feat: implement Iteration 3 — product/receipt/dish recognition

Backend:
- gemini/client.go: refactor to shared callGroq transport; add
  generateVisionContent using llama-3.2-11b-vision-preview model
- gemini/recognition.go: RecognizeReceipt, RecognizeProducts,
  RecognizeDish (vision), ClassifyIngredient (text); shared parseJSON helper
- ingredient/repository.go: add FuzzyMatch (wraps Search, returns best hit)
- recognition/handler.go: POST /ai/recognize-receipt, /ai/recognize-products,
  /ai/recognize-dish; enrichItems with fuzzy match + AI classify fallback;
  parallel multi-image processing with deduplication
- server.go + main.go: wire recognition handler under /ai routes

Flutter:
- pubspec.yaml: add image_picker ^1.1.0
- AndroidManifest.xml: add CAMERA and READ_EXTERNAL_STORAGE permissions
- Info.plist: add NSCameraUsageDescription and NSPhotoLibraryUsageDescription
- recognition_service.dart: RecognitionService wrapping /ai/* endpoints;
  RecognizedItem, ReceiptResult, DishResult models
- scan_screen.dart: mode selector (receipt / products / dish / manual);
  image source picker; loading overlay; navigates to confirm or dish screen
- recognition_confirm_screen.dart: editable list of recognized items;
  inline qty/unit editing; swipe-to-delete; batch-add to pantry
- dish_result_screen.dart: dish name, KBZHU breakdown, similar dishes chips
- app_router.dart: /scan, /scan/confirm, /scan/dish routes (no bottom nav)
- products_screen.dart: FAB now shows bottom sheet with Manual / Scan options

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

backend/internal/gemini/client.go

@@ -11,9 +11,15 @@ import (
 )
 
 const (
-	// Groq — OpenAI-compatible API, free tier, no billing required.
+	// groqAPIURL is the Groq OpenAI-compatible endpoint (free tier, no billing required).
 	groqAPIURL = "https://api.groq.com/openai/v1/chat/completions"
-	groqModel  = "llama-3.3-70b-versatile"
+
+	// groqModel is the default text generation model.
+	groqModel = "llama-3.3-70b-versatile"
+
+	// groqVisionModel supports image inputs in OpenAI vision format.
+	groqVisionModel = "llama-3.2-11b-vision-preview"
+
 	maxRetries = 3
 )
 
@@ -28,16 +34,49 @@ func NewClient(apiKey string) *Client {
 	return &Client{
 		apiKey: apiKey,
 		httpClient: &http.Client{
-			Timeout: 60 * time.Second,
+			Timeout: 90 * time.Second,
 		},
 	}
 }
 
-// generateContent sends a user prompt to Groq and returns the assistant text.
+// generateContent sends text messages to the text-only model.
 func (c *Client) generateContent(ctx context.Context, messages []map[string]string) (string, error) {
+	return c.callGroq(ctx, groqModel, 0.7, messages)
+}
+
+// generateVisionContent sends an image + text prompt to the vision model.
+// imageBase64 must be the raw base64-encoded image data (no data URI prefix).
+// mimeType defaults to "image/jpeg" if empty.
+func (c *Client) generateVisionContent(ctx context.Context, prompt, imageBase64, mimeType string) (string, error) {
+	if mimeType == "" {
+		mimeType = "image/jpeg"
+	}
+	dataURL := fmt.Sprintf("data:%s;base64,%s", mimeType, imageBase64)
+
+	messages := []any{
+		map[string]any{
+			"role": "user",
+			"content": []any{
+				map[string]any{
+					"type":      "image_url",
+					"image_url": map[string]string{"url": dataURL},
+				},
+				map[string]any{
+					"type": "text",
+					"text": prompt,
+				},
+			},
+		},
+	}
+	return c.callGroq(ctx, groqVisionModel, 0.1, messages)
+}
+
+// callGroq is the shared HTTP transport for all Groq requests.
+// messages can be []map[string]string (text) or []any (vision with image content).
+func (c *Client) callGroq(ctx context.Context, model string, temperature float64, messages any) (string, error) {
 	body := map[string]any{
-		"model":       groqModel,
-		"temperature": 0.7,
+		"model":       model,
+		"temperature": temperature,
 		"messages":    messages,
 	}