Add the supervised YOLO baseline to the comparison harness (#51)

[jonfroehlich]

The issue #51 architecture-vs-data ablation: does a generic detector trained on
the RampNet dataset also beat the zero-shot field, or is RampNet's keypoint
architecture doing the work?

This is the code half only. The training record it produced is #77, opened
concurrently by another session — the two are disjoint after a rewrite of this
branch (see "Overlap with #77" below).

What's here

• YoloDetector + yolo_results_to_boxes in detectors.py, plus the yolo
  branch in build_detector. Identity is the weights file stem, so the
  detection cache key is machine-independent; signature() hashes the weights
  bytes, so a retrain busts the cache. It emits scored boxes, so YOLO gets AP,
  PR curves and --sweep like the other open detectors.
• compare.py: --yolo-model, --yolo-conf (0.05 floor), --yolo-iou,
  --yolo-imgsz. requirements-vlm.txt: ultralytics.
• prepare_yolo_dataset.py — tiles+pano dataset builder with pluggable box-size
  strategies and deterministic background thinning.
• run_yolo_{data,prep,train}.slurm and hyak_yolo_runbook.sh (repo root,
  beside hyak_qwen_runbook.sh).
• +10 tests (187 pass, up from main's 177).

Box size is a train-only knob: evaluation reduces predictions to centre
points and matches at radius 0.022, so width and height are never scored.

Overlap with #77 — one thing to fix at merge time

This branch originally also carried the training record; I dropped it once #77
appeared, so the two PRs now touch disjoint files. One conflict survives and
is worth catching whichever merges second:

#77 adds scripts/model_comparison/yolo_baseline/run_yolo_train.slurm and
.../yolo_baseline/hyak_yolo_runbook.sh. This PR adds the canonical
scripts/model_comparison/run_yolo_train.slurm and the repo-root
hyak_yolo_runbook.sh. Git will not flag it — different paths — but merging
both leaves two copies of each script, and:

• the two run_yolo_train.slurm files are byte-identical;
• the two runbooks are not: Preserve the supervised-YOLO baseline training record (#51) #77's copy is three lines older, predating the
  ARG_MAX fix to the dataset count display (ls "$DATA/$s"/*.jpg | wc -l
  overflowed on the ~150k-file train split and printed a false train: 0).

Suggested resolution: keep the canonical copies from this PR, drop the two under
yolo_baseline/, and have that README point at them. The provenance those copies
existed to preserve is already in git history. The difference is display-only and
cannot affect the recorded runs, but shipping the older text as a second live copy
would be a real trap.

Status

Training is still running; nothing here is a result. The benchmark-city
evaluation against RampNet has not been run — see #77 for what the runs so far
do and do not support.

🤖 Generated with Claude Code

[jonfroehlich]

Review fixes pushed in ebbc27b:

• Runbook train: ACCOUNT now defaults to ckpt-makelab (the .slurm pins -p ckpt-g2 -q ckpt-gpu, which the old gpu-l40s-makelab default cannot submit to), and sub() now pins per-config BATCH (6/12/6/4/2/4) — unpinned, runbook-launched jobs ran batch=-1 (auto), defeating the identical-LR-schedule rationale the .slurm header states.
• Runbook: BOX_SIZE default is now pitch + RAMP_SIZE_M=1.8, matching run_yolo_prep.slurm's as-run defaults, and the multi-hour data/prep stages now point at their compute-node sbatch wrappers (login nodes reap heavy processes).
• prepare_yolo_dataset.py now stamps the full prep config into data.yaml as comments, with a regression test (188 pass). Box size is a train-only knob that no downstream args.yaml records — exactly the gap that made the Preserve the supervised-YOLO baseline training record (#51) #77 provenance question unanswerable from disk.
• Minor: overlay views render once instead of twice; --overlay-n help says PER SPLIT; hashlib import hoisted to module top.

The "Overlap with #77" merge-time task is resolved ahead of time: #77 dropped its two script copies and its README points at the canonical paths here, so no dedup is needed whichever merges second.

🤖 Generated with Claude Code