@myzonerocks/gosslens 0.10.1 → 0.11.1

package/dist/src/mic-input.d.ts

@@ -1,4 +1,4 @@
-import type { GossSession } from "./index";
+import type { GossSession } from "./index.js";
 export declare class GossMicInput {
     private session;
     private context;

package/dist/src/mic-input.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"mic-input.d.ts","sourceRoot":"","sources":["../../src/mic-input.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,SAAS,CAAC;AAiB3C,qBAAa,YAAY;IACvB,OAAO,CAAC,OAAO,CAAc;IAC7B,OAAO,CAAC,OAAO,CAA6B;IAC5C,OAAO,CAAC,MAAM,CAA4B;IAC1C,OAAO,CAAC,IAAI,CAAiC;IAC7C,OAAO,CAAC,MAAM,CAA2C;IAEzD,YAAY,OAAO,EAAE,WAAW,EAE/B;IAKK,KAAK,IAAI,OAAO,CAAC,IAAI,CAAC,CAuB3B;IAGK,IAAI,IAAI,OAAO,CAAC,IAAI,CAAC,CAU1B;CACF"}
+{"version":3,"file":"mic-input.d.ts","sourceRoot":"","sources":["../../src/mic-input.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,YAAY,CAAC;AAiB9C,qBAAa,YAAY;IACvB,OAAO,CAAC,OAAO,CAAc;IAC7B,OAAO,CAAC,OAAO,CAA6B;IAC5C,OAAO,CAAC,MAAM,CAA4B;IAC1C,OAAO,CAAC,IAAI,CAAiC;IAC7C,OAAO,CAAC,MAAM,CAA2C;IAEzD,YAAY,OAAO,EAAE,WAAW,EAE/B;IAKK,KAAK,IAAI,OAAO,CAAC,IAAI,CAAC,CAuB3B;IAGK,IAAI,IAAI,OAAO,CAAC,IAAI,CAAC,CAU1B;CACF"}

package/dist/src/video-texture.d.ts

@@ -1,4 +1,4 @@
-import type { GossSession } from "./index";
+import type { GossSession } from "./index.js";
 export declare class GossVideoTexture {
     private session;
     private name;

package/dist/src/video-texture.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"video-texture.d.ts","sourceRoot":"","sources":["../../src/video-texture.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,SAAS,CAAC;AAM3C,qBAAa,gBAAgB;IAC3B,OAAO,CAAC,OAAO,CAAc;IAC7B,OAAO,CAAC,IAAI,CAAS;IACrB,OAAO,CAAC,KAAK,CAAmB;IAChC,OAAO,CAAC,MAAM,CAAoB;IAClC,OAAO,CAAC,GAAG,CAAkC;IAC7C,OAAO,CAAC,OAAO,CAAS;IAExB,YAAY,OAAO,EAAE,WAAW,EAAE,IAAI,EAAE,MAAM,EAAE,GAAG,EAAE,MAAM,EAU1D;IAIK,KAAK,IAAI,OAAO,CAAC,IAAI,CAAC,CAK3B;IAKD,IAAI,IAAI,IAAI,CAYX;IAGD,IAAI,IAAI,IAAI,CAKX;CACF"}
+{"version":3,"file":"video-texture.d.ts","sourceRoot":"","sources":["../../src/video-texture.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,YAAY,CAAC;AAM9C,qBAAa,gBAAgB;IAC3B,OAAO,CAAC,OAAO,CAAc;IAC7B,OAAO,CAAC,IAAI,CAAS;IACrB,OAAO,CAAC,KAAK,CAAmB;IAChC,OAAO,CAAC,MAAM,CAAoB;IAClC,OAAO,CAAC,GAAG,CAAkC;IAC7C,OAAO,CAAC,OAAO,CAAS;IAExB,YAAY,OAAO,EAAE,WAAW,EAAE,IAAI,EAAE,MAAM,EAAE,GAAG,EAAE,MAAM,EAU1D;IAIK,KAAK,IAAI,OAAO,CAAC,IAAI,CAAC,CAK3B;IAKD,IAAI,IAAI,IAAI,CAYX;IAGD,IAAI,IAAI,IAAI,CAKX;CACF"}

package/dist/src/world.d.ts

@@ -1,4 +1,4 @@
-import type { GossSession } from "./index";
+import type { GossSession } from "./index.js";
 interface XRRigidTransformLike {
     matrix: Float32Array;
 }

package/dist/src/world.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"world.d.ts","sourceRoot":"","sources":["../../src/world.ts"],"names":[],"mappings":"AAIA,OAAO,KAAK,EAAE,WAAW,EAAwC,MAAM,SAAS,CAAC;AAIjF,UAAU,oBAAoB;IAC5B,MAAM,EAAE,YAAY,CAAC;CACtB;AAED,UAAU,UAAU;IAClB,gBAAgB,EAAE,YAAY,CAAC;CAChC;AAED,UAAU,gBAAgB;IACxB,SAAS,EAAE,oBAAoB,CAAC;IAChC,KAAK,EAAE,UAAU,EAAE,CAAC;IACpB,gBAAgB,CAAC,EAAE,OAAO,CAAC;CAC5B;AAED,UAAU,WAAW;IACnB,UAAU,EAAE,OAAO,CAAC;CACrB;AAED,UAAU,YAAY;IACpB,WAAW,EAAE,OAAO,CAAC;CACtB;AAED,UAAU,UAAU;IAClB,SAAS,EAAE,oBAAoB,CAAC;CACjC;AAED,MAAM,WAAW,eAAe;IAC9B,aAAa,CAAC,cAAc,EAAE,OAAO,GAAG,gBAAgB,GAAG,IAAI,CAAC;IAChE,OAAO,CAAC,KAAK,EAAE,OAAO,EAAE,cAAc,EAAE,OAAO,GAAG,UAAU,GAAG,IAAI,CAAC;IACpE,cAAc,CAAC,EAAE,GAAG,CAAC,WAAW,CAAC,CAAC;IAClC,cAAc,CAAC,EAAE,GAAG,CAAC,YAAY,CAAC,CAAC;CACpC;AAED,qBAAa,oBAAoB;IAKnB,OAAO,CAAC,QAAQ,CAAC,OAAO;IAJpC,OAAO,CAAC,QAAQ,CAAkC;IAClD,OAAO,CAAC,SAAS,CAAmC;IACpD,OAAO,CAAC,MAAM,CAAK;IAEnB,YAA6B,OAAO,EAAE,WAAW,EAAI;IAIrD,OAAO,CAAC,KAAK,EAAE,eAAe,EAAE,cAAc,EAAE,OAAO,EAAE,WAAW,EAAE,MAAM,GAAG,IAAI,CA6ClF;IAED,OAAO,CAAC,KAAK;CAQd"}
+{"version":3,"file":"world.d.ts","sourceRoot":"","sources":["../../src/world.ts"],"names":[],"mappings":"AAIA,OAAO,KAAK,EAAE,WAAW,EAAwC,MAAM,YAAY,CAAC;AAIpF,UAAU,oBAAoB;IAC5B,MAAM,EAAE,YAAY,CAAC;CACtB;AAED,UAAU,UAAU;IAClB,gBAAgB,EAAE,YAAY,CAAC;CAChC;AAED,UAAU,gBAAgB;IACxB,SAAS,EAAE,oBAAoB,CAAC;IAChC,KAAK,EAAE,UAAU,EAAE,CAAC;IACpB,gBAAgB,CAAC,EAAE,OAAO,CAAC;CAC5B;AAED,UAAU,WAAW;IACnB,UAAU,EAAE,OAAO,CAAC;CACrB;AAED,UAAU,YAAY;IACpB,WAAW,EAAE,OAAO,CAAC;CACtB;AAED,UAAU,UAAU;IAClB,SAAS,EAAE,oBAAoB,CAAC;CACjC;AAED,MAAM,WAAW,eAAe;IAC9B,aAAa,CAAC,cAAc,EAAE,OAAO,GAAG,gBAAgB,GAAG,IAAI,CAAC;IAChE,OAAO,CAAC,KAAK,EAAE,OAAO,EAAE,cAAc,EAAE,OAAO,GAAG,UAAU,GAAG,IAAI,CAAC;IACpE,cAAc,CAAC,EAAE,GAAG,CAAC,WAAW,CAAC,CAAC;IAClC,cAAc,CAAC,EAAE,GAAG,CAAC,YAAY,CAAC,CAAC;CACpC;AAED,qBAAa,oBAAoB;IAKnB,OAAO,CAAC,QAAQ,CAAC,OAAO;IAJpC,OAAO,CAAC,QAAQ,CAAkC;IAClD,OAAO,CAAC,SAAS,CAAmC;IACpD,OAAO,CAAC,MAAM,CAAK;IAEnB,YAA6B,OAAO,EAAE,WAAW,EAAI;IAIrD,OAAO,CAAC,KAAK,EAAE,eAAe,EAAE,cAAc,EAAE,OAAO,EAAE,WAAW,EAAE,MAAM,GAAG,IAAI,CA6ClF;IAED,OAAO,CAAC,KAAK;CAQd"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@myzonerocks/gosslens",
-  "version": "0.10.1",
+  "version": "0.11.1",
   "description": "Web SDK for the gosslens camera and AR engine, over the wasm core.",
   "license": "Apache-2.0",
   "repository": {

package/src/audio-output.ts

@@ -1,4 +1,4 @@
-import type { GossSession } from "./index";
+import type { GossSession } from "./index.js";
 
 /// The worklet drains chunks the page posts it, so the audio thread
 /// never touches the wasm heap and the engine's graph-thread pull

package/src/index.ts

@@ -163,6 +163,28 @@ export interface GossPoseOut {
 
 export const GOSS_HAND_LANDMARK_COUNT = 21;
 export const GOSS_HAND_MAX = 2;
+
+export const GOSS_CAP_TRACKING = 1 << 0;
+export const GOSS_CAP_SEGMENTATION = 1 << 1;
+export const GOSS_CAP_ML_INFER = 1 << 2;
+export const GOSS_CAP_DIFFUSION = 1 << 3;
+export const GOSS_CAP_BEAUTY = 1 << 4;
+export const GOSS_CAP_PHYSICS = 1 << 5;
+export const GOSS_CAP_VIDEO_TEXTURES = 1 << 6;
+export const GOSS_CAP_PHOTO_CAPTURE = 1 << 7;
+export const GOSS_CAP_RECORDING = 1 << 8;
+export const GOSS_CAP_FILE_IO = 1 << 9;
+
+/// A hand handed to submitHands: the tracker's presence, handedness, gesture,
+/// and GOSS_HAND_LANDMARK_COUNT * 3 frame-pixel landmark floats. The shape
+/// GossHandTracker.process returns plugs in directly.
+export interface GossHandInput {
+  presence?: number;
+  handedness?: number;
+  gesture?: number;
+  gestureScore?: number;
+  landmarks: Float32Array | number[];
+}
 const HAND_RESULT_BYTES = 560;
 const HAND_ONE_BYTES = 268;
 
@@ -371,6 +393,13 @@ export class Gosslens {
     return this.version;
   }
 
+  /// Which capabilities this build compiled real, as GOSS_CAP_* bits. A stub
+  /// build shares the full one's filename and abi version, so check the rail
+  /// you need here before feeding real bytes to an enable call.
+  capabilities(): number {
+    return this.mod.ccall("goss_capabilities", "number", [], []) as number;
+  }
+
   /// Loads gosslens_web.js and checks its ABI major version. A
   /// dynamic import, not static: bun's bundler would otherwise inline
   /// this file, breaking Emscripten's own import.meta.url-relative
@@ -498,8 +527,12 @@ export class GossEngine {
     // itself, via emscripten_webgl_create_context - passing
     // webGLContextAttributes here has no effect regardless,
     // preserveDrawingBuffer stays false. Worked around in capturePixels.
-    const rendererStatus = await mod.ccall("goss_engine_init_renderer", "number", ["number", "number"], [this.handle, rendererDescPtr], { async: true });
-    mod.ccall("goss_free", null, ["number", "number"], [rendererDescPtr, 12]);
+    let rendererStatus: number;
+    try {
+      rendererStatus = await mod.ccall("goss_engine_init_renderer", "number", ["number", "number"], [this.handle, rendererDescPtr], { async: true });
+    } finally {
+      mod.ccall("goss_free", null, ["number", "number"], [rendererDescPtr, 12]);
+    }
     if (rendererStatus !== GOSS_OK) throw new Error(`renderer init failed: ${rendererStatus}`);
 
     // Emscripten's C++ side just created this canvas's own rendering
@@ -1103,7 +1136,8 @@ export class GossSession {
     if (bytes > this.scratchCapacity) {
       if (this.scratchPtr !== 0) this.mod.ccall("goss_free", null, ["number", "number"], [this.scratchPtr, this.scratchCapacity]);
       this.scratchPtr = this.mod.ccall("goss_alloc", "number", ["number"], [bytes]) as number;
-      this.scratchCapacity = bytes;
+      this.scratchCapacity = this.scratchPtr === 0 ? 0 : bytes;
+      if (this.scratchPtr === 0) throw new Error("wasm allocation failed");
     }
     return this.scratchPtr;
   }
@@ -1264,7 +1298,7 @@ export class GossSession {
   /// muxed recording track, unused by the web preview path.
   submitAudio(samples: Float32Array, frameCount: number, sampleRate: number, channels: number, timestampUs: bigint = 0n): void {
     const bytes = samples.length * 4;
-    const ptr = this.mod.ccall("goss_alloc", "number", ["number"], [bytes]) as number;
+    const ptr = this.scratch(bytes);
     this.mod.HEAPF32.set(samples, ptr >> 2);
     this.mod.ccall(
       "goss_session_submit_audio",
@@ -1272,18 +1306,15 @@ export class GossSession {
       ["number", "number", "number", "number", "number", "number"],
       [this.handle, ptr, frameCount, sampleRate, channels, timestampUs],
     );
-    this.mod.ccall("goss_free", null, ["number", "number"], [ptr, bytes]);
   }
 
   /// Pulls the next block of mixed lens audio (frames interleaved s16) that
   /// play_sound triggers produced, for the page to feed into WebAudio.
   pullAudio(frames: number): Int16Array {
     const byteLen = frames * 2;
-    const ptr = this.mod.ccall("goss_alloc", "number", ["number"], [byteLen]);
+    const ptr = this.scratch(byteLen);
     this.mod.ccall("goss_session_pull_audio", "number", ["number", "number", "number"], [this.handle, ptr, frames]);
-    const out = new Int16Array(this.mod.HEAP16.buffer, ptr, frames).slice();
-    this.mod.ccall("goss_free", null, ["number", "number"], [ptr, byteLen]);
-    return out;
+    return new Int16Array(this.mod.HEAP16.buffer, ptr, frames).slice();
   }
 
   /// Folds the active lens sound into the caller's outgoing call/live track:
@@ -1292,12 +1323,12 @@ export class GossSession {
   /// mixed interleaved s16. Advances the mixer once, replacing `pullAudio`.
   mixOutputAudio(mic: Float32Array | null, frameCount: number, sampleRate: number, channels: number): Int16Array {
     const outLen = frameCount * channels;
-    const outBytes = outLen * 2;
-    const outPtr = this.mod.ccall("goss_alloc", "number", ["number"], [outBytes]);
-    let micPtr = 0;
+    const outBytes = (outLen * 2 + 3) & ~3;
     const micBytes = mic ? mic.length * 4 : 0;
+    const outPtr = this.scratch(outBytes + micBytes);
+    let micPtr = 0;
     if (mic) {
-      micPtr = this.mod.ccall("goss_alloc", "number", ["number"], [micBytes]);
+      micPtr = outPtr + outBytes;
       this.mod.HEAPF32.set(mic, micPtr >> 2);
     }
     this.mod.ccall(
@@ -1306,10 +1337,7 @@ export class GossSession {
       ["number", "number", "number", "number", "number", "number"],
       [this.handle, micPtr, outPtr, frameCount, sampleRate, channels],
     );
-    const out = new Int16Array(this.mod.HEAP16.buffer, outPtr, outLen).slice();
-    this.mod.ccall("goss_free", null, ["number", "number"], [outPtr, outBytes]);
-    if (mic) this.mod.ccall("goss_free", null, ["number", "number"], [micPtr, micBytes]);
-    return out;
+    return new Int16Array(this.mod.HEAP16.buffer, outPtr, outLen).slice();
   }
 
   /// Stores validated camera-hardware intent; the engine normalizes it. Read it
@@ -1454,7 +1482,7 @@ export class GossSession {
   /// Uploads one RGBA/BGRA frame into a named source (pixelFormat 3 BGRA, 4 RGBA).
   submitSourceFrameRgba(name: string, rgba: Uint8Array, width: number, height: number, stride: number, pixelFormat: GossPixelFormat = GossPixelFormat.Rgba8): void {
     const byteLen = stride * height;
-    const rgbaPtr = this.mod.ccall("goss_alloc", "number", ["number"], [byteLen]) as number;
+    const rgbaPtr = this.scratch(byteLen);
     this.mod.HEAPU8.set(rgba.subarray(0, byteLen), rgbaPtr);
     this.mod.setValue(this.frameDescPtr, width, "i32");
     this.mod.setValue(this.frameDescPtr + 4, height, "i32");
@@ -1466,7 +1494,6 @@ export class GossSession {
     this.mod.setValue(this.frameDescPtr + 28, 0, "i32");
     this.withName(name, (ptr, len) =>
       this.mod.ccall("goss_session_submit_source_frame_rgba_copy", "number", ["number", "number", "number", "number", "number", "number"], [this.handle, ptr, len, this.frameDescPtr, rgbaPtr, stride]));
-    this.mod.ccall("goss_free", null, ["number", "number"], [rgbaPtr, byteLen]);
   }
 
   /// Arranges the camera and named sources: 0 custom, 1 side-by-side, 2 top-bottom, 3 pip, 4 grid.
@@ -1635,7 +1662,7 @@ export class GossSession {
     this.mod.ccall("goss_session_brush_clear", "number", ["number"], [this.handle]);
   }
 
-  /// The brush preset the next stroke opens with: 0 pen, 1 highlighter, 2 marker, 3 neon.
+  /// The brush preset the next stroke opens with: 0 pen, 1 highlighter, 2 marker, 3 neon, 4 stamp.
   setBrushMode(mode: number): void {
     this.mod.ccall("goss_session_brush_set_mode", "number", ["number", "number"], [this.handle, mode]);
   }
@@ -1769,7 +1796,7 @@ export class GossSession {
     }
     const scaleX = this.frameWidth / sourceWidth;
     const scaleY = this.frameHeight / sourceHeight;
-    const pointCount = landmarks.length / 3;
+    const pointCount = Math.min(landmarks.length / 3, GOSS_FACE_LANDMARK_COUNT);
     const base = this.landmarksPtr >> 2;
     for (let at = 0; at < pointCount; at += 1) {
       this.mod.HEAPF32[base + at * 3] = landmarks[at * 3]! * scaleX;
@@ -1800,15 +1827,15 @@ export class GossSession {
       const off = i * FACE_RESULT_BYTES;
       const base = ptr + off;
       const f = faces[i]!;
-      const count = f.landmarks.length / 3;
+      const count = Math.min(f.landmarks.length / 3, GOSS_FACE_LANDMARK_COUNT);
       dv.setBigUint64(off, BigInt(f.frameSerial ?? 0), true);
       dv.setBigInt64(off + 8, BigInt(f.timestampUs ?? 0), true);
       dv.setFloat32(off + 16, f.presence ?? 1, true);
       dv.setUint32(off + 20, count, true);
-      this.mod.HEAPF32.set(f.landmarks, (base + 24) >> 2);
+      this.mod.HEAPF32.set(f.landmarks.subarray(0, count * 3), (base + 24) >> 2);
       const bsStart = (base + 24 + GOSS_FACE_LANDMARK_COUNT * 3 * 4) >> 2;
       this.mod.HEAPF32.fill(0, bsStart, bsStart + GOSS_FACE_BLENDSHAPE_COUNT);
-      if (f.blendshapes) this.mod.HEAPF32.set(f.blendshapes, bsStart);
+      if (f.blendshapes) this.mod.HEAPF32.set(f.blendshapes.subarray(0, GOSS_FACE_BLENDSHAPE_COUNT), bsStart);
     }
     this.mod.ccall("goss_session_submit_faces", "number", ["number", "number", "number"], [this.handle, ptr, faces.length]);
   }
@@ -1865,20 +1892,48 @@ export class GossSession {
       const off = i * POSE_RESULT_BYTES;
       const base = ptr + off;
       const b = bodies[i]!;
-      const count = b.landmarks.length / 3;
+      const count = Math.min(b.landmarks.length / 3, GOSS_POSE_LANDMARK_COUNT);
       dv.setBigUint64(off, BigInt(b.frameSerial ?? 0), true);
       dv.setBigInt64(off + 8, BigInt(b.timestampUs ?? 0), true);
       dv.setFloat32(off + 16, b.presence ?? 1, true);
       dv.setUint32(off + 20, count, true);
-      this.mod.HEAPF32.set(b.landmarks, (base + 24) >> 2);
+      this.mod.HEAPF32.set(b.landmarks.subarray(0, count * 3), (base + 24) >> 2);
       const visStart = (base + 24 + GOSS_POSE_LANDMARK_COUNT * 3 * 4) >> 2;
       this.mod.HEAPF32.fill(0, visStart, visStart + GOSS_POSE_LANDMARK_COUNT * 2);
-      if (b.visibilities) this.mod.HEAPF32.set(b.visibilities, visStart);
-      if (b.presences) this.mod.HEAPF32.set(b.presences, visStart + GOSS_POSE_LANDMARK_COUNT);
+      if (b.visibilities) this.mod.HEAPF32.set(b.visibilities.subarray(0, GOSS_POSE_LANDMARK_COUNT), visStart);
+      if (b.presences) this.mod.HEAPF32.set(b.presences.subarray(0, GOSS_POSE_LANDMARK_COUNT), visStart + GOSS_POSE_LANDMARK_COUNT);
     }
     this.mod.ccall("goss_session_submit_bodies", "number", ["number", "number", "number"], [this.handle, ptr, bodies.length]);
   }
 
+  /// Submits the hands tracked this frame from the page's own tracker (the
+  /// GossHandTracker's result plugs in directly), so hands.gesture and
+  /// hands.pinch lenses and handJoint work with no in-engine worker; the
+  /// submitted hands win over the worker while set. Null clears the path.
+  submitHands(hands: { hands: GossHandInput[]; timestampUs?: bigint } | null): void {
+    if (hands === null) {
+      this.mod.ccall("goss_session_submit_hands", "number", ["number", "number"], [this.handle, 0]);
+      return;
+    }
+    const ptr = this.scratch(HAND_RESULT_BYTES);
+    this.mod.HEAPU8.fill(0, ptr, ptr + HAND_RESULT_BYTES);
+    const dv = new DataView(this.mod.HEAPU8.buffer, ptr, HAND_RESULT_BYTES);
+    dv.setBigInt64(8, hands.timestampUs ?? 0n, true);
+    const list = hands.hands.slice(0, GOSS_HAND_MAX);
+    dv.setUint32(16, list.length, true);
+    for (let i = 0; i < list.length; i += 1) {
+      const off = 24 + i * HAND_ONE_BYTES;
+      const h = list[i]!;
+      dv.setFloat32(off, h.presence ?? 1, true);
+      dv.setFloat32(off + 4, h.handedness ?? 0.5, true);
+      dv.setUint32(off + 8, h.gesture ?? 0, true);
+      dv.setFloat32(off + 12, h.gestureScore ?? 0, true);
+      const floats = Math.min(h.landmarks.length, GOSS_HAND_LANDMARK_COUNT * 3);
+      for (let k = 0; k < floats; k += 1) dv.setFloat32(off + 16 + k * 4, h.landmarks[k]!, true);
+    }
+    this.mod.ccall("goss_session_submit_hands", "number", ["number", "number"], [this.handle, ptr]);
+  }
+
   /// Submits one frame's depth map from the host AR backend (WebXR depth-
   /// sensing): width by height metres per pixel, row major, with the near and
   /// far metres that bound it. An empty array clears it. Kept for depth
@@ -2039,6 +2094,75 @@ export class GossSession {
     return result;
   }
 
+  /// Hands the engine one bundle asset's bytes under its manifest name ahead
+  /// of activateLens, so a lens's models, references, and label files load
+  /// from memory - the door the web's inference nodes activate through. An
+  /// empty array removes a staged name; the engine keeps its own copy.
+  provideLensAsset(name: string, bytes: Uint8Array): void {
+    const id = new TextEncoder().encode(name);
+    const idPtr = this.mod.ccall("goss_alloc", "number", ["number"], [id.length || 1]) as number;
+    this.mod.HEAPU8.set(id, idPtr);
+    let bytesPtr = 0;
+    if (bytes.length > 0) {
+      bytesPtr = this.mod.ccall("goss_alloc", "number", ["number"], [bytes.length]) as number;
+      this.mod.HEAPU8.set(bytes, bytesPtr);
+    }
+    const args = ["number", "number", "number", "number", "number"];
+    const status = this.mod.ccall("goss_session_provide_lens_asset", "number", args, [this.handle, idPtr, id.length, bytesPtr, bytes.length]) as number;
+    if (bytesPtr !== 0) this.mod.ccall("goss_free", null, ["number", "number"], [bytesPtr, bytes.length]);
+    this.mod.ccall("goss_free", null, ["number", "number"], [idPtr, id.length || 1]);
+    if (status !== 0) throw new Error(`provideLensAsset failed (${status})`);
+  }
+
+  /// One ml.infer node's whole published output tensor, by the node's id and
+  /// tensor index, or null before the model's first publish or when the node
+  /// or tensor does not exist. A length probe sizes the read, so detections,
+  /// embeddings, and logits come back at their own length.
+  mlOutput(nodeId: string, tensor = 0): Float32Array | null {
+    const id = new TextEncoder().encode(nodeId);
+    const idPtr = this.mod.ccall("goss_alloc", "number", ["number"], [id.length || 1]) as number;
+    this.mod.HEAPU8.set(id, idPtr);
+    const lenPtr = this.mod.ccall("goss_alloc", "number", ["number"], [4]) as number;
+    const readLen = () => new DataView(this.mod.HEAPU8.buffer, lenPtr, 4).getUint32(0, true);
+    const args = ["number", "number", "number", "number", "number", "number", "number"];
+    this.mod.ccall("goss_session_ml_output", "number", args, [this.handle, idPtr, id.length, tensor, 0, 0, lenPtr]);
+    const needed = readLen();
+    let result: Float32Array | null = null;
+    if (needed > 0) {
+      const outPtr = this.mod.ccall("goss_alloc", "number", ["number"], [needed * 4]) as number;
+      const status = this.mod.ccall("goss_session_ml_output", "number", args, [this.handle, idPtr, id.length, tensor, outPtr, needed, lenPtr]) as number;
+      if (status === 0) result = this.mod.HEAPF32.slice(outPtr >> 2, (outPtr >> 2) + needed);
+      this.mod.ccall("goss_free", null, ["number", "number"], [outPtr, needed * 4]);
+    }
+    this.mod.ccall("goss_free", null, ["number", "number"], [lenPtr, 4]);
+    this.mod.ccall("goss_free", null, ["number", "number"], [idPtr, id.length || 1]);
+    return result;
+  }
+
+  /// One ml.infer node's mask-bound output resampled to the fixed
+  /// segmentation plane, or null when the node has no mask binding or the
+  /// model has not published yet.
+  mlMask(nodeId: string): Float32Array | null {
+    const id = new TextEncoder().encode(nodeId);
+    const idPtr = this.mod.ccall("goss_alloc", "number", ["number"], [id.length || 1]) as number;
+    this.mod.HEAPU8.set(id, idPtr);
+    const lenPtr = this.mod.ccall("goss_alloc", "number", ["number"], [4]) as number;
+    const readLen = () => new DataView(this.mod.HEAPU8.buffer, lenPtr, 4).getUint32(0, true);
+    const args = ["number", "number", "number", "number", "number", "number"];
+    this.mod.ccall("goss_session_ml_mask", "number", args, [this.handle, idPtr, id.length, 0, 0, lenPtr]);
+    const needed = readLen();
+    let result: Float32Array | null = null;
+    if (needed > 0) {
+      const outPtr = this.mod.ccall("goss_alloc", "number", ["number"], [needed * 4]) as number;
+      const status = this.mod.ccall("goss_session_ml_mask", "number", args, [this.handle, idPtr, id.length, outPtr, needed, lenPtr]) as number;
+      if (status === 0) result = this.mod.HEAPF32.slice(outPtr >> 2, (outPtr >> 2) + needed);
+      this.mod.ccall("goss_free", null, ["number", "number"], [outPtr, needed * 4]);
+    }
+    this.mod.ccall("goss_free", null, ["number", "number"], [lenPtr, 4]);
+    this.mod.ccall("goss_free", null, ["number", "number"], [idPtr, id.length || 1]);
+    return result;
+  }
+
   /// The recent diarized caption segment at index (0 the newest), or null when
   /// the index is past the segments held: the times it spanned, the speaker who
   /// spoke it, and its text, a speaker-tagged transcript for diarized subtitles.
@@ -2089,9 +2213,9 @@ export class GossSession {
   /// Submits one NV12 exposure of an HDR bracket, fed to bracket-source
   /// temporal.fuse nodes. y and uv are the plane bytes as submitFrameCopy takes.
   submitFrameBracket(y: Uint8Array, yStride: number, uv: Uint8Array, uvStride: number, width: number, height: number, colorStandard = 1, colorRange = 0): void {
-    const yPtr = this.mod.ccall("goss_alloc", "number", ["number"], [y.length]) as number;
+    const yPtr = this.scratch(y.length + uv.length);
     this.mod.HEAPU8.set(y, yPtr);
-    const uvPtr = this.mod.ccall("goss_alloc", "number", ["number"], [uv.length]) as number;
+    const uvPtr = yPtr + y.length;
     this.mod.HEAPU8.set(uv, uvPtr);
     this.mod.setValue(this.frameDescPtr, width, "i32");
     this.mod.setValue(this.frameDescPtr + 4, height, "i32");
@@ -2102,14 +2226,12 @@ export class GossSession {
     this.mod.setValue(this.frameDescPtr + 24, 0, "i32");
     this.mod.setValue(this.frameDescPtr + 28, 0, "i32");
     this.mod.ccall("goss_session_submit_frame_bracket", "number", ["number", "number", "number", "number", "number", "number"], [this.handle, this.frameDescPtr, yPtr, yStride, uvPtr, uvStride]);
-    this.mod.ccall("goss_free", null, ["number", "number"], [yPtr, y.length]);
-    this.mod.ccall("goss_free", null, ["number", "number"], [uvPtr, uv.length]);
   }
 
   private submitPlanes(sym: string, y: Uint8Array, yStride: number, uv: Uint8Array, uvStride: number, width: number, height: number, colorStandard: number, colorRange: number): void {
-    const yPtr = this.mod.ccall("goss_alloc", "number", ["number"], [y.length]) as number;
+    const yPtr = this.scratch(y.length + uv.length);
     this.mod.HEAPU8.set(y, yPtr);
-    const uvPtr = this.mod.ccall("goss_alloc", "number", ["number"], [uv.length]) as number;
+    const uvPtr = yPtr + y.length;
     this.mod.HEAPU8.set(uv, uvPtr);
     this.mod.setValue(this.frameDescPtr, width, "i32");
     this.mod.setValue(this.frameDescPtr + 4, height, "i32");
@@ -2120,8 +2242,6 @@ export class GossSession {
     this.mod.setValue(this.frameDescPtr + 24, 0, "i32");
     this.mod.setValue(this.frameDescPtr + 28, 0, "i32");
     this.mod.ccall(sym, "number", ["number", "number", "number", "number", "number", "number"], [this.handle, this.frameDescPtr, yPtr, yStride, uvPtr, uvStride]);
-    this.mod.ccall("goss_free", null, ["number", "number"], [yPtr, y.length]);
-    this.mod.ccall("goss_free", null, ["number", "number"], [uvPtr, uv.length]);
   }
 
   /// Feeds one NV12 frame to the in-engine trackers only (no render), for a host
@@ -2500,7 +2620,8 @@ export class GossSession {
     if (this.framePixelsCapacity >= byteLength) return;
     if (this.framePixelsPtr !== 0) this.mod.ccall("goss_free", null, ["number", "number"], [this.framePixelsPtr, this.framePixelsCapacity]);
     this.framePixelsPtr = this.mod.ccall("goss_alloc", "number", ["number"], [byteLength]);
-    this.framePixelsCapacity = byteLength;
+    this.framePixelsCapacity = this.framePixelsPtr === 0 ? 0 : byteLength;
+    if (this.framePixelsPtr === 0) throw new Error("wasm allocation failed");
   }
 
   /// rotationDegrees omitted means the setVideoFlip state decides (a
@@ -2603,7 +2724,8 @@ export class GossSession {
     if (this.worldScratchLen >= byteLength) return;
     if (this.worldScratchPtr !== 0) this.mod.ccall("goss_free", null, ["number", "number"], [this.worldScratchPtr, this.worldScratchLen]);
     this.worldScratchPtr = this.mod.ccall("goss_alloc", "number", ["number"], [byteLength]);
-    this.worldScratchLen = byteLength;
+    this.worldScratchLen = this.worldScratchPtr === 0 ? 0 : byteLength;
+    if (this.worldScratchPtr === 0) throw new Error("wasm allocation failed");
   }
 
   /// Reports one finished frame: measured whole-pipeline time plus
@@ -2679,9 +2801,14 @@ export class GossPreviewSession {
   static async create(canvas: HTMLCanvasElement, wasmJsUrl: string | URL, events: GossSessionEvents = {}): Promise<GossPreviewSession> {
     const gosslens = await Gosslens.load(canvas, wasmJsUrl);
     const engine = GossEngine.create(gosslens);
-    await engine.initRenderer(canvas);
-    const session = GossSession.create(engine);
-    return new GossPreviewSession(gosslens, engine, session, events);
+    try {
+      await engine.initRenderer(canvas);
+      const session = GossSession.create(engine);
+      return new GossPreviewSession(gosslens, engine, session, events);
+    } catch (err) {
+      engine.destroy();
+      throw err;
+    }
   }
 
   abiVersion(): number {
@@ -2810,8 +2937,10 @@ export class GossPreviewSession {
       this.cameraFrames += 1;
       const width = this.video.videoWidth;
       const height = this.video.videoHeight;
-      this.scratchCanvas.width = width;
-      this.scratchCanvas.height = height;
+      if (this.scratchCanvas.width !== width || this.scratchCanvas.height !== height) {
+        this.scratchCanvas.width = width;
+        this.scratchCanvas.height = height;
+      }
       this.scratchCtx.drawImage(this.video, 0, 0, width, height);
       const pixels = this.scratchCtx.getImageData(0, 0, width, height);
       // Not mirrored here - the demo page's own CSS mirrors the canvas
@@ -2852,8 +2981,8 @@ export class GossPreviewSession {
   }
 }
 
-export { GossWebXRWorldSource } from "./world";
-export type { GossXRFrameLike } from "./world";
-export { GossAudioOutput } from "./audio-output";
-export { GossMicInput } from "./mic-input";
-export { GossVideoTexture } from "./video-texture";
+export { GossWebXRWorldSource } from "./world.js";
+export type { GossXRFrameLike } from "./world.js";
+export { GossAudioOutput } from "./audio-output.js";
+export { GossMicInput } from "./mic-input.js";
+export { GossVideoTexture } from "./video-texture.js";

package/src/mic-input.ts

@@ -1,4 +1,4 @@
-import type { GossSession } from "./index";
+import type { GossSession } from "./index.js";
 
 /// Captures the microphone and feeds it to the engine's audio analysis,
 /// so the level and beat triggers a lens reads fire on web the same way

package/src/video-texture.ts

@@ -1,4 +1,4 @@
-import type { GossSession } from "./index";
+import type { GossSession } from "./index.js";
 
 /// Plays an MP4 through a hidden video element, the browser's own
 /// decoder, and feeds each frame to a named engine source a lens

package/src/world.ts

@@ -2,7 +2,7 @@
 /// planes, and tracked anchors off each XR frame and feeds them into
 /// the session. The page owns the XR session; this source only reads.
 
-import type { GossSession, GossWorldPlane, GossWorldAnchorInput } from "./index";
+import type { GossSession, GossWorldPlane, GossWorldAnchorInput } from "./index.js";
 
 // WebXR's types live outside the DOM lib; the narrow slice this source
 // reads is declared here so no type package enters the build.