@graphty/webgpu-graph-algorithms 0.5.0 → 0.6.0

package/dist/src/wgsl/fa2-repulsion-exact.wgsl.js.map

@@ -1 +1 @@
-{"version":3,"file":"fa2-repulsion-exact.wgsl.js","sourceRoot":"","sources":["../../../src/wgsl/fa2-repulsion-exact.wgsl.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AACH,MAAM,CAAC,MAAM,qBAAqB,GAAG,UAAU,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CAmE/C,CAAC"}
+{"version":3,"file":"fa2-repulsion-exact.wgsl.js","sourceRoot":"","sources":["../../../src/wgsl/fa2-repulsion-exact.wgsl.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AACH,MAAM,CAAC,MAAM,qBAAqB,GAAG,UAAU,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA0E/C,CAAC"}

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

@@ -4,11 +4,22 @@
  * 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
  * is the target of the K1 sabotage mutations (test/helpers/sabotage.ts, P3-T5); amend the contract before editing.
  */
 /** The K1 body: entry point `stats_finalize`; calls the reduction helpers (`needs: ["subgroups"]`, contract 4.3). */
-export declare const fa2StatsFinalizeWgsl = "// K1: folds the previous integrate's partials into the state block (spec 7.4); one workgroup\n@compute @workgroup_size(WG)\nfn stats_finalize(@builtin(local_invocation_id) lid: vec3<u32>) {\n    let groups = (P.n + WG - 1u) / WG;\n    let fold = (P.flags & FA2_FLAG_FIRST) == 0u;    // the first iteration after load() keeps the host-written state\n    var sum = vec4f(0.0);\n    var lo = vec4f(F32_MAX);\n    var hi = vec4f(-F32_MAX);\n    var disp = 0.0;\n    var free = 0u;\n    if (fold) {\n        for (var g = lid.x; g < groups; g = g + WG) {   // sequential per lane in index order: deterministic\n            let q = partials[g];\n            sum = sum + q.sum;\n            lo = min(lo, q.min);\n            hi = max(hi, q.max);\n            disp = disp + q.dispFree.x;\n            free = free + u32(q.dispFree.y);\n        }\n    }\n    let tSum = wg_reduce_vec4(sum, lid.x, 0u);\n    let tLo = wg_reduce_vec4(lo, lid.x, 1u);\n    let tHi = wg_reduce_vec4(hi, lid.x, 2u);\n    let tDisp = wg_reduce_f32(disp, lid.x, 0u);\n    let tFree = wg_reduce_u32(free, lid.x, 0u);\n    if (lid.x == 0u) {\n        if (fold) {\n            let n = f32(P.n);\n            let c = tSum.xyz / n;\n            S.centroid = vec4f(c, 0.0);\n            S.rmsRadius = sqrt(max(tSum.w, 0.0) / n);                  // RMS radius about the previous centroid (7.17)\n            S.min = vec4f(tLo.xyz, 0.0);\n            S.max = vec4f(tHi.xyz, 0.0);\n            S.radius = sqrt(max(tHi.w, 0.0));                          // max |p - centroid| about the same previous centroid as rmsRadius (K5 puts |q|^2 in max.w)\n            let meanDisp = select(tDisp / f32(tFree), 0.0, tFree == 0u);  // all-fixed: 0, never NaN (7.4)\n            S.meanDisplacement = meanDisp;\n            S.settledCount = select(0u, S.settledCount + 1u, meanDisp <= P.settleThreshold * S.rmsRadius);\n        }\n        S.iteration = S.iteration + 1u;\n        T[P.iterationIndex].meanDisplacement = S.meanDisplacement;\n        T[P.iterationIndex].settledCount = S.settledCount;\n        T[P.iterationIndex].iteration = S.iteration;\n    }\n}";
+export declare const fa2StatsFinalizeWgsl = "// K1: folds the previous integrate's partials into the state block (spec 7.4); one workgroup\n@compute @workgroup_size(WG)\nfn stats_finalize(@builtin(local_invocation_id) lid: vec3<u32>) {\n    let groups = (P.n + WG - 1u) / WG;\n    let fold = (P.flags & FA2_FLAG_FIRST) == 0u;    // the first iteration after load() keeps the host-written state\n    var sum = vec4f(0.0);\n    var lo = vec4f(F32_MAX);\n    var hi = vec4f(-F32_MAX);\n    var disp = 0.0;\n    var free = 0u;\n    var ke = 0.0;\n    if (fold) {\n        for (var g = lid.x; g < groups; g = g + WG) {   // sequential per lane in index order: deterministic\n            let q = partials[g];\n            sum = sum + q.sum;\n            lo = min(lo, q.min);\n            hi = max(hi, q.max);\n            disp = disp + q.dispFree.x;\n            free = free + u32(q.dispFree.y);\n            ke = ke + q.swingTraction.x;\n        }\n    }\n    let tSum = wg_reduce_vec4(sum, lid.x, 0u);\n    let tLo = wg_reduce_vec4(lo, lid.x, 1u);\n    let tHi = wg_reduce_vec4(hi, lid.x, 2u);\n    let tDisp = wg_reduce_f32(disp, lid.x, 0u);\n    let tFree = wg_reduce_u32(free, lid.x, 0u);\n    let tKe = wg_reduce_f32(ke, lid.x, 0u);\n    if (lid.x == 0u) {\n        if (fold) {\n            let n = f32(P.n);\n            let c = tSum.xyz / n;\n            S.centroid = vec4f(c, 0.0);\n            S.rmsRadius = sqrt(max(tSum.w, 0.0) / n);                  // RMS radius about the previous centroid (7.17)\n            S.min = vec4f(tLo.xyz, 0.0);\n            S.max = vec4f(tHi.xyz, 0.0);\n            S.radius = sqrt(max(tHi.w, 0.0));                          // max |p - centroid| about the same previous centroid as rmsRadius (K5 puts |q|^2 in max.w)\n            let meanDisp = select(tDisp / f32(tFree), 0.0, tFree == 0u);  // all-fixed: 0, never NaN (7.4)\n            S.meanDisplacement = meanDisp;\n            S.settledCount = select(0u, S.settledCount + 1u, meanDisp <= P.settleThreshold * S.rmsRadius);\n        }\n        S.iteration = S.iteration + 1u;\n        T[P.iterationIndex].meanDisplacement = S.meanDisplacement;\n        T[P.iterationIndex].settledCount = S.settledCount;\n        T[P.iterationIndex].iteration = S.iteration;\n        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)\n            let cells = P.gridMax * P.gridMax * select(1u, P.gridMax, P.dim == 3u);\n            if (fold) {\n                let box = (S.max.xyz - S.min.xyz) * GRID_BBOX_MARGIN;\n                var bboxExtent = max(box.x, box.y);\n                if (P.dim == 3u) { bboxExtent = max(bboxExtent, box.z); }\n                let extent = max(min(bboxExtent, P.extentFactor * S.rmsRadius), GRID_EXTENT_FLOOR);   // min(bbox, extentFactor x rms), floored (7.7)\n                let cellSize = extent / f32(P.gridMax);\n                S.gridMin = vec4f(S.centroid.xyz - vec3f(0.5 * extent), cellSize);   // gridMin.w carries cellSize\n                S.invCellSize = 1.0 / cellSize;\n                S.eps = 0.25 * cellSize;\n            }\n            S.outsideGrid = cellHist[cells];                         // the previous iteration's pseudo-cell count (0 after load)\n            S.maxCellOccupancy = atomicLoad(&hubCounters[1]);\n            atomicStore(&hubCounters[0], 0u);\n            atomicStore(&hubCounters[1], 0u);\n        }\n        if (STATS_MODE == 1u) {                                    // FR: this iteration's temperature (7.20) into the state and the trace\n            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\n                if (fold) {\n                    var t = S.temperature;\n                    if (tKe < S.frEnergy) {\n                        S.frProgress = S.frProgress + 1u;\n                        if (S.frProgress >= FR_COOLING_PATIENCE) { S.frProgress = 0u; t = t / FR_COOLING_STEP; }\n                    } else {\n                        S.frProgress = 0u;\n                        t = t * FR_COOLING_STEP;\n                    }\n                    S.frEnergy = tKe;\n                    S.temperature = t;\n                }\n            } else {\n                S.temperature = P.temperature;\n            }\n            T[P.iterationIndex].modelScalar = S.temperature;\n        }\n        if (STATS_MODE == 2u) {                                    // spring-electrical: the kinetic energy K5 folded into partials B (PD-4); 0 on the first iteration after load()\n            S.kineticEnergy = tKe;\n            T[P.iterationIndex].modelScalar = tKe;\n        }\n    }\n}";
 //# sourceMappingURL=fa2-stats-finalize.wgsl.d.ts.map

package/dist/src/wgsl/fa2-stats-finalize.wgsl.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"fa2-stats-finalize.wgsl.d.ts","sourceRoot":"","sources":["../../../src/wgsl/fa2-stats-finalize.wgsl.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH,qHAAqH;AACrH,eAAO,MAAM,oBAAoB,gkEA2C/B,CAAC"}
+{"version":3,"file":"fa2-stats-finalize.wgsl.d.ts","sourceRoot":"","sources":["../../../src/wgsl/fa2-stats-finalize.wgsl.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;GAqBG;AAEH,qHAAqH;AACrH,eAAO,MAAM,oBAAoB,snJAsF/B,CAAC"}

package/dist/src/wgsl/fa2-stats-finalize.wgsl.js

@@ -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
@@ -20,6 +31,7 @@ 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];
@@ -28,6 +40,7 @@ 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);
@@ -35,6 +48,7 @@ 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);
@@ -52,6 +66,46 @@ 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;
+        }
     }
 }`;
 //# sourceMappingURL=fa2-stats-finalize.wgsl.js.map

package/dist/src/wgsl/fa2-stats-finalize.wgsl.js.map

@@ -1 +1 @@
-{"version":3,"file":"fa2-stats-finalize.wgsl.js","sourceRoot":"","sources":["../../../src/wgsl/fa2-stats-finalize.wgsl.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH,qHAAqH;AACrH,MAAM,CAAC,MAAM,oBAAoB,GAAG,UAAU,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;EA2C7C,CAAC"}
+{"version":3,"file":"fa2-stats-finalize.wgsl.js","sourceRoot":"","sources":["../../../src/wgsl/fa2-stats-finalize.wgsl.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;GAqBG;AAEH,qHAAqH;AACrH,MAAM,CAAC,MAAM,oBAAoB,GAAG,UAAU,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;EAsF7C,CAAC"}

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

@@ -0,0 +1,8 @@
+/**
+ * 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 declare const gridCellKeyWgsl = "\nfn grid_cells() -> u32 { return P.gridMax * P.gridMax * select(1u, P.gridMax, P.dim == 3u); }\n\n@compute @workgroup_size(WG)\nfn grid_cell_key(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_id) lid: vec3<u32>) {\n    let i = linear_id(wid, lid.x);\n    if (i >= P.n) { return; }                                      // no barrier follows\n    let cells = grid_cells();\n    let gf = f32(P.gridMax);\n    let q = (pos[i].xyz - S.gridMin.xyz) * S.invCellSize;          // PD-10: never a division\n    let c = vec3<i32>(floor(clamp(q, vec3f(-1.0), vec3f(gf + 1.0))));\n    let g = i32(P.gridMax);\n    var inside = c.x >= 0 && c.x < g && c.y >= 0 && c.y < g;\n    if (P.dim == 3u) { inside = inside && c.z >= 0 && c.z < g; }\n    var key = cells;                                               // the outside pseudo-cell (7.7)\n    if (inside) {\n        key = u32(c.x) + P.gridMax * u32(c.y);\n        if (P.dim == 3u) { key = key + P.gridMax * P.gridMax * u32(c.z); }\n    }\n    cellKey[i] = key;\n    cellVal[i] = i;\n}\n";
+//# sourceMappingURL=grid-cell-key.wgsl.d.ts.map

package/dist/src/wgsl/grid-cell-key.wgsl.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"grid-cell-key.wgsl.d.ts","sourceRoot":"","sources":["../../../src/wgsl/grid-cell-key.wgsl.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AACH,eAAO,MAAM,eAAe,uhCAsB3B,CAAC"}

package/dist/src/wgsl/grid-cell-key.wgsl.js

@@ -0,0 +1,30 @@
+/**
+ * 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;
+}
+`;
+//# sourceMappingURL=grid-cell-key.wgsl.js.map

package/dist/src/wgsl/grid-cell-key.wgsl.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"grid-cell-key.wgsl.js","sourceRoot":"","sources":["../../../src/wgsl/grid-cell-key.wgsl.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AACH,MAAM,CAAC,MAAM,eAAe,GAAG,UAAU,CAAC;;;;;;;;;;;;;;;;;;;;;;CAsBzC,CAAC"}

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

@@ -0,0 +1,8 @@
+/**
+ * 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 declare const gridCentroidHubWgsl = "\n@compute @workgroup_size(WG)\nfn grid_centroid_hub(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_id) lid: vec3<u32>) {\n    let h = group_id(wid);\n    let valid = h < hubCount[0];                                   // a workgroup past the count sums nothing\n    var c = 0u;\n    var start = 0u;\n    var count = 0u;\n    if (valid) {\n        c = hubList[h];\n        start = cellStart[c];\n        count = cellStart[c + 1u] - start;\n    }\n    var acc = vec4f(0.0);\n    for (var k = start + lid.x; k < start + count; k = k + WG) {   // strided over the cell's sorted range\n        let p = pos[sortedIdx[k]];\n        acc = acc + vec4f(p.xyz * p.w, p.w);\n    }\n    let t = wg_reduce_vec4(acc, lid.x, 0u);                        // uniform control flow: 256 -> 1\n    if (valid && lid.x == 0u) { pyramid[c] = t; }\n}\n";
+//# sourceMappingURL=grid-centroid-hub.wgsl.d.ts.map

package/dist/src/wgsl/grid-centroid-hub.wgsl.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"grid-centroid-hub.wgsl.d.ts","sourceRoot":"","sources":["../../../src/wgsl/grid-centroid-hub.wgsl.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AACH,eAAO,MAAM,mBAAmB,i1BAqB/B,CAAC"}

package/dist/src/wgsl/grid-centroid-hub.wgsl.js

@@ -0,0 +1,29 @@
+/**
+ * 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; }
+}
+`;
+//# sourceMappingURL=grid-centroid-hub.wgsl.js.map

package/dist/src/wgsl/grid-centroid-hub.wgsl.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"grid-centroid-hub.wgsl.js","sourceRoot":"","sources":["../../../src/wgsl/grid-centroid-hub.wgsl.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AACH,MAAM,CAAC,MAAM,mBAAmB,GAAG,UAAU,CAAC;;;;;;;;;;;;;;;;;;;;;CAqB7C,CAAC"}

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

@@ -0,0 +1,8 @@
+/**
+ * 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 declare const gridCentroidWgsl = "\nfn grid_cells() -> u32 { return P.gridMax * P.gridMax * select(1u, P.gridMax, P.dim == 3u); }\n\n@compute @workgroup_size(WG)\nfn grid_centroid(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_id) lid: vec3<u32>) {\n    let c = linear_id(wid, lid.x);\n    if (c > grid_cells()) { return; }                              // cells [0, cells]: the pseudo-cell is index cells; no barrier follows\n    let start = cellStart[c];\n    let count = cellStart[c + 1u] - start;\n    atomicMax(&hubCounters[1], count);                             // maxCellOccupancy, read by K1 next iteration\n    if (count > GRID_HUB_CELL) {                                   // a hub cell: G4b sums it (PD-13)\n        hubList[atomicAdd(&hubCounters[0], 1u)] = c;\n        return;\n    }\n    var acc = vec4f(0.0);\n    for (var k = start; k < start + count; k = k + 1u) {           // sorted order: deterministic\n        let p = pos[sortedIdx[k]];\n        acc = acc + vec4f(p.xyz * p.w, p.w);                       // (sum m x, sum m y, sum m z, sum m)\n    }\n    pyramid[c] = acc;\n}\n";
+//# sourceMappingURL=grid-centroid.wgsl.d.ts.map

package/dist/src/wgsl/grid-centroid.wgsl.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"grid-centroid.wgsl.d.ts","sourceRoot":"","sources":["../../../src/wgsl/grid-centroid.wgsl.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AACH,eAAO,MAAM,gBAAgB,+jCAqB5B,CAAC"}

package/dist/src/wgsl/grid-centroid.wgsl.js

@@ -0,0 +1,29 @@
+/**
+ * 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;
+}
+`;
+//# sourceMappingURL=grid-centroid.wgsl.js.map

package/dist/src/wgsl/grid-centroid.wgsl.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"grid-centroid.wgsl.js","sourceRoot":"","sources":["../../../src/wgsl/grid-centroid.wgsl.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AACH,MAAM,CAAC,MAAM,gBAAgB,GAAG,UAAU,CAAC;;;;;;;;;;;;;;;;;;;;;CAqB1C,CAAC"}

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

@@ -0,0 +1,7 @@
+/**
+ * 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 declare const gridDownsampleWgsl = "\n@compute @workgroup_size(WG)\nfn grid_downsample(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_id) lid: vec3<u32>) {\n    let pc = linear_id(wid, lid.x);                                // the parent cell inside its level\n    if (pc >= P.parentCells) { return; }                           // no barrier follows\n    let side = P.parentSide;\n    let cs = 2u * side;                                            // the child level's side\n    let px = pc % side;\n    let py = (pc / side) % side;\n    let pz = pc / (side * side);\n    var acc = vec4f(0.0);\n    for (var dz = 0u; dz < P.depth; dz = dz + 1u) {\n        for (var dy = 0u; dy < 2u; dy = dy + 1u) {\n            for (var dx = 0u; dx < 2u; dx = dx + 1u) {\n                let child = (2u * px + dx) + cs * ((2u * py + dy) + cs * (2u * pz + dz));\n                acc = acc + pyramid[P.childBase + child];\n            }\n        }\n    }\n    pyramid[P.parentBase + pc] = acc;\n}\n";
+//# sourceMappingURL=grid-downsample.wgsl.d.ts.map

package/dist/src/wgsl/grid-downsample.wgsl.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"grid-downsample.wgsl.d.ts","sourceRoot":"","sources":["../../../src/wgsl/grid-downsample.wgsl.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AACH,eAAO,MAAM,kBAAkB,w8BAqB9B,CAAC"}

package/dist/src/wgsl/grid-downsample.wgsl.js

@@ -0,0 +1,28 @@
+/**
+ * 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;
+}
+`;
+//# sourceMappingURL=grid-downsample.wgsl.js.map

package/dist/src/wgsl/grid-downsample.wgsl.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"grid-downsample.wgsl.js","sourceRoot":"","sources":["../../../src/wgsl/grid-downsample.wgsl.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AACH,MAAM,CAAC,MAAM,kBAAkB,GAAG,UAAU,CAAC;;;;;;;;;;;;;;;;;;;;;CAqB5C,CAAC"}

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

@@ -0,0 +1,13 @@
+/**
+ * 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 declare const gridFarFieldWgsl = "\nfn load_force(i: u32) -> vec3f { return vec3f(force[3u * i], force[3u * i + 1u], force[3u * i + 2u]); }\nfn store_force(i: u32, f: vec3f) {\n    force[3u * i] = f.x;\n    force[3u * i + 1u] = f.y;\n    force[3u * i + 2u] = f.z;\n}\nfn grid_cells() -> u32 { return P.gridMax * P.gridMax * select(1u, P.gridMax, P.dim == 3u); }\nfn grid_side(level: u32) -> u32 { return P.gridMax >> level; }\nfn level_base(level: u32) -> u32 {                             // the pyramid index of level L's cell 0 (level 0 carries the pseudo-cell at index cells)\n    var base = 0u;\n    for (var l = 0u; l < level; l = l + 1u) {\n        let s = grid_side(l);\n        base = base + s * s * select(1u, s, P.dim == 3u) + select(0u, 1u, l == 0u);\n    }\n    return base;\n}\nfn cell_at(level: u32, cx: i32, cy: i32, cz: i32) -> u32 {\n    let s = grid_side(level);\n    return level_base(level) + u32(cx) + s * (u32(cy) + select(0u, s * u32(cz), P.dim == 3u));\n}\nfn cell_force(pi: vec4f, q: vec4f) -> vec3f {                  // one far-field term, softened by state.eps (7.7)\n    if (q.w <= 0.0) { return vec3f(0.0); }                     // an empty cell\n    let d = pi.xyz - q.xyz / q.w;                              // to the mass-weighted centroid\n    let d2 = dot(d, d) + S.eps * S.eps;\n    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\n    if (LAW == 2u) { return d * (-P.coulomb * pi.w * q.w / (d2 * sqrt(d2))); }     // LAW 2 (coulomb): -g m_i M_cell / d^2\n    return d * (P.scalingRatio * pi.w * q.w / d2);             // LAW 0 (FA2): |F| = k m_i M_cell / d\n}\n\n@compute @workgroup_size(WG)\nfn grid_far_field(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_id) lid: vec3<u32>) {\n    let t = linear_id(wid, lid.x);\n    if (t >= P.n) { return; }                                      // no barrier follows\n    let i = sortedIdx[t];                                          // sorted order (D24)\n    let pi = pos[i];\n    let gf = f32(P.gridMax);\n    let q = (pi.xyz - S.gridMin.xyz) * S.invCellSize;              // PD-10\n    var c0 = vec3<i32>(floor(clamp(q, vec3f(-1.0), vec3f(gf + 1.0))));\n    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\n    let g = i32(P.gridMax);\n    var inside = c0.x >= 0 && c0.x < g && c0.y >= 0 && c0.y < g;\n    if (P.dim == 3u) { inside = inside && c0.z >= 0 && c0.z < g; }\n    let top = P.levels - 1u;\n    let ts = i32(grid_side(top));                                  // the coarsest side (4)\n    let zTop = select(0, ts - 1, P.dim == 3u);                     // z ranges: one plane in 2D\n    var f = vec3f(0.0);\n    if (inside) {\n        let ct = c0 / i32(1u << top);                              // the node's coarsest cell\n        for (var cz = 0; cz <= zTop; cz = cz + 1) {\n            for (var cy = 0; cy < ts; cy = cy + 1) {\n                for (var cx = 0; cx < ts; cx = cx + 1) {\n                    if (abs(cx - ct.x) <= 1 && abs(cy - ct.y) <= 1 && abs(cz - ct.z) <= 1) { continue; }   // the 3x3(x3) is finer levels' work\n                    f = f + cell_force(pi, pyramid[cell_at(top, cx, cy, cz)]);\n                }\n            }\n        }\n        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\n            let level = l - 1u;\n            let cl = c0 / i32(1u << level);\n            let cp = cl / 2;\n            let side = i32(grid_side(level));\n            let zLo = select(0, max(0, 2 * (cp.z - 1)), P.dim == 3u);\n            let zHi = select(0, min(side - 1, 2 * (cp.z + 1) + 1), P.dim == 3u);\n            for (var cz = zLo; cz <= zHi; cz = cz + 1) {\n                for (var cy = max(0, 2 * (cp.y - 1)); cy <= min(side - 1, 2 * (cp.y + 1) + 1); cy = cy + 1) {\n                    for (var cx = max(0, 2 * (cp.x - 1)); cx <= min(side - 1, 2 * (cp.x + 1) + 1); cx = cx + 1) {\n                        if (abs(cx - cl.x) <= 1 && abs(cy - cl.y) <= 1 && abs(cz - cl.z) <= 1) { continue; }\n                        f = f + cell_force(pi, pyramid[cell_at(level, cx, cy, cz)]);\n                    }\n                }\n            }\n        }\n        f = f + cell_force(pi, pyramid[grid_cells()]);             // the outside pseudo-cell as one far-field term\n    } else {\n        for (var cz = 0; cz <= zTop; cz = cz + 1) {                // an outside node: the coarsest level in full, no pseudo-cell (it would include itself)\n            for (var cy = 0; cy < ts; cy = cy + 1) {\n                for (var cx = 0; cx < ts; cx = cx + 1) {\n                    f = f + cell_force(pi, pyramid[cell_at(top, cx, cy, cz)]);\n                }\n            }\n        }\n    }\n    store_force(i, load_force(i) + f);\n}\n";
+//# sourceMappingURL=grid-far-field.wgsl.d.ts.map

package/dist/src/wgsl/grid-far-field.wgsl.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"grid-far-field.wgsl.d.ts","sourceRoot":"","sources":["../../../src/wgsl/grid-far-field.wgsl.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AACH,eAAO,MAAM,gBAAgB,mzJAqF5B,CAAC"}

package/dist/src/wgsl/grid-far-field.wgsl.js

@@ -0,0 +1,98 @@
+/**
+ * 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);
+}
+`;
+//# sourceMappingURL=grid-far-field.wgsl.js.map

package/dist/src/wgsl/grid-far-field.wgsl.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"grid-far-field.wgsl.js","sourceRoot":"","sources":["../../../src/wgsl/grid-far-field.wgsl.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AACH,MAAM,CAAC,MAAM,gBAAgB,GAAG,UAAU,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CAqF1C,CAAC"}

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

@@ -0,0 +1,19 @@
+/**
+ * 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 declare const gridNearFieldWgsl = "\nfn load_force(i: u32) -> vec3f { return vec3f(force[3u * i], force[3u * i + 1u], force[3u * i + 2u]); }\nfn store_force(i: u32, f: vec3f) {\n    force[3u * i] = f.x;\n    force[3u * i + 1u] = f.y;\n    force[3u * i + 2u] = f.z;\n}\nfn load_old(i: u32) -> vec3f { return vec3f(oldForce[3u * i], oldForce[3u * i + 1u], oldForce[3u * i + 2u]); }\nfn gravity_force(pi: vec4f) -> vec3f {                         // spec 7.9: centroid (GRAVITY_CENTER 0) or origin (1); regular or strong\n    var q = pi.xyz;\n    if (GRAVITY_CENTER == 0u) { q = pi.xyz - S.centroid.xyz; }\n    if (STRONG_GRAVITY) { return -P.gravity * pi.w * q; }\n    let d = length(q);\n    if (d > FA2_DIST_FLOOR) { return -P.gravity * pi.w * q / d; }\n    return vec3f(0.0);\n}\nfn grid_cells() -> u32 { return P.gridMax * P.gridMax * select(1u, P.gridMax, P.dim == 3u); }\nfn kick_magnitude(mi: f32, mj: f32) -> f32 {                   // the law's magnitude at d = FA2_DIST_FLOOR (the P5 plan's PD-10)\n    if (LAW == 1u) { return P.frK * P.frK / FA2_DIST_FLOOR; }\n    if (LAW == 2u) { return -P.coulomb * mi * mj / FA2_DIST_FLOOR_SQ; }\n    return P.scalingRatio * mi * mj / FA2_DIST_FLOOR;\n}\nfn 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\n    let d = pi.xyz - o.xyz;\n    var d2 = dot(d, d);\n    if (d2 < FA2_COINCIDENT_SQ) { return kick_dir(i, jj, P.dim) * kick_magnitude(pi.w, o.w); }\n    if (LAW == 0u) { d2 = max(d2, FA2_DIST_FLOOR_SQ); }\n    let k = P.scalingRatio * pi.w * o.w;\n    if (LAW == 1u) { return d * (P.frK * P.frK / d2); }\n    if (LAW == 2u) { return d * (-P.coulomb * pi.w * o.w / (d2 * sqrt(d2))); }\n    return d * (k / d2);\n}\nfn cell_sum(i: u32, pi: vec4f, c: u32, own: bool) -> vec3f {     // one finest cell: exact below nearMax entries, Horvitz-Thompson above (PD-15)\n    let start = cellStart[c];\n    let count = cellStart[c + 1u] - start;\n    var f = vec3f(0.0);\n    if (count <= P.nearMax) {\n        for (var k = start; k < start + count; k = k + 1u) {\n            let jj = sortedIdx[k];\n            if (jj != i) { f = f + pair_force(i, pi, jj, pos[jj]); }\n        }\n        return f;\n    }\n    let base = (c ^ (P.iterationIndex * 0x9E3779B9u)) ^ P.seed;                  // the per-iteration draw seed (7.16)\n    var sampled = 0u;\n    for (var k = 0u; k < P.nearMax; k = k + 1u) {\n        let jj = sortedIdx[start + (lowbias32(base ^ (k * 0x85EBCA6Bu)) % count)];   // draw k: independent inclusion, with replacement (PD-15)\n        if (jj == i) { continue; }\n        f = f + pair_force(i, pi, jj, pos[jj]);\n        sampled = sampled + 1u;\n    }\n    if (sampled == 0u) { return vec3f(0.0); }\n    let others = select(count, count - 1u, own);\n    return f * (f32(others) / f32(sampled));                     // others / sampled over the realised sample (DEP-P4-K)\n}\n\n@compute @workgroup_size(WG)\nfn grid_near_field(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_id) lid: vec3<u32>) {\n    let t = linear_id(wid, lid.x);\n    let valid = t < P.n;\n    var i = 0u;\n    var pi = vec4f(0.0);\n    var f = vec3f(0.0);\n    if (valid) {\n        i = sortedIdx[t];                                          // sorted order (D24)\n        pi = pos[i];\n        let gf = f32(P.gridMax);\n        let q = (pi.xyz - S.gridMin.xyz) * S.invCellSize;\n        var c0 = vec3<i32>(floor(clamp(q, vec3f(-1.0), vec3f(gf + 1.0))));\n        if (P.dim == 2u) { c0.z = 0; }                             // 2D: the one z plane\n        let g = i32(P.gridMax);\n        var inside = c0.x >= 0 && c0.x < g && c0.y >= 0 && c0.y < g;\n        if (P.dim == 3u) { inside = inside && c0.z >= 0 && c0.z < g; }\n        if (inside) {\n            let zr = select(0, 1, P.dim == 3u);\n            for (var dz = -zr; dz <= zr; dz = dz + 1) {\n                for (var dy = -1; dy <= 1; dy = dy + 1) {\n                    for (var dx = -1; dx <= 1; dx = dx + 1) {\n                        let cx = c0.x + dx;\n                        let cy = c0.y + dy;\n                        let cz = c0.z + dz;\n                        if (cx < 0 || cx >= g || cy < 0 || cy >= g || cz < 0 || cz >= g) { continue; }\n                        let c = u32(cx) + P.gridMax * (u32(cy) + select(0u, P.gridMax * u32(cz), P.dim == 3u));\n                        f = f + cell_sum(i, pi, c, dx == 0 && dy == 0 && dz == 0);\n                    }\n                }\n            }\n        } else {\n            f = cell_sum(i, pi, grid_cells(), true);               // an outside node: the pseudo-cell alone\n        }\n    }\n    // epilogue (7.9, 7.10): gravity and force += under the guard, the swing / traction reduction outside it (K3's text)\n    var sw = 0.0;\n    var tr = 0.0;\n    if (valid) {\n        f = f + gravity_force(pi);\n        let fnew = load_force(i) + f;\n        store_force(i, fnew);\n        if (SWING_MODE == 1u) {                                    // NetworkX: positions and forces mixed, every node (7.2)\n            sw = pi.w * length(pi.xyz - fnew);\n            tr = 0.5 * pi.w * length(pi.xyz + fnew);\n        } else if (!mask_bit(fixedMask[i >> 5u], i)) {             // paper: free nodes only\n            let fold = load_old(i);\n            sw = pi.w * length(fnew - fold);\n            tr = 0.5 * pi.w * length(fnew + fold);\n        }\n    }\n    let tt = wg_reduce_vec4(vec4f(sw, tr, 0.0, 0.0), lid.x, 0u);   // uniform control flow: 256 -> 1\n    if (lid.x == 0u) { partials[group_id(wid)].swingTraction = tt.xy; }\n}\n";
+//# sourceMappingURL=grid-near-field.wgsl.d.ts.map

package/dist/src/wgsl/grid-near-field.wgsl.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"grid-near-field.wgsl.d.ts","sourceRoot":"","sources":["../../../src/wgsl/grid-near-field.wgsl.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AACH,eAAO,MAAM,iBAAiB,27KA8G7B,CAAC"}