feat: wire Gemma 4 vision into browser_tool for screenshots (#816)

Resolves #816. Updated _get_vision_model() to prefer Gemma 4
native multimodal for screenshot analysis.

Priority chain:
1. AUXILIARY_VISION_MODEL (explicit override)
2. GEMMA_VISION_MODEL (Gemma 4 native multimodal)
3. OLLAMA_VISION_MODEL (local Ollama vision)
4. None (text-only fallback)

Reduces latency by eliminating model switching for vision tasks.
Backward compatible — existing AUXILIARY_VISION_MODEL still works.

tools/browser_tool.py

@@ -201,8 +201,31 @@ def _get_command_timeout() -> int:
 
 
 def _get_vision_model() -> Optional[str]:
-    """Model for browser_vision (screenshot analysis — multimodal)."""
-    return os.getenv("AUXILIARY_VISION_MODEL", "").strip() or None
+    """Model for browser_vision (screenshot analysis — multimodal).
+
+    Priority:
+    1. AUXILIARY_VISION_MODEL env var (explicit override)
+    2. Gemma 4 (native multimodal, no model switching)
+    3. Ollama local vision models
+    4. None (fallback to text-only snapshot)
+    """
+    # Explicit override always wins
+    explicit = os.getenv("AUXILIARY_VISION_MODEL", "").strip()
+    if explicit:
+        return explicit
+
+    # Prefer Gemma 4 (native multimodal — no separate vision model needed)
+    gemma = os.getenv("GEMMA_VISION_MODEL", "").strip()
+    if gemma:
+        return gemma
+
+    # Check for Ollama vision models
+    ollama_vision = os.getenv("OLLAMA_VISION_MODEL", "").strip()
+    if ollama_vision:
+        return ollama_vision
+
+    # Default: None (text-only fallback)
+    return None
 
 
 def _get_extraction_model() -> Optional[str]: