feat: add vision model preprocessor with persistent media cache

[StepanchukYI]

Summary

• Adds vision model support as a preprocessing step: images in messages are transcribed by a vision-capable model before being sent to the main agent LLM
• Introduces MediaCache — a persistent SHA-256 keyed JSON cache that avoids re-transcribing the same images
• Configurable via vision_model in agent defaults (e.g. vision_model: "openai/gpt-4o")
• Auto-detects vision-capable models by pattern matching (gpt-4o, claude-3, gemini, etc.)

How it works

1. When a message contains images and the main model doesn't support vision, the vision model transcribes them
2. Transcriptions are cached by SHA-256 hash of the image content
3. The text description replaces the image in the message sent to the main model

Test plan

• Existing tests pass (339/339)
• Verify vision preprocessing works with image messages
• Verify cache persists and avoids redundant API calls
• Verify pass-through when main model supports vision natively

🤖 Generated with Claude Code