fix: repair recognition — migrate vision model and fix XFile handling

- Replace decommissioned llama-3.2-11b-vision-preview with
  meta-llama/llama-4-scout-17b-16e-instruct (Groq deprecation)
- Use XFile.readAsBytes() instead of File(path).readAsBytes() so
  Android content URIs (from gallery picks) are read correctly
- Add maxWidth/maxHeight constraints to image picker calls to reduce
  payload size
- Increase receiveTimeout from 30s to 120s to accommodate slow vision AI
- Log recognition errors via debugPrint instead of swallowing them

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

backend/internal/gemini/client.go

@@ -18,7 +18,7 @@ const (
 	groqModel = "llama-3.3-70b-versatile"
 
 	// groqVisionModel supports image inputs in OpenAI vision format.
-	groqVisionModel = "llama-3.2-11b-vision-preview"
+	groqVisionModel = "meta-llama/llama-4-scout-17b-16e-instruct"
 
 	maxRetries = 3
 )