhS/src/node/compile.ts
+// ../../../../../../../tmp/ldr-build-HYKc29/src/node/compile.ts
 class BfcState {
   accum;
   ccw = true;
@@ -2840,7 +2840,7 @@ function encodePack(parts, studKinds, weld, extra, options = {}) {
   return Buffer.concat([head, ...chunks.map((c) => new Uint8Array(c.buffer, c.byteOffset, c.byteLength))]);
 }
 
-// ../../../../../../../tmp/ldr-build-Fw2uhS/src/node/build-pack.ts
+// ../../../../../../../tmp/ldr-build-HYKc29/src/node/build-pack.ts
 function packCompilerOpts(o) {
   return { studs: o.studs ?? "low", prims: o.primitives ?? "full", weld: o.weld ?? true, separateStuds: true, cond: !!o.cond, tex: !!o.textures, ...o.partWorkLimit ? { workLimit: o.partWorkLimit } : {} };
 }
@@ -2913,7 +2913,7 @@ function buildPack(o) {
   return { data: encodePack(parts, kinds, comp.opts.weld, extra, { texture, lod: o.lod !== false, reorder: !!o.reorder }), parts: parts.size, studKinds: kinds.size, missing: [...missing], ...warnings.length ? { warnings } : {} };
 }
 
-// ../../../../../../../tmp/ldr-build-Fw2uhS/src/cli.ts
+// ../../../../../../../tmp/ldr-build-HYKc29/src/cli.ts
 var HELP = `ldr pack — build a parts pack for the ldr three.js renderer
 
 Usage:

package/dist/cull.d.ts

@@ -21,17 +21,21 @@ export declare function mul12To(a: ArrayLike<number>, b: ArrayLike<number>, bo:
 export declare const det12: (m: ArrayLike<number>) => number;
 /** Singular placement (determinant ~0, relative to the row lengths): BFC can't cull it (article 415 "Degenerate Matrices"). */
 export declare const singular12: (m: ArrayLike<number>, d?: number) => boolean;
-/** Open-addressing hash table from int32 triples to int32 values (-1 = absent); grows as needed. */
+/** Open-addressing hash table from int32 triples to int32 values (-1 = absent); grows as needed. Keys and value sit interleaved in one
+ *  Int32Array (x, y, z, value), and a byte per slot holds 8 more bits of the key's hash (0 = empty), so a miss usually reads only the
+ *  byte array and a hit one cache line of the keys, where five parallel arrays cost five. With `shift` >= 0, keys close in x / z (a
+ *  4 x 4 block of (1 << shift)-unit squares) start in neighbouring slots, so the studs of one brick, or the cells under it, share cache
+ *  lines. The stud-top table of a 400k-brick city has millions of entries and its lookups were the culling pass's biggest cost
+ *  (Round 73). */
 export declare class Int3Table {
-    a: Int32Array;
-    b: Int32Array;
-    c: Int32Array;
-    vals: Int32Array;
-    used: Uint8Array;
+    t: Int32Array;
+    tag: Uint8Array;
     mask: number;
     size: number;
-    constructor(cap?: number);
+    shift: number;
+    constructor(cap?: number, shift?: number);
     private slot;
+    private static tagOf;
     get(x: number, y: number, z: number): number;
     /** Sets (x, y, z) -> v and returns the previous value (-1 if none). */
     put(x: number, y: number, z: number, v: number): number;