chore: checkpoint web Blender migration source, tools and docs

Accumulated uncommitted work: GHOST Web backend, WebGL/GPU web layer,
web workflow engine and protocol, tooling, docs and task records.
Large generated evidence, user-testing archives, build snapshots and
imported test assets remain untracked.

Co-Authored-By: Claude Code <noreply@anthropic.com>

web/protocol/gpu-compute.ts

@@ -0,0 +1,76 @@
+/** Native buffer and storage-image compute. Kernels execute on a real WebGPU device. */
+export type ComputeBindingKind = "storage" | "read-only-storage" | "uniform" | "storage-image" | "sampled-texture";
+/** Tight native rows; r8uint uses one exact uint32 and rg11b10ufloat three float32 values per texel. */
+export interface ComputeImage {
+  format: "r8uint" | "r32uint" | "r32float" | "r16float" | "rg16float" | "rg11b10ufloat" | "rgba16float" | "rgba32float" | "rgba8unorm";
+  width: number;
+  height: number;
+  /** Layer-major tight rows. Omitted fields retain the original 2D wire format. */
+  layers?: number;
+  dimension?: "2d" | "2d-array" | "cube" | "3d";
+  access: "read-only" | "write-only" | "read-write";
+  /** Sampled snapshots are mip-major, then layer-major tight rows. Each gets
+   * a separate sampler binding after all payload bindings, in payload order. */
+  mipLevelCount?: number;
+  sampler?: {
+    addressModeU: "clamp-to-edge" | "repeat" | "mirror-repeat";
+    addressModeV: "clamp-to-edge" | "repeat" | "mirror-repeat";
+    addressModeW: "clamp-to-edge" | "repeat" | "mirror-repeat";
+    magFilter: "nearest" | "linear";
+    minFilter: "nearest" | "linear";
+    mipmapFilter: "nearest" | "linear";
+    lodMinClamp: number;
+    lodMaxClamp: number;
+  };
+}
+export interface ComputeStep {
+  entryPoint?: string;
+  groups?: [number, number, number];
+  indirect?: { buffer: number; offset: number };
+  constants?: Record<string, number>;
+  bindingTypes?: ComputeBindingKind[];
+}
+export interface ComputeJob {
+  language: "wgsl" | "glsl";
+  source: string;
+  entryPoint: string;
+  groups: [number, number, number];
+  kinds: ComputeBindingKind[];
+  steps: ComputeStep[] | null;
+  images?: Record<string, ComputeImage>;
+  /** Native allocation lifetime tokens, never WASM addresses. Null entries are transient. */
+  owners?: (string | null)[];
+  /** Native API arguments at byte offset zero, snapshotted after COMMAND visibility.
+   * This is an INDIRECT-only allocation, never an extra shader binding. */
+  nativeIndirect?: { owner: string; arguments: [number, number, number] };
+}
+export interface ComputeRelease { releaseOwners: string[] }
+/** Effective limits of the device used by this broker, capped by transport budgets.
+ * Keep the word order stable for the shared-memory capability reply. */
+export const COMPUTE_CAPABILITY_FIELDS = [
+  "storageBuffers", "storageBufferBytes", "storageBufferAlignment", "workgroups",
+  "workgroupSizeX", "workgroupSizeY", "workgroupSizeZ", "uniformBufferBytes",
+] as const;
+export type ComputeCapabilities = Record<typeof COMPUTE_CAPABILITY_FIELDS[number], number>;
+export interface ComputeCapabilitiesRequest {
+  capabilities: true;
+  control: SharedArrayBuffer;
+  error: SharedArrayBuffer;
+  data: SharedArrayBuffer;
+}
+export interface ComputeRequest {
+  job: ComputeJob;
+  sizes: number[];
+  /** Control[0]: 0 pending, 1 committed, -1 error, 2 abandoned. */
+  control: SharedArrayBuffer;
+  error: SharedArrayBuffer;
+  data: SharedArrayBuffer;
+}
+export const COMPUTE_LIMITS = {
+  sourceBytes: 1024 * 1024,
+  buffers: 32,
+  bufferBytes: 64 * 1024 * 1024,
+  totalBytes: 256 * 1024 * 1024,
+  steps: 64,
+  errorBytes: 4096,
+} as const;