feat(video-conf): ship MediaPipe assets locally for airgapped support

The WASM runtime and TFLite segmentation models were loaded from
cdn.jsdelivr.net and storage.googleapis.com, which broke on cloud
deployments (CSP blocks the CDN) and on airgapped workspaces (no
internet). Ship them from public/mediapipe/ instead, following the
same pattern as the RNNoise assets in public/noise-suppressor/.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: rodrigok

apps/meteor/client/views/videoConference/livekit/backgroundBlurProcessor.ts

@@ -17,16 +17,15 @@ import { supportsBackgroundBlur } from './backgroundBlurSupport';
  *   **15.6 MB** against 244 KB, and roughly twice the work per frame.
  * - **selfie** is one class at 256×144 — the landscape shape a call actually is, cheap, and blunter around hair.
  *
- * Both are fetched from Google's CDN on first use, so a workspace with no way out to the internet gets a caught
- * failure and no blur. Serving them from `public/`, as the RNNoise assets are, is what would fix that.
+ * Both are served from `public/mediapipe/`, alongside the WASM runtime, so airgapped workspaces work out of the box.
  */
 export const SEGMENTER_MODELS = {
 	multiclass: {
-		url: 'https://storage.googleapis.com/mediapipe-models/image_segmenter/selfie_multiclass_256x256/float32/latest/selfie_multiclass_256x256.tflite',
+		url: '/mediapipe/selfie_multiclass_256x256.tflite',
 		input: { width: 256, height: 256 },
 	},
 	selfie: {
-		url: 'https://storage.googleapis.com/mediapipe-models/image_segmenter/selfie_segmenter_landscape/float16/latest/selfie_segmenter_landscape.tflite',
+		url: '/mediapipe/selfie_segmenter_landscape.tflite',
 		input: { width: 256, height: 144 },
 	},
 } as const;
@@ -38,10 +37,10 @@ export const SEGMENTER = SEGMENTER_MODELS.multiclass;
  * MediaPipe's WASM, which has to match the version of `@mediapipe/tasks-vision` this app depends on — it is the
  * runtime for the JS in the package, not an independent thing. Keep the two in step when the package moves.
  *
- * Both this and the model come from a CDN, so a workspace with no way out to the internet gets a caught failure and
- * no blur. Serving them from `public/`, as the RNNoise assets are, is what would fix that.
+ * Served from `public/mediapipe/wasm/`, copied from the npm package at build time. When `@mediapipe/tasks-vision`
+ * is updated, re-copy the wasm directory contents.
  */
-export const SEGMENTER_WASM = 'https://cdn.jsdelivr.net/npm/@mediapipe/tasks-vision@0.10.14/wasm';
+export const SEGMENTER_WASM = '/mediapipe/wasm';
 
 /**
  * Which confidence mask describes the person, and whether it has to be read inside out.

apps/meteor/public/mediapipe/selfie_multiclass_256x256.tflite

@@ -228,9 +228,9 @@ Two ways of doing it, and which one runs is whichever can — the same arrangeme
 - **The camera's own**, via the `backgroundBlur` constraint. Free: the platform does it before the frames reach us.
   Some platforms let the app control it; others expose an OS-controlled effect that the app can only observe.
 - **Ours**, `BackgroundBlurProcessor` in `apps/meteor/client/views/videoConference/livekit/`: MediaPipe segmentation
-  and a WebGL2 compositor, as a LiveKit `TrackProcessor`. Works anywhere with the modern APIs, and it is not free — it fetches
-  MediaPipe's WASM and a model from a CDN (`cdn.jsdelivr.net`, `storage.googleapis.com`) the first time a level is
-  picked. A workspace with no way out to the internet gets the caught failure and blur stays off.
+  and a WebGL2 compositor, as a LiveKit `TrackProcessor`. Works anywhere with the modern APIs, and it is not free — it loads
+  MediaPipe's WASM runtime and a TFLite model from `public/mediapipe/` the first time a level is picked. Both are
+  shipped with the application, so airgapped workspaces work out of the box.
 
 **Ask `getCapabilities()`, never `applyConstraints`.** `applyConstraints({ backgroundBlur: true })` *resolves
 happily* on a browser that has never heard of the constraint — an unrecognised non-required constraint is dropped