ntk 8.17.7 → 8.18.0

package/README.md

@@ -144,6 +144,10 @@ parser, say — can import just that piece, and loads only what it needs:
 Each is the same module the root re-exports, so importing a name both ways
 gives the same function or class: nothing is loaded, or instantiated, twice.
 
+`ntk/package.json` is exported too, for the tools that read a dependency's
+version or metadata that way — bundlers, React Native's resolver, version
+reporters. It loads nothing.
+
 ## 3d graphics
 
 Two backends, chosen by `glPolicy`

package/lib/fontconfig.js

@@ -397,7 +397,7 @@ function censusBytes() {
   const deadline = performance.now() + PREWARM_WAIT_MS;
   let status = prewarmStatus(census.base);
   while (status === null && !census.exited && performance.now() < deadline) {
-    Atomics.wait(nap, 0, 0, 1);
+    nap();
     status = prewarmStatus(census.base);
   }
   census.bytes = null;
@@ -818,8 +818,16 @@ function spawnToFiles(fcs, bests = []) {
   return entries;
 }
 
-// the synchronous path's nap while a prewarm finishes
-const nap = new Int32Array(new SharedArrayBuffer(4));
+// the synchronous path's nap while a prewarm finishes: a millisecond's
+// Atomics.wait on a cell made by the first nap. Not at load: a page that is
+// not cross-origin isolated has no SharedArrayBuffer, so one made here was a
+// ReferenceError in every browser bundle of ntk, which never naps — there is
+// no prewarm without node (NOT_NODE)
+let napCell;
+function nap() {
+  napCell ??= new Int32Array(new SharedArrayBuffer(4));
+  Atomics.wait(napCell, 0, 0, 1);
+}
 
 // How long the synchronous path waits on a prewarm before asking fc-match
 // itself. fc-match answers in tens of milliseconds, and a prewarm that is
@@ -836,7 +844,7 @@ function answerSync(fc, entry) {
   let left = readPrewarm(entry.base);
   // a child that is gone has left all it ever will
   while (left === null && !entry.exited && performance.now() < deadline) {
-    Atomics.wait(nap, 0, 0, 1);
+    nap();
     left = readPrewarm(entry.base);
   }
   if (inflight.get(fc) === entry) inflight.delete(fc);
@@ -866,7 +874,7 @@ function firstOfPrewarm(spec, entry) {
   const deadline = performance.now() + PREWARM_WAIT_MS;
   let status = prewarmStatus(entry.base);
   while (status === null && !entry.exited && performance.now() < deadline) {
-    Atomics.wait(nap, 0, 0, 1);
+    nap();
     status = prewarmStatus(entry.base);
   }
   // not answered, or not yet: the whole read is what decides, and reports
@@ -893,7 +901,7 @@ function bestOfPrewarm(spec, entry) {
   const deadline = performance.now() + PREWARM_WAIT_MS;
   let left = readPrewarm(entry.base);
   while (left === null && !entry.exited && performance.now() < deadline) {
-    Atomics.wait(nap, 0, 0, 1);
+    nap();
     left = readPrewarm(entry.base);
   }
   if (left === null) return null;

package/lib/index.js

@@ -47,6 +47,7 @@ import {
   defaultRasterizer,
   setDefaultRasterizer
 } from './rasterize.js';
+import { PreciseRasterizer } from './precise.js';
 import { DEFAULT_MASK_POLICY } from './maskcluster.js';
 import {
   DEFAULT_SHADOW_POLICY,
@@ -284,6 +285,7 @@ export {
   setDefaultFontSource,
   CoverageAccumulator,
   ScanlineRasterizer,
+  PreciseRasterizer,
   defaultRasterizer,
   setDefaultRasterizer,
   DEFAULT_RASTER_POLICY,

package/lib/precise.js

@@ -0,0 +1,452 @@
+// RENDER's Precise rasterization, done here: the a8 coverage an X server
+// draws for a Triangles or AddTraps request, computed on this side of the
+// wire from the same numbers.
+//
+// A mask is rasterized either here or by the server (routeRaster in
+// rasterize.js). The analytic ScanlineRasterizer and the server's point
+// sampling come out a few levels apart at every edge, so whenever the same
+// pixels were drawn by different routes in different passes — a stroke cut
+// to the runs inside a clip against the whole of it, a stroke drawn straight
+// to the destination against the same stroke through the mask — a partial
+// repaint and a full one disagreed (issue #462). With this as the local
+// rasterizer they cannot: both routes are one function of the wire geometry,
+// and the route is a question of cost alone.
+//
+// RENDER's default poly-mode is Precise, and the protocol defines it
+// (renderproto, "Polygon Rasterization"): for an 8-bit alpha, a regular grid
+// 17 samples wide and 15 high, centred in the pixel; each trapezoid or
+// triangle is Add-combined into the mask. 17 × 15 = 255, so a pixel's alpha
+// is its count of covered samples. pixman implements it, and pixman is what
+// rasterizes Triangles, Trapezoids and AddTraps in fb (Xvfb, XQuartz) and in
+// glamor (Xorg's modesetting driver, Xwayland). This is a port of pixman's
+// code — pixman_edge_init and _step, pixman_sample_ceil_y and _floor_y,
+// rasterize_edges_8, triangle_to_trapezoids — down to its rounding. pixman
+// sits up to one 16.16 unit off the spec's sample positions, and its edge
+// step leaves the error term alone when a step carries nothing; matching it
+// is matching the servers, so both are kept. Measured byte-identical against
+// XQuartz on 300 random drawings, 24.5 M pixels.
+//
+// One liberty, which changes no byte: pixman adds each trapezoid into the
+// mask a sample row at a time with a saturating add. Every count it adds is
+// non-negative, so the saturating sum is the plain sum clamped once. Each
+// sample row's span goes into a difference array instead — four cells,
+// however wide the span — and each pixel row is integrated and clamped at
+// the end.
+//
+// The input is the wire's integers. `fixed16` is what a coordinate becomes
+// in a request, and the 2d context sends coordinates already on that grid
+// (`snapFixed16`), so that the conversion node-x11 applies has nothing left
+// to round.
+
+import { trapezoidize } from './trapezoid.js';
+
+// the sample grid for an 8-bit alpha, as pixman-private.h derives it
+const STEP_Y_SMALL = 4369; // 65536 / 15
+const STEP_Y_BIG = 4370; // 65536 - 14 * STEP_Y_SMALL
+const Y_FRAC_FIRST = 2185; // STEP_Y_BIG / 2
+const Y_FRAC_LAST = 63351; // Y_FRAC_FIRST + 14 * STEP_Y_SMALL
+const STEP_X_SMALL = 3855; // 65536 / 17
+const X_FRAC_FIRST = 1928; // (65536 - 16 * STEP_X_SMALL) / 2
+const N_X_FRAC = 17;
+
+/**
+ * A coordinate as the 16.16 fixed-point int32 a request carries: truncated
+ * toward zero, as node-x11 converts. Every coordinate the 2d context sends
+ * is in reach of the int32 (lib/cliprect.js's WIRE_REACH).
+ */
+export function fixed16(v) {
+  return Math.trunc(v * 65536) | 0;
+}
+
+/**
+ * The nearest coordinate toward zero that 16.16 holds exactly. Sending this
+ * rather than `v` makes the request's integer `fixed16(v)` however the
+ * client library converts — node-x11's `parseInt(v * 65536)` reads a tiny
+ * product's exponent notation, so 4e-11 became 4 where it should be 0.
+ */
+export function snapFixed16(v) {
+  return Math.trunc(v * 65536) / 65536;
+}
+
+// C's integer division, which truncates toward zero, exact for |a| < 2^53.
+// A float quotient can land on the wrong side of an integer, so the
+// remainder settles it.
+function idiv(a, b) {
+  let q = Math.trunc(a / b);
+  const r = a - q * b;
+  if (a >= 0) {
+    if (r < 0) q--;
+    else if (r >= b) q++;
+  } else if (r > 0) q++;
+  else if (r <= -b) q--;
+  return q;
+}
+
+// pixman's DIV: division rounding toward -infinity, for b > 0
+function floorDiv(a, b) {
+  return a < 0 ? idiv(a - b + 1, b) : idiv(a, b);
+}
+
+/** pixman_sample_ceil_y: the first sample row at or below y */
+function sampleCeilY(y) {
+  let f = y & 0xffff;
+  let i = y - f;
+  f = floorDiv(f - Y_FRAC_FIRST + (STEP_Y_SMALL - 1), STEP_Y_SMALL) * STEP_Y_SMALL + Y_FRAC_FIRST;
+  if (f > Y_FRAC_LAST) {
+    if (i >> 16 === 0x7fff) {
+      f = 0xffff;
+    } else {
+      f = Y_FRAC_FIRST;
+      i += 65536;
+    }
+  }
+  return (i | f) | 0;
+}
+
+/** pixman_sample_floor_y: the last sample row strictly above y */
+function sampleFloorY(y) {
+  let f = y & 0xffff;
+  let i = y - f;
+  f = floorDiv(f - 1 - Y_FRAC_FIRST, STEP_Y_SMALL) * STEP_Y_SMALL + Y_FRAC_FIRST;
+  if (f < Y_FRAC_FIRST) {
+    if (i >> 16 === -32768) {
+      f = 0;
+    } else {
+      f = Y_FRAC_LAST;
+      i -= 65536;
+    }
+  }
+  return (i | f) | 0;
+}
+
+// pixman_edge_t, as eight int32 fields
+const X = 0;
+const E = 1;
+const DY = 2;
+const SIGNDX = 3;
+const STEPX_SMALL = 4;
+const DX_SMALL = 5;
+const STEPX_BIG = 6;
+const DX_BIG = 7;
+const left = new Int32Array(8);
+const right = new Int32Array(8);
+
+/**
+ * pixman_edge_init followed by its pixman_edge_step to `yStart`: the edge
+ * from (xTop, yTop) down to (xBot, yBot), placed on its first sample row.
+ */
+function edgeInit(edge, yStart, xTop, yTop, xBot, yBot) {
+  const dx = (xBot - xTop) | 0;
+  const dy = (yBot - yTop) | 0;
+  let x = xTop;
+  let e = 0;
+  let stepx = 0;
+  let signdx = 0;
+  let rem = 0;
+  if (dy) {
+    if (dx >= 0) {
+      signdx = 1;
+      stepx = idiv(dx, dy);
+      rem = dx % dy;
+      e = -dy;
+    } else {
+      signdx = -1;
+      stepx = -idiv(-dx, dy);
+      rem = -dx % dy;
+    }
+    // _pixman_edge_multi_init for the small and the big step; n is at most
+    // 4370, so every product here is exact in a double
+    let ne = STEP_Y_SMALL * rem;
+    let sx = (STEP_Y_SMALL * stepx) | 0;
+    if (ne > 0) {
+      const nx = idiv(ne, dy) | 0;
+      ne -= nx * dy;
+      sx = (sx + nx * signdx) | 0;
+    }
+    edge[STEPX_SMALL] = sx;
+    edge[DX_SMALL] = ne;
+    ne = STEP_Y_BIG * rem;
+    sx = (STEP_Y_BIG * stepx) | 0;
+    if (ne > 0) {
+      const nx = idiv(ne, dy) | 0;
+      ne -= nx * dy;
+      sx = (sx + nx * signdx) | 0;
+    }
+    edge[STEPX_BIG] = sx;
+    edge[DX_BIG] = ne;
+  }
+
+  // pixman_edge_step(e, yStart - yTop). Its products are 64-bit in C and
+  // pass 2^53 for a tall edge entering the grid far below its top, so past
+  // that they are done in BigInt.
+  const n = (yStart - yTop) | 0;
+  const step = n * stepx;
+  x = Number.isSafeInteger(step)
+    ? (x + step) | 0
+    : Number(BigInt.asIntN(32, BigInt(x) + BigInt(n) * BigInt(stepx)));
+  const carry = n * rem;
+  if (Number.isSafeInteger(carry) && Number.isSafeInteger(e + carry + dy)) {
+    const ne = e + carry;
+    if (n >= 0) {
+      if (ne > 0) {
+        const nx = idiv(ne + dy - 1, dy) | 0;
+        e = (ne - nx * dy) | 0;
+        x = (x + nx * signdx) | 0;
+      }
+    } else if (ne <= -dy) {
+      const nx = idiv(-ne, dy) | 0;
+      e = (ne + nx * dy) | 0;
+      x = (x - nx * signdx) | 0;
+    }
+  } else {
+    const ne = BigInt(e) + BigInt(n) * BigInt(rem);
+    const bdy = BigInt(dy);
+    if (n >= 0) {
+      if (ne > 0n) {
+        const nx = BigInt.asIntN(32, (ne + bdy - 1n) / bdy);
+        e = Number(BigInt.asIntN(32, ne - nx * bdy));
+        x = (x + Number(nx) * signdx) | 0;
+      }
+    } else if (ne <= -bdy) {
+      const nx = BigInt.asIntN(32, -ne / bdy);
+      e = Number(BigInt.asIntN(32, ne + nx * bdy));
+      x = (x - Number(nx) * signdx) | 0;
+    }
+  }
+  edge[X] = x;
+  edge[E] = e;
+  edge[DY] = dy;
+  edge[SIGNDX] = signdx;
+}
+
+/**
+ * rasterize_edges_8 between `left` and `right`, from sample row `t` to `b`
+ * inclusive, into the difference array `acc` (rows `stride` apart).
+ */
+function addSpans(acc, stride, width, t, b) {
+  let y = t;
+  let row = (y >> 16) * stride;
+  let lx = left[X];
+  let le = left[E];
+  const ldy = left[DY];
+  const lsign = left[SIGNDX];
+  const lsxs = left[STEPX_SMALL];
+  const ldxs = left[DX_SMALL];
+  const lsxb = left[STEPX_BIG];
+  const ldxb = left[DX_BIG];
+  let rx = right[X];
+  let re = right[E];
+  const rdy = right[DY];
+  const rsign = right[SIGNDX];
+  const rsxs = right[STEPX_SMALL];
+  const rdxs = right[DX_SMALL];
+  const rsxb = right[STEPX_BIG];
+  const rdxb = right[DX_BIG];
+  // pixman clips the span to the image: from 0, and to the last pixel of
+  // the row taken whole. Both are what an unclipped span would add there.
+  const lastX = (width << 16) - 1;
+  for (;;) {
+    const l = lx < 0 ? 0 : lx;
+    const r = rx >> 16 >= width ? lastX : rx;
+    if (r > l) {
+      // the span adds 17 - ls to the pixel it starts in, 17 to each pixel
+      // after, rs to the pixel it ends in and nothing past it: the
+      // difference of two steps, each two cells of `acc`
+      const li = l >> 16;
+      const ri = r >> 16;
+      const ls = (((l & 0xffff) + X_FRAC_FIRST) / STEP_X_SMALL) | 0;
+      const rs = (((r & 0xffff) + X_FRAC_FIRST) / STEP_X_SMALL) | 0;
+      acc[row + li] += N_X_FRAC - ls;
+      acc[row + li + 1] += ls;
+      acc[row + ri] -= N_X_FRAC - rs;
+      acc[row + ri + 1] -= rs;
+    }
+    if (y === b) return;
+    if ((y & 0xffff) !== Y_FRAC_LAST) {
+      lx = (lx + lsxs) | 0;
+      le = (le + ldxs) | 0;
+      if (le > 0) {
+        le = (le - ldy) | 0;
+        lx = (lx + lsign) | 0;
+      }
+      rx = (rx + rsxs) | 0;
+      re = (re + rdxs) | 0;
+      if (re > 0) {
+        re = (re - rdy) | 0;
+        rx = (rx + rsign) | 0;
+      }
+      y = (y + STEP_Y_SMALL) | 0;
+    } else {
+      lx = (lx + lsxb) | 0;
+      le = (le + ldxb) | 0;
+      if (le > 0) {
+        le = (le - ldy) | 0;
+        lx = (lx + lsign) | 0;
+      }
+      rx = (rx + rsxb) | 0;
+      re = (re + rdxb) | 0;
+      if (re > 0) {
+        re = (re - rdy) | 0;
+        rx = (rx + rsign) | 0;
+      }
+      y = (y + STEP_Y_BIG) | 0;
+      row += stride;
+    }
+  }
+}
+
+/** the first sample row of a band from `top`, inside the grid */
+function firstRow(top) {
+  return sampleCeilY(top < 0 ? 0 : top);
+}
+
+/** the last sample row of a band to `bottom`, inside a grid `height` high */
+function lastRow(bottom, height) {
+  return sampleFloorY(bottom >> 16 >= height ? (height << 16) - 1 : bottom);
+}
+
+/**
+ * pixman_rasterize_trapezoid: a pixman_trapezoid_t — top and bottom, each
+ * side a line through two points — already offset into the grid
+ */
+function addTrapezoid(acc, stride, width, height, top, bottom, l1x, l1y, l2x, l2y, r1x, r1y, r2x, r2y) {
+  if (l1y === l2y || r1y === r2y || !(bottom > top)) return;
+  const t = firstRow(top);
+  const b = lastRow(bottom, height);
+  if (b < t) return;
+  if (l1y <= l2y) edgeInit(left, t, l1x, l1y, l2x, l2y);
+  else edgeInit(left, t, l2x, l2y, l1x, l1y);
+  if (r1y <= r2y) edgeInit(right, t, r1x, r1y, r2x, r2y);
+  else edgeInit(right, t, r2x, r2y, r1x, r1y);
+  addSpans(acc, stride, width, t, b);
+}
+
+// whether b is clockwise of a about ref in y-down space, as pixman asks it
+function clockwise(refX, refY, ax, ay, bx, by) {
+  const adx = ax - refX;
+  const ady = ay - refY;
+  const bdx = bx - refX;
+  const bdy = by - refY;
+  const p = bdy * adx;
+  const q = ady * bdx;
+  if (Number.isSafeInteger(p) && Number.isSafeInteger(q)) return p - q < 0;
+  return BigInt(bdy) * BigInt(adx) - BigInt(ady) * BigInt(bdx) < 0n;
+}
+
+/**
+ * Add a Triangles request's triangles — float device coordinates, six per
+ * triangle — to the grid, each split into two trapezoids as
+ * triangle_to_trapezoids does.
+ */
+function addTriangles(acc, stride, width, height, tris, dx, dy) {
+  const xo = dx * 65536;
+  const yo = dy * 65536;
+  for (let i = 0; i + 5 < tris.length; i += 6) {
+    let tx = (fixed16(tris[i]) + xo) | 0;
+    let ty = (fixed16(tris[i + 1]) + yo) | 0;
+    let lx = (fixed16(tris[i + 2]) + xo) | 0;
+    let ly = (fixed16(tris[i + 3]) + yo) | 0;
+    let rx = (fixed16(tris[i + 4]) + xo) | 0;
+    let ry = (fixed16(tris[i + 5]) + yo) | 0;
+    let s;
+    // the topmost point first, ties to the left; then left before right
+    if (ty === ly ? tx > lx : ty > ly) {
+      s = lx; lx = tx; tx = s;
+      s = ly; ly = ty; ty = s;
+    }
+    if (ty === ry ? tx > rx : ty > ry) {
+      s = rx; rx = tx; tx = s;
+      s = ry; ry = ty; ty = s;
+    }
+    if (clockwise(tx, ty, rx, ry, lx, ly)) {
+      s = rx; rx = lx; lx = s;
+      s = ry; ry = ly; ly = s;
+    }
+    // down from the top to the higher of the other two, then on to the
+    // lower, with the side that turned there continuing along the third edge
+    const mid = ry < ly ? ry : ly;
+    addTrapezoid(acc, stride, width, height, ty, mid, tx, ty, lx, ly, tx, ty, rx, ry);
+    if (ry < ly) addTrapezoid(acc, stride, width, height, ry, ly, tx, ty, lx, ly, rx, ry, lx, ly);
+    else addTrapezoid(acc, stride, width, height, ly, ry, lx, ly, rx, ry, tx, ty, rx, ry);
+  }
+}
+
+/**
+ * Add an AddTraps request's trapezoids — float device coordinates, six per
+ * trapezoid: top left, right, y, then bottom left, right, y — to the grid,
+ * as pixman_add_traps does.
+ */
+function addTraps(acc, stride, width, height, traps, dx, dy) {
+  const xo = dx * 65536;
+  const yo = dy * 65536;
+  for (let i = 0; i + 5 < traps.length; i += 6) {
+    const topY = (fixed16(traps[i + 2]) + yo) | 0;
+    const botY = (fixed16(traps[i + 5]) + yo) | 0;
+    const t = firstRow(topY);
+    const b = lastRow(botY, height);
+    if (b < t) continue;
+    const tl = (fixed16(traps[i]) + xo) | 0;
+    const tr = (fixed16(traps[i + 1]) + xo) | 0;
+    const bl = (fixed16(traps[i + 3]) + xo) | 0;
+    const br = (fixed16(traps[i + 4]) + xo) | 0;
+    edgeInit(left, t, tl, topY, bl, botY);
+    edgeInit(right, t, tr, topY, br, botY);
+    addSpans(acc, stride, width, t, b);
+  }
+}
+
+/**
+ * Each row's running sum, clamped: the saturating adds, all at once. The
+ * cells are zeroed as they are read, so the accumulator is ready for the
+ * next mask without a pass of its own.
+ */
+function integrate(acc, stride, width, height) {
+  const out = new Uint8Array(width * height);
+  for (let y = 0; y < height; y++) {
+    let sum = 0;
+    const from = y * stride;
+    const to = y * width;
+    for (let x = 0; x < width; x++) {
+      sum += acc[from + x];
+      acc[from + x] = 0;
+      out[to + x] = sum > 255 ? 255 : sum;
+    }
+    // the slack cells past the row, which a span ending at its last pixel
+    // writes and nothing reads
+    acc[from + width] = 0;
+    acc[from + width + 1] = 0;
+  }
+  return out;
+}
+
+// One accumulator for every mask, grown to the largest asked for: masks are
+// rasterized one at a time, and allocating a fresh one each time was a
+// tenth of the cost of a wall of small icons, in the garbage collector.
+let scratch = new Int32Array(0);
+
+function accumulator(cells) {
+  if (scratch.length < cells) scratch = new Int32Array(Math.max(cells, scratch.length * 2));
+  return scratch;
+}
+
+/**
+ * A Rasterizer (docs/context-2d.md) whose masks are the ones the X server
+ * draws from the same geometry: `triangles` exactly as the 2d context sends
+ * them with Triangles, `polys` trapezoidized exactly as it sends them with
+ * AddTraps. A drawing gets the same bytes whichever side rasterizes it.
+ *
+ * `dx`/`dy` are whole pixels, which is all the 2d context passes; anything
+ * else is declined, and goes to the server.
+ */
+export class PreciseRasterizer {
+  rasterize({ polys, triangles, width, height, rule, dx = 0, dy = 0 }) {
+    if (!Number.isInteger(dx) || !Number.isInteger(dy)) return null;
+    // one cell of slack: a span ending in the last column clears the cell
+    // past it
+    const stride = width + 2;
+    const acc = accumulator(stride * height);
+    if (triangles) addTriangles(acc, stride, width, height, triangles, dx, dy);
+    else addTraps(acc, stride, width, height, trapezoidize(polys, 0, 0, [], rule), dx, dy);
+    return integrate(acc, stride, width, height);
+  }
+}

package/lib/rasterize.js

@@ -1,11 +1,14 @@
 // Pure-JS coverage rasterization: outlines and polygons in, 8-bit alpha out.
 //
-// Two consumers, one algorithm:
-//
 //   rasterizePath()   glyph outlines -> a8 bitmap for XRender AddGlyphs
-//   rasterizePolys()  device-space path/stroke geometry -> a8 coverage mask,
-//                     uploaded by the 2d context instead of asking the server
-//                     to rasterize trapezoids (see routeRaster below)
+//   rasterizePolys()  device-space polygons -> an a8 coverage mask
+//
+// The 2d context's masks are rasterized locally or by the server
+// (routeRaster below), locally by default with PreciseRasterizer
+// (lib/precise.js), which samples as the server does so that the route
+// never shows. The algorithm here is the analytic one: ScanlineRasterizer,
+// a Rasterizer that can still be plugged in, and what glyph bitmaps use,
+// since a glyph has no server route to agree with.
 //
 // The algorithm is signed-area accumulation, the one font-rs and
 // stb_truetype v2 use. Each edge deposits the exact area it sweeps through
@@ -15,6 +18,8 @@
 // the result is closer to a reference rasterizer than 4x4 supersampling was,
 // at a fraction of the cost (~25 microseconds for a 20px icon against ~600).
 
+import { PreciseRasterizer } from './precise.js';
+
 /**
  * Antialiased coverage accumulator over a width x height pixel grid.
  *
@@ -421,7 +426,12 @@ function curveSteps(...coords) {
 //     A rasterizer that only handles, say, non-zero fills can return null
 //     for everything else and stay correct.
 
-/** the in-repo rasterizer: CoverageAccumulator, no dependencies */
+/**
+ * The analytic rasterizer: CoverageAccumulator, exact area coverage. Its
+ * edges are a few levels away from the server's point sampling, so a drawing
+ * it rasterizes and one the server does can disagree where they meet — the
+ * reason the default is PreciseRasterizer (lib/precise.js).
+ */
 export class ScanlineRasterizer {
   rasterize({ polys, triangles, width, height, rule, dx = 0, dy = 0 }) {
     const acc = new CoverageAccumulator(width, height);
@@ -433,9 +443,12 @@ export class ScanlineRasterizer {
 
 let _defaultRasterizer = null;
 
-/** the process-wide default Rasterizer */
+/**
+ * The process-wide default Rasterizer: PreciseRasterizer, whose masks are
+ * byte-identical to the server's for the same geometry (lib/precise.js).
+ */
 export function defaultRasterizer() {
-  if (!_defaultRasterizer) _defaultRasterizer = new ScanlineRasterizer();
+  if (!_defaultRasterizer) _defaultRasterizer = new PreciseRasterizer();
   return _defaultRasterizer;
 }