Add Ox Alpha stealth model with bbox_2d pixel detection format

[SkalskiP]

Summary

• Adds the Ox Alpha stealth model (stealth/ox-alpha on OpenRouter) to configs/models.yaml, along with a new stealth lab entry. Pricing is 0.00/0.00: stealth models are free on OpenRouter during preview.
• Adds stealth/ox-alpha to _REASONING_REQUIRED_MODELS in the OpenRouter provider; the model rejects reasoning.enabled=False, so low-effort runs send {"effort": "low"} instead.
• Adds a new detection coordinate format xyxy_absolute_original_image_bbox_2d: a Qwen2.5-VL-style prompt asking for {"bbox_2d": [x1, y1, x2, y2], "label": ...} in original-image pixel coordinates, parsed by the existing absolute-pixel parser via a bbox_2d key fallback.
• Documents the new format in AGENTS.md and README.md; adds tests for the prompt, the parser, and the reasoning config.

Coordinate format justification

Ox Alpha is a stealth model with no official documentation (source: https://openrouter.ai/stealth/ox-alpha), so the format was chosen via a local 20-image detection probe. An open-ended dump showed the model natively emits original-image pixel boxes under bbox/box/bbox_2d keys. mAP@50 across the probe matrix:

Format	mAP@50
xyxy_absolute_original_image_bbox_2d	0.4068
yxyx_normalized_0_to_1000	0.3436
xyxy_absolute_original_image	0.0475
xyxy_normalized_0_to_100	0.0363
xyxy_normalized_0_to_1000	0.0027

Notes

• No benchmark results are committed; Ox Alpha stays off the public leaderboards while it remains a stealth model. Its runs live in local probe directories.
• web/benchmark_summary.json was regenerated and is byte-identical (no results in results/), so it is not part of the diff.

Test plan

• ruff check src tests and ruff format --check src tests pass
• pytest tests/ passes (205 tests)
• Bugbot review of the diff: no findings

Made with Cursor