Add Clovis validation split to the HF benchmark dataset

Add the Clovis split to the published HF benchmark

The Clovis validation split landed in gt/clovis-benchmark (branch, PR pending):
benchmark/clovis/ with 125 GT panos and verdicts.json. It scores P 0.889 /
R 0.650 on the unbiased subset (P 0.914 / R 0.713 across all 125) — a second
out-of-distribution Mapillary city alongside Richmond, and the weakest split so far.

scripts/build_benchmark_dataset.py currently packs only bend + richmond. To include
Clovis:

• Add "clovis": "mapillary" to CITY_IMAGERY.
• Training-overlap check: Clovis is Mapillary (unseen imagery source), so like
  Richmond it should have zero overlap with projectsidewalk/rampnet-dataset
  train/val. Confirm and document (no LEAKED_* set expected).
• Verify the reload-and-rescore step reproduces P 0.889 / R 0.650 for the split.
• Update the dataset card: Clovis is an out-of-distribution deployment city whose
  lower recall is explained by imagery — it is 100% GoPro Fusion (soft 2018
  consumer 360), vs Richmond's mostly-pro NCTECH iSTAR Pulsar + GoPro Max.

Opportunity: expose camera provenance as a dataset column

The records now carry full Mapillary provenance — camera_make / camera_model /
camera_type plus a verbatim source_metadata dump (added in the auto-labeler,
sidewalk-auto-labeler export-benchmark-verify). Since camera model is the field that
actually predicts image quality across cities (quality_score is compressed near the top
and resolution barely varies), consider adding camera_model to the dataset FEATURES
so the OOD story is analyzable directly from the published benchmark.

Related: #21 (HF benchmark), #26 (GT tool), and the Clovis split PR.

Start with scripts/build_benchmark_dataset.py, especially CITY_IMAGERY and the dataset FEATURES, then inspect the reload-and-rescore workflow. Add the Clovis split, verify its train/val overlap and reported precision/recall, and update the dataset card with its OOD and camera context; determine whether the camera_model opportunity is in scope.