Guard the silent failures found in the RF100-VL yolox campaigns

[EHxuban11]

Harness half of the 2026-08 campaign postmortem. Library half is LibreYOLO/libreyolo#710.

Every problem in those campaigns was silent. Crash handling is genuinely good (~300 dataset-runs, 0 failures, dataset+epoch resume, OOM to grad-accum). What was missing was anything that notices a wrong result.

What

• ETA lane count (rf100vl_dash.py). lanes = len(gpus_active) counted GPUs, not workers, so an 8-GPU box at --jobs-per-gpu 2 divided remaining work by 8 instead of 16 and reported roughly double. Now counts running datasets.
• -1 sentinel exclusion (_aggregate_metrics). pycocotools returns -1 for an area range with no ground truth. Averaging it in published a yolox-nano mAP_small of -0.0535, a negative average precision.
• Selection-vs-test divergence warning (new). Compares each dataset's valid_mAP50_95 from stats.json against its reported test AP, warns above 0.15 AP, records rf100vl.selection_test_divergence.

Why the third one matters most

yolox-nano selected on valid 0.5663 and published test 0.1620 on ball, while yolox-tiny agreed to 0.002 on the same datasets. Cause was BatchNorm eps reverting on a class-count rebuild (LibreYOLO#700): trained at 1e-5, evaluated at 1e-3, ruinous only for the depthwise variant. A full 100-dataset campaign published 0.3601 where the truth was 0.4853, and the harness said nothing.

A 16-point gap between adjacent sizes of one family is not a result, it is a bug report. Adjacent sizes land ~3 points apart.

Check

tests/test_campaign_guards.py (new, 7 tests): sentinel excluded from the mean, published negative mAP_small cannot recur, metric defined nowhere stays -1, the real yolox-nano numbers are flagged, the real yolox-tiny numbers are not, missing stats and absent weights-root are non-errors. All passing.

Not verified

• Warn-only by design. A genuinely hard dataset can drift, and a run that refuses to emit results is worse than one that emits them loudly caveated. The 0.15 threshold is a judgement call from two campaigns, not a tuned value.
• The ETA fix undercounts lanes once the queue drains and fewer datasets run than the pool size. Little work remains by then, but it is not exact.
• Not fixed here, and worth a follow-up: sync-artifacts reports uploaded N, skipped 0 while silently shipping nothing outside the paths it knows. It dropped every checkpoint and a provenance note from a corrected run because the weights root was flat rather than .runs-shaped. It also never uploads the recipe JSON, so a run whose recipe_sha256 points at a box-local file is unreproducible once that box is gone.

[EHxuban11]

Reviewed (Cursor agent, on Xuban's behalf). LGTM — merging.

• Lanes fix is correct: len(running) is the pool size while anything is pending, and the residual undercount once the queue drains is honestly documented and immaterial.
• Sentinel exclusion is right: >= 0.0 keeps genuine zero scores, exclusion is per-metric, and an all-sentinel metric stays -1 instead of fabricating a number. The published negative mAP_small cannot recur, and the test asserts exactly that.
• Divergence gate is the highest-value change: warn-only is the correct posture for a benchmark harness, the threshold is documented as a judgement call, the result is recorded as selection_test_divergence in the output JSON, and the tests pin the real campaign numbers (nano's 0.5663/0.1620 flagged, tiny's 0.6198/0.6091 quiet). Verified the helper's imports (load_json, Path, os, json, Any) all resolve in va_bench/rf100vl.py.

Agreed that sync-artifacts (skip-report vs explicit manifest) stays a follow-up — it needs a design call and shouldn't ride in a bug-fix PR. Given it nearly published an unreproducible correction today, it's the top remaining harness item.

CI pytest is green.