Add YOLOE (real-time open-vocab) to the zero-shot detector bucket

Motivation

The open-vocab bucket (#39 / PR #40) currently holds OWLv2 and Grounding DINO, both 2024-era. YOLOE ("Real-Time Seeing Anything," Tsinghua, Mar 2025) is the current SOTA real-time open-vocab detector — +3.5 AP over YOLO-Worldv2 on LVIS at ~1.4× speed, with text / visual / prompt-free modes. Adding it (a) modernizes the bucket, (b) adds a real-time speed axis against the slow VLMs, (c) tests whether a newer open-vocab model is a more efficient recall complement than OWLv2.

Expectation-setting (from PR #40)

The "purpose-built open-vocab beats chat VLMs" hypothesis was refuted — OWLv2 / gdino did worse (best-sweep F1 owlv2 0.184 vs gemini-3.6 0.634 vs rampnet 0.855); text-prompted open-vocab isn't selective enough for curb ramps. So YOLOE will most likely lose on F1 too. Its real value is currency + the recall-oracle angle: OWLv2 recovered 69/72 of RampNet's misses (union recall 0.990) but at 36–128 FP per recovered ramp — a 6–20× less efficient complement than Gemini (#35). Open question: does YOLOE's visual-prompt / prompt-free mode complement RampNet more efficiently?

What to run

Add a yoloe provider to compare.py, text prompt "curb ramp" / "wheelchair ramp" (plus try visual-prompt mode with a few ramp exemplars). Feed via the existing reprojection rig; boxes → center points → matcher at 0.022. Calibrated scores → AP / PR / --sweep like the other open detectors. Low effort — Ultralytics-packaged, no new eval path.

Caveats

Honor the nadir/hood masking follow-up (#47 — a third of each view is vehicle hood), and the parser/rig cache-key gap (bust .model_cache on any parser change).

Relates to #39 (open detectors) and #20 (harness). Companion supervised baseline: #51.

Start in compare.py and trace the existing open-vocabulary providers through the reprojection rig, box-to-center conversion, matcher threshold 0.022, and calibrated AP/PR and --sweep evaluation. Add the YOLOE provider with the specified text prompts and investigate visual-prompt mode using a few ramp exemplars. Done means YOLOE runs through the existing path, produces comparable metrics, and respects nadir/hood masking and cache invalidation after parser changes.