Add the 1k manual gold-standard set as a model-comparison split (bigger, independently-labeled, un-anchored GT)

Motivation

The model-comparison harness (#34 / #36 / #40) scores RampNet vs the VLM/open-detector field on the richmond/bend/clovis verdict GT. That GT is deliberately deployment/OOD-focused, but it is explicitly "RampNet-anchored" — derived from a review of RampNet's own detections plus a missed-ramp pass (the docs/model_comparison.md caveat, and the bias behind #55).

We already have a bigger, independently-labeled, un-anchored gold set and aren't using it in the comparison: the 1,000-panorama Manual Ground Truth Set (3,919 curb ramps). It is:

• Fully and manually labeled from scratch by humans (not seeded by RampNet) → no RampNet-anchoring, so it sidesteps #55 entirely for this data and is fairer to the VLMs too (their recall isn't measured against a GT built from RampNet's hits).
• 4× the size of the current benchmark (1,000 vs 234) → tighter Wilson CIs.
• The paper's canonical gold standard — README.md (§"Evaluating") says we "place more emphasis on [it]… less prone to errors… directly from a human source."

Where the data is

• Images: in projectsidewalk/rampnet-dataset on HF (the manual-eval panos are held out from train/val, so they live in the test split — see README data-leakage note).
• Labels: in-repo at manual_labels/*.txt, YOLO format (class cx cy w h, normalized), keyed by pano id.

Important caveat — it's in-distribution (GSV)

The gold set is GSV, sampled from cities including some in training (NYC/Portland/Bend; docs/data_provenance.md flags in-training cities as optimistically biased). It is held out from Stage-2 training, but it's the same imagery source + overlapping geography RampNet was built for, so RampNet has home-field advantage and its margin here will likely be larger than on OOD Mapillary. Frame it as an in-distribution reference that complements the OOD cities (richmond/clovis), not a replacement — don't over-read an in-distribution win.

What it takes (harness is split-generic)

• Fetch the 1,000 gold-set images from projectsidewalk/rampnet-dataset (by pano id from manual_labels/) into a benchmark/manual_gold/ split (panos/ git-ignored like the cities).
• GT adapter: convert manual_labels YOLO boxes → point GT in the harness's derived-GT shape — gt_points = box centers, ignore_points = [] (no unsure), fn_confirmed = True (fully labeled ⇒ every pano recall-eligible). Same box→center reduction the harness already applies to VLMs; curb-ramp boxes are small, so center is a fine point proxy at radius 0.022.
• Run the full model set (rampnet, gemini, qwen, owlv2/gdino, molmo) on the split; add a results row to docs/model_comparison.md with P/R + AP + Wilson CIs. (RampNet + open detectors free; Gemini on 1,000 panos is more paid calls but within the pre-approved budget.)
• Sanity-check RampNet's harness numbers against the known gold-set figures (HF card 0.949/0.873, with TTA + one-to-one matching) — expect a small offset since the harness path is no-TTA + perspective-tiled.

Notes

• Multi-city mixed set (not a single "city").
• Possible pano overlap with the existing bend split (different sample; likely disjoint — confirm and note).

Refs #34, #20, #39, #45, #55.

Start with the split-generic model-comparison harness, manual_labels/*.txt, README.md, and docs/data_provenance.md. Fetch the 1,000-panorama set into benchmark/manual_gold/, add the YOLO-to-point GT adapter, run the listed model set, and update docs/model_comparison.md with metrics, confidence intervals, and the stated in-distribution caveats.