fix(heal): upgrade inpaint architecture to hybrid 2-tier LaMa ONNX and fix starburst artifacts

- backend: preserve aspect ratio with edge-replicated square padding into 512x512 LaMa model
- backend: fix inverted mask threshold (>0.85*plane) and prune CUDA/TensorRT binaries for slim image
- frontend: enable server AI by default and route masks >3.5% or edge-crossing to LaMa Tier 2
- frontend: enforce 60%-80% context padding (min 64px) aligned to multiples of 8
- frontend: add 3-5px binary mask dilation and high-frequency sensor grain matching
- frontend: anchor deletion badge to topmost point of brush stroke
- build & export: updated docker containers and recipescam-images.tar.gz (357MB)

docker/backend/package.json

@@ -9,7 +9,8 @@
     "start": "node dist/server.js",
     "dev": "tsx watch src/server.ts",
     "test": "node test/security.mjs",
-    "test:immich": "node test/immich.mjs"
+    "test:immich": "node test/immich.mjs",
+    "test:inpaint": "node test/inpaint-onnx.mjs"
   },
   "dependencies": {
     "better-sqlite3": "^12.11.1",