feat: add recognition engine with CLIP matching and Ollama vision LLM

- OllamaClient: analyze_image, compare_images, is_available
- Recognition pipeline: CLIP embedding -> cosine similarity -> threshold
  routing (known >0.85, unknown <0.60, ambiguous -> Ollama fallback)
- Vision dispatcher: get_description() supports haiku/ollama/both backends
- 9 embedding serialization tests

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

app/vision.py

@@ -1,7 +1,11 @@
 import base64
 import logging
+from typing import Optional
+
 import anthropic
 
+from ollama_client import OllamaClient
+
 logger = logging.getLogger(__name__)
 
 PROMPT = (
@@ -40,3 +44,66 @@ def analyze_snapshot(client: anthropic.Anthropic, image_bytes: bytes) -> str | N
     except Exception:
         logger.exception("Claude Vision API failed")
         return None
+
+
+async def analyze_snapshot_ollama(
+    ollama_client: OllamaClient, image_bytes: bytes
+) -> str | None:
+    """Send a snapshot to Ollama vision model for description.
+
+    Args:
+        ollama_client: Configured OllamaClient instance.
+        image_bytes: Raw image bytes (JPEG).
+
+    Returns:
+        Description string or None on failure.
+    """
+    return await ollama_client.analyze_image(image_bytes, PROMPT)
+
+
+async def get_description(
+    image_bytes: bytes,
+    vision_backend: str,
+    claude_client: Optional[anthropic.Anthropic] = None,
+    ollama_client: Optional[OllamaClient] = None,
+) -> str | None:
+    """Get a scene description using the configured vision backend.
+
+    Args:
+        image_bytes: Raw image bytes (JPEG).
+        vision_backend: One of "haiku", "ollama", or "both".
+            "both" tries Ollama first, falls back to Haiku.
+        claude_client: Anthropic client (required for "haiku" and "both").
+        ollama_client: OllamaClient instance (required for "ollama" and "both").
+
+    Returns:
+        Description string or None if all backends fail.
+    """
+    if vision_backend == "ollama":
+        if ollama_client is None:
+            logger.error("Ollama backend selected but no client provided")
+            return None
+        return await analyze_snapshot_ollama(ollama_client, image_bytes)
+
+    if vision_backend == "haiku":
+        if claude_client is None:
+            logger.error("Haiku backend selected but no client provided")
+            return None
+        return analyze_snapshot(claude_client, image_bytes)
+
+    if vision_backend == "both":
+        # Try Ollama first, fall back to Haiku
+        if ollama_client is not None:
+            result = await analyze_snapshot_ollama(ollama_client, image_bytes)
+            if result is not None:
+                return result
+            logger.info("Ollama failed, falling back to Haiku")
+
+        if claude_client is not None:
+            return analyze_snapshot(claude_client, image_bytes)
+
+        logger.error("Both backend selected but no clients provided")
+        return None
+
+    logger.error("Unknown vision backend: %s", vision_backend)
+    return None