Add ask_vlm method for cloud VLM alert verification

[srnangi]

What

Adds Groundlight.ask_vlm(images, query, model_id) which verifies one or two images against a natural-language query by calling the new POST /v1/vlm-queries cloud endpoint. Returns a VLMVerificationResult dataclass with verdict (YES/NO/UNSURE), confidence, reasoning, and token cost.

Pairs with the janzu endpoint (separate PR). The VLM runs in the Groundlight cloud (Bedrock) — no local inference.

How

• Accepts a single image or [full_frame, roi] for the dual-image strategy, reusing parse_supported_image_types for encoding (numpy BGR, PIL, bytes, filename, etc.).
• POSTs multipart images parts with query/model_id as query params, using the existing requests-based pattern from internalapi.py.
• Moves import requests to module level.
• Exports VLMVerificationResult from the package.

Usage

gl = Groundlight()
result = gl.ask_vlm(images=[full_frame, roi], query="Is there a fire?")
if result.verdict == "YES":
    emit_alert()
print(result.confidence, result.reasoning, result.total_cost_usd)

Testing

• 7 unit tests with mocked HTTP (image encoding, dual-image, model_id passthrough, >2-image guard, result parsing) — all passing.

🤖 Generated with Claude Code