@graphty/webgpu-graph-algorithms 0.5.0 → 0.6.0

package/src/wgsl/fa2-integrate.wgsl.ts

@@ -4,7 +4,12 @@
  * networkx mode `m |F|`, contract 4.6), the position update with no displacement clamp (D25), z never integrated
  * in 2D (spec 7.13), `oldForce` stored in paper mode for fixed nodes too (spec 7.11), and the partials A
  * (sum p, sum |p - c|^2, min, max with max |p - c|^2 in `max.w`) and C (sum |dp| over free rows, free count) in
- * uniform control flow.
+ * uniform control flow. `APPLY` (P5, spec 7.20) picks the integrator: 0 = the FA2 text, 1 = the Fruchterman-Reingold
+ * temperature cap `min(|F|, t)` along F, 2 = ngraph's semi-implicit Euler step with drag and the unit speed clamp
+ * over the velocity that lives in the `oldForce` slot (PD-2), which also folds the free kinetic energy into
+ * `partials.swingTraction.x` (PD-4); mode 1 folds the free force energy `sum |F|^2` into the same slot for K1's
+ * adaptive cooling and takes its temperature from the state block when `FA2_FLAG_ADAPTIVE` is set. The energy
+ * reduction runs under every mode; only its write is conditional.
  *
  * Body only (spec 3.5, D9); normative text (contract 4.5); the K5 sabotage mutations (P3-T5) are textual edits of it.
  */
@@ -25,6 +30,7 @@ fn integrate(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_id
     var p = vec4f(0.0);
     var free = false;
     var valid = false;
+    var ke = 0.0;
     if (i < P.n) {
         valid = true;
         let f = load_force(i);
@@ -33,7 +39,25 @@ fn integrate(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_id
         if (SWING_MODE == 0u) { swing_i = p.w * length(f - load_old(i)); }  // paper: m |F(t) - F(t-1)|, recomputed inline (7.2)
         let factor = S.speed / (1.0 + sqrt(S.speed * swing_i));
         let fixed = mask_bit(fixedMask[i >> 5u], i);
-        dp = select(f * factor, vec3f(0.0), fixed);                        // no clamp on dp (D25)
+        dp = select(f * factor, vec3f(0.0), fixed);                        // APPLY 0 (FA2): no clamp on dp (D25)
+        if (APPLY == 1u) {                                                  // APPLY 1 (FR, 7.20): move along F by min(|F|, t); a fixed node stays
+            let mag = length(f);
+            let t = select(P.temperature, S.temperature, (P.flags & FA2_FLAG_ADAPTIVE) != 0u);   // adaptive cooling: K1's state temperature
+            dp = vec3f(0.0);
+            if (mag > 0.0 && !fixed) { dp = f * (min(mag, t) / mag); }
+            ke = select(0.0, dot(f, f), !fixed);                              // the force energy of Hu's step control, folded like the spring preset's kinetic energy
+        }
+        if (APPLY == 2u) {                                                  // APPLY 2 (spring-electrical): ngraph's Euler step over the velocity in the oldForce slot (PD-2)
+            var v = load_old(i);
+            let fd = f - P.dragCoefficient * v;                             // drag (generateCreateDragForce.js:18)
+            v = v + (P.timeStep / p.w) * fd;                                 // v += (dt / m) F (generateIntegrator.js:27-29)
+            let sp = length(v);
+            if (sp > 1.0) { v = v / sp; }                                     // the unit speed clamp (generateIntegrator.js:33-37)
+            if (P.dim == 2u) { v.z = 0.0; }
+            dp = select(P.timeStep * v, vec3f(0.0), fixed);                  // dp = dt v; a pinned body is skipped (generateIntegrator.js:21, 39-41)
+            if (!fixed) { store_old(i, v); }
+            ke = select(0.0, 0.5 * p.w * dot(v, v), !fixed);                 // partials B under APPLY 2 (PD-4)
+        }
         if (P.dim == 2u) { dp.z = 0.0; }                                   // 2D never integrates z (7.13)
         p = vec4f(p.xyz + dp, p.w);
         pos[i] = p;
@@ -59,11 +83,13 @@ fn integrate(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_id
     let tHi = wg_reduce_vec4(hi, lid.x, 2u);
     let tDl = wg_reduce_f32(dl, lid.x, 0u);
     let tFr = wg_reduce_u32(fr, lid.x, 0u);
+    let tKe = wg_reduce_f32(ke, lid.x, 0u);
     if (lid.x == 0u) {
         let g = group_id(wid);
         partials[g].sum = tSum;
         partials[g].min = tLo;
         partials[g].max = tHi;
         partials[g].dispFree = vec2f(tDl, f32(tFr));
+        if (APPLY != 0u) { partials[g].swingTraction = vec2f(tKe, 0.0); }   // overwrites K3's epilogue: K4 never runs under FR or the preset (PD-4)
     }
 }`;

package/src/wgsl/fa2-repulsion-exact.wgsl.ts

@@ -1,6 +1,8 @@
 /**
  * The `fa2-repulsion-exact` kernel body (K3; contract 4.5; spec 7.6, 7.9, 7.10): the tiled all-pairs repulsion
- * `|F| = k m_i m_j / d` with the 0.01 distance floor and the antisymmetric coincident kick, the gravity epilogue
+ * `|F| = k m_i m_j / d` with the 0.01 distance floor and the antisymmetric coincident kick (`LAW` 0; P5's `LAW` 1
+ * is Fruchterman-Reingold `k^2 / d` and `LAW` 2 ngraph's Coulomb `-g m_i m_j / d^2`, both unfloored, spec 7.20,
+ * with the kick's magnitude taken from the law at d = 0.01, PD-10), the gravity epilogue
  * (GRAVITY_CENTER 0 = centroid, 1 = origin; STRONG_GRAVITY) added under the `valid` guard, `force += f`, and the
  * swing / traction workgroup reduction in uniform control flow (SWING_MODE 0 = paper: free nodes only against
  * `oldForce`; 1 = NetworkX: positions and forces mixed, every node) whose lane 0 writes
@@ -18,6 +20,11 @@ fn store_force(i: u32, f: vec3f) {
     force[3u * i + 2u] = f.z;
 }
 fn load_old(i: u32) -> vec3f { return vec3f(oldForce[3u * i], oldForce[3u * i + 1u], oldForce[3u * i + 2u]); }
+fn kick_magnitude(mi: f32, mj: f32) -> f32 {                   // the law's magnitude at d = FA2_DIST_FLOOR (PD-10)
+    if (LAW == 1u) { return P.frK * P.frK / FA2_DIST_FLOOR; }
+    if (LAW == 2u) { return -P.coulomb * mi * mj / FA2_DIST_FLOOR_SQ; }
+    return P.scalingRatio * mi * mj / FA2_DIST_FLOOR;
+}
 fn gravity_force(pi: vec4f) -> vec3f {                         // spec 7.9: centroid (GRAVITY_CENTER 0) or origin (1); regular or strong
     var q = pi.xyz;
     if (GRAVITY_CENTER == 0u) { q = pi.xyz - S.centroid.xyz; }
@@ -45,13 +52,15 @@ fn repulsion(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_id
             if (o.w > 0.0 && jj != i) {                                           // mass > 0 for every real node, 0 for the pad
                 let d = pi.xyz - o.xyz;
                 var d2 = dot(d, d);
-                if (d2 < FA2_COINCIDENT_SQ) {                                     // coincident: antisymmetric unit kick of magnitude k m_i m_j / 0.01 (7.2)
-                    f = f + kick_dir(i, jj, P.dim) * (P.scalingRatio * pi.w * o.w / FA2_DIST_FLOOR);
+                if (d2 < FA2_COINCIDENT_SQ) {                                     // coincident: antisymmetric unit kick of the law's magnitude at d = 0.01 (7.2; PD-10)
+                    f = f + kick_dir(i, jj, P.dim) * kick_magnitude(pi.w, o.w);
                     continue;
                 }
-                d2 = max(d2, FA2_DIST_FLOOR_SQ);                                  // d >= 0.01
+                if (LAW == 0u) { d2 = max(d2, FA2_DIST_FLOOR_SQ); }               // FA2 alone floors d >= 0.01 (7.2); FR and coulomb are unfloored (7.20)
                 let k = P.scalingRatio * pi.w * o.w;
-                f = f + d * (k / d2);                                             // |F| = k m_i m_j / d along d / d
+                if (LAW == 0u) { f = f + d * (k / d2); }                          // LAW 0 (FA2): |F| = k m_i m_j / d along d / d
+                if (LAW == 1u) { f = f + d * (P.frK * P.frK / d2); }              // LAW 1 (FR, 7.20): |F| = k^2 / d, mass ignored
+                if (LAW == 2u) { f = f + d * (-P.coulomb * pi.w * o.w / (d2 * sqrt(d2))); }   // LAW 2 (coulomb, ngraph generateQuadTree.js:131-132): |F| = -g m_i m_j / d^2
             }
         }
         workgroupBarrier();

package/src/wgsl/fa2-stats-finalize.wgsl.ts

@@ -4,6 +4,17 @@
  * exact spec 3.3 value from `partials.max.w`), bounding box, mean displacement over free nodes (0 when every node
  * is fixed), the settle counter -- increments the iteration counter and writes the K1 half of the trace record.
  * On the first iteration after load() (`FA2_FLAG_FIRST`) it folds nothing and keeps the host-written state.
+ * `STATS_MODE` (P5, spec 7.20) adds the model statistic: 0 = the FA2 text, 1 = the Fruchterman-Reingold temperature
+ * of this iteration into `S.temperature` and the trace's `modelScalar` -- the uniform's under the linear schedule,
+ * or, with `FA2_FLAG_ADAPTIVE` set, the adaptive one: the previous iteration's force energy (K5 folds sum |F|^2 over
+ * free nodes into `partials.swingTraction.x`) against `S.frEnergy` grows the temperature by 1 / FR_COOLING_STEP after
+ * FR_COOLING_PATIENCE consecutive falls and shrinks it by FR_COOLING_STEP on a rise (Yifan Hu 2005, section 3.2);
+ * 2 = the spring-electrical kinetic energy K5 folded into `partials.swingTraction.x` (PD-4) into `S.kineticEnergy`
+ * and the trace. On the grid tier (`P.gridMax > 0`, P4-T10, PD-14) it also derives the grid frame of the next build
+ * from the fold (`extent = max(min(bboxExtent * GRID_BBOX_MARGIN, extentFactor * rmsRadius), GRID_EXTENT_FLOOR)`,
+ * `cellSize = extent / G`, `gridMin = centroid - extent / 2` with `cellSize` in `.w`, `invCellSize`, `eps = 0.25
+ * cellSize`), copies the previous iteration's pseudo-cell count and occupancy max into the state, and resets the
+ * hub counters; the exact tier writes `gridMax: 0` and binds two dummies, so the block is dead there.
  *
  * This file holds the kernel BODY only (spec 3.5, D9): no bind-group lines and no `override` lines -- the composer
  * emits them from the registry entry in src/kernels.ts (contract 3.10.1). The text is normative (contract 4.5) and
@@ -21,6 +32,7 @@ fn stats_finalize(@builtin(local_invocation_id) lid: vec3<u32>) {
     var hi = vec4f(-F32_MAX);
     var disp = 0.0;
     var free = 0u;
+    var ke = 0.0;
     if (fold) {
         for (var g = lid.x; g < groups; g = g + WG) {   // sequential per lane in index order: deterministic
             let q = partials[g];
@@ -29,6 +41,7 @@ fn stats_finalize(@builtin(local_invocation_id) lid: vec3<u32>) {
             hi = max(hi, q.max);
             disp = disp + q.dispFree.x;
             free = free + u32(q.dispFree.y);
+            ke = ke + q.swingTraction.x;
         }
     }
     let tSum = wg_reduce_vec4(sum, lid.x, 0u);
@@ -36,6 +49,7 @@ fn stats_finalize(@builtin(local_invocation_id) lid: vec3<u32>) {
     let tHi = wg_reduce_vec4(hi, lid.x, 2u);
     let tDisp = wg_reduce_f32(disp, lid.x, 0u);
     let tFree = wg_reduce_u32(free, lid.x, 0u);
+    let tKe = wg_reduce_f32(ke, lid.x, 0u);
     if (lid.x == 0u) {
         if (fold) {
             let n = f32(P.n);
@@ -53,5 +67,45 @@ fn stats_finalize(@builtin(local_invocation_id) lid: vec3<u32>) {
         T[P.iterationIndex].meanDisplacement = S.meanDisplacement;
         T[P.iterationIndex].settledCount = S.settledCount;
         T[P.iterationIndex].iteration = S.iteration;
+        if (P.gridMax > 0u) {                                      // the grid tier (7.7): the robust extent, the cell size, eps, last iteration's counts, the hub counter reset (PD-14)
+            let cells = P.gridMax * P.gridMax * select(1u, P.gridMax, P.dim == 3u);
+            if (fold) {
+                let box = (S.max.xyz - S.min.xyz) * GRID_BBOX_MARGIN;
+                var bboxExtent = max(box.x, box.y);
+                if (P.dim == 3u) { bboxExtent = max(bboxExtent, box.z); }
+                let extent = max(min(bboxExtent, P.extentFactor * S.rmsRadius), GRID_EXTENT_FLOOR);   // min(bbox, extentFactor x rms), floored (7.7)
+                let cellSize = extent / f32(P.gridMax);
+                S.gridMin = vec4f(S.centroid.xyz - vec3f(0.5 * extent), cellSize);   // gridMin.w carries cellSize
+                S.invCellSize = 1.0 / cellSize;
+                S.eps = 0.25 * cellSize;
+            }
+            S.outsideGrid = cellHist[cells];                         // the previous iteration's pseudo-cell count (0 after load)
+            S.maxCellOccupancy = atomicLoad(&hubCounters[1]);
+            atomicStore(&hubCounters[0], 0u);
+            atomicStore(&hubCounters[1], 0u);
+        }
+        if (STATS_MODE == 1u) {                                    // FR: this iteration's temperature (7.20) into the state and the trace
+            if ((P.flags & FA2_FLAG_ADAPTIVE) != 0u) {                // adaptive cooling (Yifan Hu 2005 3.2): tKe is the previous iteration's sum |F|^2 over free nodes
+                if (fold) {
+                    var t = S.temperature;
+                    if (tKe < S.frEnergy) {
+                        S.frProgress = S.frProgress + 1u;
+                        if (S.frProgress >= FR_COOLING_PATIENCE) { S.frProgress = 0u; t = t / FR_COOLING_STEP; }
+                    } else {
+                        S.frProgress = 0u;
+                        t = t * FR_COOLING_STEP;
+                    }
+                    S.frEnergy = tKe;
+                    S.temperature = t;
+                }
+            } else {
+                S.temperature = P.temperature;
+            }
+            T[P.iterationIndex].modelScalar = S.temperature;
+        }
+        if (STATS_MODE == 2u) {                                    // spring-electrical: the kinetic energy K5 folded into partials B (PD-4); 0 on the first iteration after load()
+            S.kineticEnergy = tKe;
+            T[P.iterationIndex].modelScalar = tKe;
+        }
     }
 }`;

package/src/wgsl/grid-cell-key.wgsl.ts

@@ -0,0 +1,29 @@
+/**
+ * G1, the `grid-cell-key` kernel body (spec 7.7; P4-T8): the finest cell of every node from the state's robust extent,
+ * `floor((p - gridMin) * invCellSize)` (a multiply, correctly rounded everywhere: PD-10), linearised when every axis
+ * is in [0, G) and the outside pseudo-cell `cells` otherwise; `cellVal[i] = i`. The clamp before the floor keeps a
+ * far-away or NaN coordinate out of an out-of-range float-to-int conversion. Body only; normative text.
+ */
+export const gridCellKeyWgsl = /* wgsl */ `
+fn grid_cells() -> u32 { return P.gridMax * P.gridMax * select(1u, P.gridMax, P.dim == 3u); }
+
+@compute @workgroup_size(WG)
+fn grid_cell_key(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_id) lid: vec3<u32>) {
+    let i = linear_id(wid, lid.x);
+    if (i >= P.n) { return; }                                      // no barrier follows
+    let cells = grid_cells();
+    let gf = f32(P.gridMax);
+    let q = (pos[i].xyz - S.gridMin.xyz) * S.invCellSize;          // PD-10: never a division
+    let c = vec3<i32>(floor(clamp(q, vec3f(-1.0), vec3f(gf + 1.0))));
+    let g = i32(P.gridMax);
+    var inside = c.x >= 0 && c.x < g && c.y >= 0 && c.y < g;
+    if (P.dim == 3u) { inside = inside && c.z >= 0 && c.z < g; }
+    var key = cells;                                               // the outside pseudo-cell (7.7)
+    if (inside) {
+        key = u32(c.x) + P.gridMax * u32(c.y);
+        if (P.dim == 3u) { key = key + P.gridMax * P.gridMax * u32(c.z); }
+    }
+    cellKey[i] = key;
+    cellVal[i] = i;
+}
+`;

package/src/wgsl/grid-centroid-hub.wgsl.ts

@@ -0,0 +1,28 @@
+/**
+ * G4b, the `grid-centroid-hub` kernel body (spec 7.7; P4-T9): one workgroup per hub cell of hubList, dispatched
+ * indirectly from hubArgs (the T1 finalize over hubCounters[0]); a WG-strided mass-weighted sum reduced by the
+ * prelude's tree. The work is guarded by `valid`, never an early return, so the reduction is uniform (PD-13). Body
+ * only; normative text.
+ */
+export const gridCentroidHubWgsl = /* wgsl */ `
+@compute @workgroup_size(WG)
+fn grid_centroid_hub(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_id) lid: vec3<u32>) {
+    let h = group_id(wid);
+    let valid = h < hubCount[0];                                   // a workgroup past the count sums nothing
+    var c = 0u;
+    var start = 0u;
+    var count = 0u;
+    if (valid) {
+        c = hubList[h];
+        start = cellStart[c];
+        count = cellStart[c + 1u] - start;
+    }
+    var acc = vec4f(0.0);
+    for (var k = start + lid.x; k < start + count; k = k + WG) {   // strided over the cell's sorted range
+        let p = pos[sortedIdx[k]];
+        acc = acc + vec4f(p.xyz * p.w, p.w);
+    }
+    let t = wg_reduce_vec4(acc, lid.x, 0u);                        // uniform control flow: 256 -> 1
+    if (valid && lid.x == 0u) { pyramid[c] = t; }
+}
+`;

package/src/wgsl/grid-centroid.wgsl.ts

@@ -0,0 +1,28 @@
+/**
+ * G4, the `grid-centroid` kernel body (spec 7.7; P4-T9): thread per finest cell, the pseudo-cell included; the
+ * mass-weighted position sum of a cell's sorted range in index order (no atomics: deterministic), the largest
+ * occupancy into hubCounters[1], and cells above GRID_HUB_CELL entries appended to hubList for G4b (PD-13). Body
+ * only; normative text.
+ */
+export const gridCentroidWgsl = /* wgsl */ `
+fn grid_cells() -> u32 { return P.gridMax * P.gridMax * select(1u, P.gridMax, P.dim == 3u); }
+
+@compute @workgroup_size(WG)
+fn grid_centroid(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_id) lid: vec3<u32>) {
+    let c = linear_id(wid, lid.x);
+    if (c > grid_cells()) { return; }                              // cells [0, cells]: the pseudo-cell is index cells; no barrier follows
+    let start = cellStart[c];
+    let count = cellStart[c + 1u] - start;
+    atomicMax(&hubCounters[1], count);                             // maxCellOccupancy, read by K1 next iteration
+    if (count > GRID_HUB_CELL) {                                   // a hub cell: G4b sums it (PD-13)
+        hubList[atomicAdd(&hubCounters[0], 1u)] = c;
+        return;
+    }
+    var acc = vec4f(0.0);
+    for (var k = start; k < start + count; k = k + 1u) {           // sorted order: deterministic
+        let p = pos[sortedIdx[k]];
+        acc = acc + vec4f(p.xyz * p.w, p.w);                       // (sum m x, sum m y, sum m z, sum m)
+    }
+    pyramid[c] = acc;
+}
+`;

package/src/wgsl/grid-downsample.wgsl.ts

@@ -0,0 +1,27 @@
+/**
+ * G5, the `grid-downsample` kernel body (spec 7.7; P4-T9): one dispatch per coarser level; every parent cell is the
+ * sum of its 4 (2D) or 8 (3D) children at the level below, read at P.childBase and written at P.parentBase (the
+ * pseudo-cell, index cells of level 0, is never a child). No atomics. Body only; normative text.
+ */
+export const gridDownsampleWgsl = /* wgsl */ `
+@compute @workgroup_size(WG)
+fn grid_downsample(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_id) lid: vec3<u32>) {
+    let pc = linear_id(wid, lid.x);                                // the parent cell inside its level
+    if (pc >= P.parentCells) { return; }                           // no barrier follows
+    let side = P.parentSide;
+    let cs = 2u * side;                                            // the child level's side
+    let px = pc % side;
+    let py = (pc / side) % side;
+    let pz = pc / (side * side);
+    var acc = vec4f(0.0);
+    for (var dz = 0u; dz < P.depth; dz = dz + 1u) {
+        for (var dy = 0u; dy < 2u; dy = dy + 1u) {
+            for (var dx = 0u; dx < 2u; dx = dx + 1u) {
+                let child = (2u * px + dx) + cs * ((2u * py + dy) + cs * (2u * pz + dz));
+                acc = acc + pyramid[P.childBase + child];
+            }
+        }
+    }
+    pyramid[P.parentBase + pc] = acc;
+}
+`;

package/src/wgsl/grid-far-field.wgsl.ts

@@ -0,0 +1,97 @@
+/**
+ * G6, the `grid-far-field` kernel body (spec 7.7; P4-T10; D24): per node `i = sortedIdx[t]`, its finest cell
+ * recomputed from `pos[i]` and the state (PD-10); for an inside node the coarsest level minus the 3x3 (3x3x3)
+ * around its coarsest cell, then at every finer level the 6x6 (6x6x6) block that is the parent's 3x3 minus this
+ * level's own 3x3 -- space tiled exactly once, no theta -- plus the outside pseudo-cell's centroid; for an outside
+ * node the coarsest level in full and no pseudo-cell. Every cell term is the per-cell law on the mass-weighted
+ * centroid (Gephi Region semantics), softened by `eps^2`: `LAW` 0 (FA2) `d * (k m_i M / d2)`, `LAW` 1 (FR, 7.20)
+ * `d * (k^2 M / d2)` (mass 1 per node, so `M` is the cell's count), `LAW` 2 (coulomb) `d * (-g m_i M / d2^1.5)`
+ * (P4-T13, PD-22). `force += f` (K2 wrote it). The loop bounds are `P.levels` and `P.gridMax` from the uniform,
+ * not a `LEVELS` override (PD-16, DEP-P4-G). Body only; normative text.
+ */
+export const gridFarFieldWgsl = /* wgsl */ `
+fn load_force(i: u32) -> vec3f { return vec3f(force[3u * i], force[3u * i + 1u], force[3u * i + 2u]); }
+fn store_force(i: u32, f: vec3f) {
+    force[3u * i] = f.x;
+    force[3u * i + 1u] = f.y;
+    force[3u * i + 2u] = f.z;
+}
+fn grid_cells() -> u32 { return P.gridMax * P.gridMax * select(1u, P.gridMax, P.dim == 3u); }
+fn grid_side(level: u32) -> u32 { return P.gridMax >> level; }
+fn level_base(level: u32) -> u32 {                             // the pyramid index of level L's cell 0 (level 0 carries the pseudo-cell at index cells)
+    var base = 0u;
+    for (var l = 0u; l < level; l = l + 1u) {
+        let s = grid_side(l);
+        base = base + s * s * select(1u, s, P.dim == 3u) + select(0u, 1u, l == 0u);
+    }
+    return base;
+}
+fn cell_at(level: u32, cx: i32, cy: i32, cz: i32) -> u32 {
+    let s = grid_side(level);
+    return level_base(level) + u32(cx) + s * (u32(cy) + select(0u, s * u32(cz), P.dim == 3u));
+}
+fn cell_force(pi: vec4f, q: vec4f) -> vec3f {                  // one far-field term, softened by state.eps (7.7)
+    if (q.w <= 0.0) { return vec3f(0.0); }                     // an empty cell
+    let d = pi.xyz - q.xyz / q.w;                              // to the mass-weighted centroid
+    let d2 = dot(d, d) + S.eps * S.eps;
+    if (LAW == 1u) { return d * (P.frK * P.frK * q.w / d2); }                       // LAW 1 (FR, 7.20): k^2 / d per node, q.w nodes at the centroid
+    if (LAW == 2u) { return d * (-P.coulomb * pi.w * q.w / (d2 * sqrt(d2))); }     // LAW 2 (coulomb): -g m_i M_cell / d^2
+    return d * (P.scalingRatio * pi.w * q.w / d2);             // LAW 0 (FA2): |F| = k m_i M_cell / d
+}
+
+@compute @workgroup_size(WG)
+fn grid_far_field(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_id) lid: vec3<u32>) {
+    let t = linear_id(wid, lid.x);
+    if (t >= P.n) { return; }                                      // no barrier follows
+    let i = sortedIdx[t];                                          // sorted order (D24)
+    let pi = pos[i];
+    let gf = f32(P.gridMax);
+    let q = (pi.xyz - S.gridMin.xyz) * S.invCellSize;              // PD-10
+    var c0 = vec3<i32>(floor(clamp(q, vec3f(-1.0), vec3f(gf + 1.0))));
+    if (P.dim == 2u) { c0.z = 0; }                                 // 2D: one z plane; the loops below visit cz = 0 only, so the 3x3 test must see cz - 0
+    let g = i32(P.gridMax);
+    var inside = c0.x >= 0 && c0.x < g && c0.y >= 0 && c0.y < g;
+    if (P.dim == 3u) { inside = inside && c0.z >= 0 && c0.z < g; }
+    let top = P.levels - 1u;
+    let ts = i32(grid_side(top));                                  // the coarsest side (4)
+    let zTop = select(0, ts - 1, P.dim == 3u);                     // z ranges: one plane in 2D
+    var f = vec3f(0.0);
+    if (inside) {
+        let ct = c0 / i32(1u << top);                              // the node's coarsest cell
+        for (var cz = 0; cz <= zTop; cz = cz + 1) {
+            for (var cy = 0; cy < ts; cy = cy + 1) {
+                for (var cx = 0; cx < ts; cx = cx + 1) {
+                    if (abs(cx - ct.x) <= 1 && abs(cy - ct.y) <= 1 && abs(cz - ct.z) <= 1) { continue; }   // the 3x3(x3) is finer levels' work
+                    f = f + cell_force(pi, pyramid[cell_at(top, cx, cy, cz)]);
+                }
+            }
+        }
+        for (var l = top; l > 0u; l = l - 1u) {                    // level l - 1: the parent's 3x3 at level l, refined, minus this level's own 3x3
+            let level = l - 1u;
+            let cl = c0 / i32(1u << level);
+            let cp = cl / 2;
+            let side = i32(grid_side(level));
+            let zLo = select(0, max(0, 2 * (cp.z - 1)), P.dim == 3u);
+            let zHi = select(0, min(side - 1, 2 * (cp.z + 1) + 1), P.dim == 3u);
+            for (var cz = zLo; cz <= zHi; cz = cz + 1) {
+                for (var cy = max(0, 2 * (cp.y - 1)); cy <= min(side - 1, 2 * (cp.y + 1) + 1); cy = cy + 1) {
+                    for (var cx = max(0, 2 * (cp.x - 1)); cx <= min(side - 1, 2 * (cp.x + 1) + 1); cx = cx + 1) {
+                        if (abs(cx - cl.x) <= 1 && abs(cy - cl.y) <= 1 && abs(cz - cl.z) <= 1) { continue; }
+                        f = f + cell_force(pi, pyramid[cell_at(level, cx, cy, cz)]);
+                    }
+                }
+            }
+        }
+        f = f + cell_force(pi, pyramid[grid_cells()]);             // the outside pseudo-cell as one far-field term
+    } else {
+        for (var cz = 0; cz <= zTop; cz = cz + 1) {                // an outside node: the coarsest level in full, no pseudo-cell (it would include itself)
+            for (var cy = 0; cy < ts; cy = cy + 1) {
+                for (var cx = 0; cx < ts; cx = cx + 1) {
+                    f = f + cell_force(pi, pyramid[cell_at(top, cx, cy, cz)]);
+                }
+            }
+        }
+    }
+    store_force(i, load_force(i) + f);
+}
+`;

package/src/wgsl/grid-near-field.wgsl.ts

@@ -0,0 +1,128 @@
+/**
+ * G7, the `grid-near-field` kernel body (spec 7.7, 7.6, 7.20; P4-T10, P4-T13; D24): per node `i = sortedIdx[t]`, the
+ * exact pair law of K3 (`LAW` 0: `|F| = k m_i m_j / d` with the 0.01 floor; `LAW` 1: FR's unfloored `k^2 / d`;
+ * `LAW` 2: the unfloored coulomb `-g m_i m_j / d^2`; the antisymmetric coincident kick at the law's magnitude at
+ * d = 0.01, PD-22) over the 9 (27) finest
+ * cells around its own, or over the outside pseudo-cell alone for an outside node; a cell above `nearMax` entries
+ * is sampled by `nearMax` INDEPENDENT draws with replacement, draw `k` reading the slot
+ * `lowbias32(((c ^ (iteration * 0x9E3779B9)) ^ seed) ^ (k * 0x85EBCA6B)) % count` (every slot's inclusion
+ * probability is `nearMax / count` whatever its position in the sorted order, so a duplicated draw is counted twice
+ * and the node itself, when drawn, is skipped and not replaced), and scaled by `others / sampled` where `sampled` is
+ * the realised number of draws that were not the node (the Horvitz-Thompson form of PD-15 / DEP-P4-K: given
+ * `sampled = s`, those `s` draws are i.i.d. uniform over the `others` slots, so the expectation of the scaled sum is
+ * the exact cell sum whenever `s >= 1`; the G4 record's G4-F2 row carries the measurement); then
+ * the fused epilogue of K3 (gravity, `force +=`, the swing / traction workgroup reduction in uniform control flow).
+ * The helpers `load_force`, `store_force`, `load_old`, `kick_magnitude`, `gravity_force` and the epilogue are K3's
+ * text. Body only; normative text.
+ */
+export const gridNearFieldWgsl = /* wgsl */ `
+fn load_force(i: u32) -> vec3f { return vec3f(force[3u * i], force[3u * i + 1u], force[3u * i + 2u]); }
+fn store_force(i: u32, f: vec3f) {
+    force[3u * i] = f.x;
+    force[3u * i + 1u] = f.y;
+    force[3u * i + 2u] = f.z;
+}
+fn load_old(i: u32) -> vec3f { return vec3f(oldForce[3u * i], oldForce[3u * i + 1u], oldForce[3u * i + 2u]); }
+fn gravity_force(pi: vec4f) -> vec3f {                         // spec 7.9: centroid (GRAVITY_CENTER 0) or origin (1); regular or strong
+    var q = pi.xyz;
+    if (GRAVITY_CENTER == 0u) { q = pi.xyz - S.centroid.xyz; }
+    if (STRONG_GRAVITY) { return -P.gravity * pi.w * q; }
+    let d = length(q);
+    if (d > FA2_DIST_FLOOR) { return -P.gravity * pi.w * q / d; }
+    return vec3f(0.0);
+}
+fn grid_cells() -> u32 { return P.gridMax * P.gridMax * select(1u, P.gridMax, P.dim == 3u); }
+fn kick_magnitude(mi: f32, mj: f32) -> f32 {                   // the law's magnitude at d = FA2_DIST_FLOOR (the P5 plan's PD-10)
+    if (LAW == 1u) { return P.frK * P.frK / FA2_DIST_FLOOR; }
+    if (LAW == 2u) { return -P.coulomb * mi * mj / FA2_DIST_FLOOR_SQ; }
+    return P.scalingRatio * mi * mj / FA2_DIST_FLOOR;
+}
+fn pair_force(i: u32, pi: vec4f, jj: u32, o: vec4f) -> vec3f {  // the exact pair law of K3 (7.6, 7.20): the floor (FA2 only), the coincident kick
+    let d = pi.xyz - o.xyz;
+    var d2 = dot(d, d);
+    if (d2 < FA2_COINCIDENT_SQ) { return kick_dir(i, jj, P.dim) * kick_magnitude(pi.w, o.w); }
+    if (LAW == 0u) { d2 = max(d2, FA2_DIST_FLOOR_SQ); }
+    let k = P.scalingRatio * pi.w * o.w;
+    if (LAW == 1u) { return d * (P.frK * P.frK / d2); }
+    if (LAW == 2u) { return d * (-P.coulomb * pi.w * o.w / (d2 * sqrt(d2))); }
+    return d * (k / d2);
+}
+fn cell_sum(i: u32, pi: vec4f, c: u32, own: bool) -> vec3f {     // one finest cell: exact below nearMax entries, Horvitz-Thompson above (PD-15)
+    let start = cellStart[c];
+    let count = cellStart[c + 1u] - start;
+    var f = vec3f(0.0);
+    if (count <= P.nearMax) {
+        for (var k = start; k < start + count; k = k + 1u) {
+            let jj = sortedIdx[k];
+            if (jj != i) { f = f + pair_force(i, pi, jj, pos[jj]); }
+        }
+        return f;
+    }
+    let base = (c ^ (P.iterationIndex * 0x9E3779B9u)) ^ P.seed;                  // the per-iteration draw seed (7.16)
+    var sampled = 0u;
+    for (var k = 0u; k < P.nearMax; k = k + 1u) {
+        let jj = sortedIdx[start + (lowbias32(base ^ (k * 0x85EBCA6Bu)) % count)];   // draw k: independent inclusion, with replacement (PD-15)
+        if (jj == i) { continue; }
+        f = f + pair_force(i, pi, jj, pos[jj]);
+        sampled = sampled + 1u;
+    }
+    if (sampled == 0u) { return vec3f(0.0); }
+    let others = select(count, count - 1u, own);
+    return f * (f32(others) / f32(sampled));                     // others / sampled over the realised sample (DEP-P4-K)
+}
+
+@compute @workgroup_size(WG)
+fn grid_near_field(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_id) lid: vec3<u32>) {
+    let t = linear_id(wid, lid.x);
+    let valid = t < P.n;
+    var i = 0u;
+    var pi = vec4f(0.0);
+    var f = vec3f(0.0);
+    if (valid) {
+        i = sortedIdx[t];                                          // sorted order (D24)
+        pi = pos[i];
+        let gf = f32(P.gridMax);
+        let q = (pi.xyz - S.gridMin.xyz) * S.invCellSize;
+        var c0 = vec3<i32>(floor(clamp(q, vec3f(-1.0), vec3f(gf + 1.0))));
+        if (P.dim == 2u) { c0.z = 0; }                             // 2D: the one z plane
+        let g = i32(P.gridMax);
+        var inside = c0.x >= 0 && c0.x < g && c0.y >= 0 && c0.y < g;
+        if (P.dim == 3u) { inside = inside && c0.z >= 0 && c0.z < g; }
+        if (inside) {
+            let zr = select(0, 1, P.dim == 3u);
+            for (var dz = -zr; dz <= zr; dz = dz + 1) {
+                for (var dy = -1; dy <= 1; dy = dy + 1) {
+                    for (var dx = -1; dx <= 1; dx = dx + 1) {
+                        let cx = c0.x + dx;
+                        let cy = c0.y + dy;
+                        let cz = c0.z + dz;
+                        if (cx < 0 || cx >= g || cy < 0 || cy >= g || cz < 0 || cz >= g) { continue; }
+                        let c = u32(cx) + P.gridMax * (u32(cy) + select(0u, P.gridMax * u32(cz), P.dim == 3u));
+                        f = f + cell_sum(i, pi, c, dx == 0 && dy == 0 && dz == 0);
+                    }
+                }
+            }
+        } else {
+            f = cell_sum(i, pi, grid_cells(), true);               // an outside node: the pseudo-cell alone
+        }
+    }
+    // epilogue (7.9, 7.10): gravity and force += under the guard, the swing / traction reduction outside it (K3's text)
+    var sw = 0.0;
+    var tr = 0.0;
+    if (valid) {
+        f = f + gravity_force(pi);
+        let fnew = load_force(i) + f;
+        store_force(i, fnew);
+        if (SWING_MODE == 1u) {                                    // NetworkX: positions and forces mixed, every node (7.2)
+            sw = pi.w * length(pi.xyz - fnew);
+            tr = 0.5 * pi.w * length(pi.xyz + fnew);
+        } else if (!mask_bit(fixedMask[i >> 5u], i)) {             // paper: free nodes only
+            let fold = load_old(i);
+            sw = pi.w * length(fnew - fold);
+            tr = 0.5 * pi.w * length(fnew + fold);
+        }
+    }
+    let tt = wg_reduce_vec4(vec4f(sw, tr, 0.0, 0.0), lid.x, 0u);   // uniform control flow: 256 -> 1
+    if (lid.x == 0u) { partials[group_id(wid)].swingTraction = tt.xy; }
+}
+`;

package/src/wgsl/histogram.wgsl.ts

@@ -0,0 +1,14 @@
+/**
+ * The `histogram` kernel body (spec 6 row 5; P4-T3): one atomicAdd per key into the global `hist` (zeroed by a fill
+ * dispatch earlier in the pass, PD-4). Order-independent, hence deterministic. A key >= P.bins is not counted (the
+ * caller's contract; the grid's keys are always < cells + 1). Body only; normative text.
+ */
+export const histogramWgsl = /* wgsl */ `
+@compute @workgroup_size(WG)
+fn histogram(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_id) lid: vec3<u32>) {
+    let i = linear_id(wid, lid.x);
+    if (i >= P.count) { return; }                                  // no barrier follows
+    let k = keys[i];
+    if (k < P.bins) { atomicAdd(&hist[k], 1u); }                   // order-independent: the count is the same whatever the schedule
+}
+`;

package/src/wgsl/indirect-finalize.wgsl.ts

@@ -0,0 +1,25 @@
+/**
+ * The `indirect-finalize` kernel body (spec 5.4; P4-T1): one lane turns the device-side count `counters[P.countIndex]`
+ * into the `(x, y, 1)` of an indirect dispatch by plan1d's rule (spec 5.2) and writes it with the count into the
+ * 16-byte slot `P.slot` of `args` (PD-2). No barrier follows the early return of the other lanes. Body only (spec 3.5,
+ * D9); the text is normative: the P4 sabotage rows are textual edits of it.
+ */
+export const indirectFinalizeWgsl = /* wgsl */ `
+@compute @workgroup_size(WG)
+fn indirect_finalize(@builtin(local_invocation_id) lid: vec3<u32>) {
+    if (lid.x != 0u) { return; }                                   // one lane; no barrier follows (3.5 rule 1)
+    let count = counters[P.countIndex];
+    let groups = count / P.wg + select(0u, 1u, count % P.wg != 0u);   // ceil(count / wg) without the u32 wrap of (count + wg - 1) above 2^32 - wg: plan1d's rule (5.2) for ANY u32 count
+    var x = groups;
+    var y = 1u;
+    if (groups > MAX_WORKGROUPS_PER_DIM) {                         // the 2D split; y <= 1,025 for any u32 count
+        x = MAX_WORKGROUPS_PER_DIM;
+        y = (groups + MAX_WORKGROUPS_PER_DIM - 1u) / MAX_WORKGROUPS_PER_DIM;
+    }
+    let base = 4u * P.slot;                                        // 16-byte slots: (x, y, 1, count) (PD-2)
+    args[base] = x;
+    args[base + 1u] = y;
+    args[base + 2u] = 1u;
+    args[base + 3u] = count;
+}
+`;

package/src/wgsl/radix-hist.wgsl.ts

@@ -0,0 +1,30 @@
+/**
+ * The `radix-hist` kernel body (spec 6 row 6; P4-T4): the RADIX_BINS-bin (2^8) digit histogram of one WG-wide block
+ * of keys, privatised in workgroup memory (atomics on workgroup memory are order-independent) and written
+ * DIGIT-MAJOR, `hist[digit * P.groups + group]`, so one exclusiveScan over the table yields, per digit, the offsets
+ * of the workgroups in workgroup order -- what a stable LSD scatter needs. The bitwise operators act on a KEY and a
+ * DIGIT, never on an arc index (house rule). Body only; normative text.
+ */
+export const radixHistWgsl = /* wgsl */ `
+var<workgroup> local: array<atomic<u32>, RADIX_BINS>;
+
+@compute @workgroup_size(WG)
+fn radix_hist(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_id) lid: vec3<u32>) {
+    for (var b = lid.x; b < RADIX_BINS; b = b + WG) { atomicStore(&local[b], 0u); }
+    workgroupBarrier();
+    let g = group_id(wid);
+    let i = g * WG + lid.x;
+    if (i < P.count) {
+        let d = (keys[i] >> P.shift) & RADIX_DIGIT_MASK;                 // the pass's digit
+        atomicAdd(&local[d], 1u);
+    }
+    workgroupBarrier();
+    // Above MAX_1D_ITEMS plan1d pads the 2D grid to x * y >= P.groups workgroups; a padding workgroup (g >= P.groups)
+    // holds an all-zero table and its digit-major slot b * P.groups + g is digit b + 1's slot of a REAL group, so it
+    // must not store. Uniform per workgroup, no barrier inside. Every test size fits 1D (the ladder tops at 2^22); no
+    // case reaches this branch, so the guard is proved by reading, not by a run.
+    if (g < P.groups) {
+        for (var b = lid.x; b < RADIX_BINS; b = b + WG) { hist[b * P.groups + g] = atomicLoad(&local[b]); }   // digit-major
+    }
+}
+`;

package/src/wgsl/radix-scatter.wgsl.ts

@@ -0,0 +1,39 @@
+/**
+ * The `radix-scatter` kernel body (spec 6 row 6; P4-T4): the stable scatter of one LSD pass. Lane 0 ranks the block's
+ * keys serially in index order with a RADIX_BINS-entry counter table (PD-5: deterministic and stable, WG steps per
+ * workgroup; ponytail: the 8-way split ranking of GraphWaGu is the upgrade if T-6 shows the sort on the critical
+ * path), then every lane writes its key and value at `offsets[digit * P.groups + group] + rank`. Body only;
+ * normative text.
+ */
+export const radixScatterWgsl = /* wgsl */ `
+var<workgroup> rank: array<u32, WG>;
+var<workgroup> cnt: array<u32, RADIX_BINS>;
+
+@compute @workgroup_size(WG)
+fn radix_scatter(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_id) lid: vec3<u32>) {
+    for (var b = lid.x; b < RADIX_BINS; b = b + WG) { cnt[b] = 0u; }
+    let g = group_id(wid);
+    let i = g * WG + lid.x;
+    var key = 0u;
+    var d = 0u;
+    if (i < P.count) {
+        key = keys[i];
+        d = (key >> P.shift) & RADIX_DIGIT_MASK;
+    }
+    workgroupBarrier();
+    if (lid.x == 0u) {                                             // serial stable ranking (PD-5)
+        let last = min(WG, P.count - g * WG);
+        for (var s = 0u; s < last; s = s + 1u) {
+            let ds = (keys[g * WG + s] >> P.shift) & RADIX_DIGIT_MASK;
+            rank[s] = cnt[ds];
+            cnt[ds] = cnt[ds] + 1u;
+        }
+    }
+    workgroupBarrier();
+    if (i < P.count) {
+        let dst = offsets[d * P.groups + g] + rank[lid.x];
+        keysOut[dst] = key;
+        valsOut[dst] = vals[i];
+    }
+}
+`;