Reference model benchmark

The reference benchmark is a staged pipeline that validates the recording, smoke-checks each pose model, benchmarks all selected models on shared detections, and writes per-frame and aggregate results.

uv run aprilpose-benchmark --max-active-frames 300

Outputs are written to outputs/reference_benchmark/:

• frames.csv contains per-frame poses, runtime, and refinement diagnostics.
• summary.json contains dataset inspection, model checks, translation and SO(3) orientation-change metrics, and timing summaries.

Available models are baseline, subpix_ippe_temporal, and internal_lines. Select a subset with:

uv run aprilpose-benchmark --models baseline internal_lines

notebooks/ref_pyfile.py remains as a compatibility launcher. The reusable implementation lives under aprilpose/benchmark/.