Mask the nadir/hood region in the reprojection rig; measure ΔAP

Motivation

equirect_tiling.default_views() uses pitch_deg=-30, so the bottom ~40% of every perspective view is the capture vehicle's hood and the black nadir cap — pixels that cannot contain a curb ramp. Roughly a third of every paid VLM call is spent there. It's not just wasteful:

• Grounding DINO's highest-scoring box in a view is the hood itself (0.40), outranking its correct 0.22 box on a real tactile pad. Because AP ranks by score, hood detections at the top of the ranking depress AP directly — a plausible part of why GDINO's AP (0.032) trails OWLv2's (0.104) at similar operating points.

Change

Mask/crop the nadir + hood band before the model call (or drop detections whose center falls inside it). The band is a function of the rig geometry (pitch_deg, fov_v_deg), so it can be computed from the view, not hand-tuned per city. Consider raising/adjusting the pitch as an alternative to masking, and compare.

Now measurable

This was argued in docs/model_comparison.md (Next increments #1); the open-detector scores make its benefit quantifiable. Report ΔAP and ΔF1 vs the current rig, per model, on both cities, and perspective vs --tiling none side by side. This is also a fairness fix — the challengers are currently charged false positives for a fixed vehicle they can't avoid seeing, and the taxonomy pass (companion issue) will show how many FPs it accounts for.

Refs #39, #43.

Start with equirect_tiling.default_views() and the current reprojection rig, then read docs/model_comparison.md and inspect the perspective versus --tiling none evaluation paths. Compare masking, dropping detections, and pitch adjustment without assuming a hand-tuned band. Done means reporting ΔAP and ΔF1 versus the current rig for each model, on both cities and for both tiling modes.