Add clovis as a third benchmark city; pursue source/geographic diversity

Motivation

The comparison rests on two cities — richmond (Mapillary 360) and bend (GSV). That's thin for a "RampNet wins" generalization claim, and it can't distinguish "RampNet is better everywhere" from "RampNet is tuned to these two US-suburban streetscapes." More cities tighten the Wilson CIs and stress-test both the lead and RampNet itself.

Immediate: finish clovis

benchmark/clovis/ is already staged — 125 Mapillary panos (Clovis, NM), records.jsonl + panos/ + gallery/ + index.csv — but has no verdicts.json yet, so it isn't scored and doesn't appear in the results. Remaining:

• GT-verify the sample with scripts/gt_gallery.py benchmark/clovis → verdicts.json (in flight; gt_gallery.py currently has uncommitted working changes).
• One cluster run of the model set → a third results row in docs/model_comparison.md.
• Add the split to the HF benchmark dataset (#21) and the README "Current splits" table.

Beyond clovis: real diversity

clovis is Mapillary again, so it doesn't add source diversity (still GSV-vs-Mapillary). A genuinely harder generalization test wants a different source / geography / camera: a dense urban core, a non-US city, or a different capture rig. The two-repo split (sidewalk-auto-labeler enumerates + detects, RampNet scores) makes adding a city cheap — a split just needs records.jsonl + verdicts.json + panos/.

Bonus

clovis can double as the dev split for the VLM prompt study (companion issue), keeping prompt tuning off the richmond/bend test set.

Refs #20, #21; docs/model_comparison.md Next increments #2.

Start with benchmark/clovis/ and run scripts/gt_gallery.py benchmark/clovis to produce the missing verdicts.json, checking the script's current working changes first. Then run one model cluster, add the Clovis row to docs/model_comparison.md, update the HF dataset and README Current splits table, and confirm the split appears in results.