Fix yolov2 support and other bugs found in deep review

[IanButterworth]

Note

This PR was written by Claude (Claude Code), from a deep review of the repo requested and supervised by @IanButterworth.

This PR fixes every confirmed bug from a full-repo review, one fix per commit.

Headline: yolov2 models fixed and re-enabled

v2_COCO (skipped: "Not all weights are read during load") and v2_tiny_COCO (skipped: "results differ") were broken by three stacked bugs:

1. reorg layer was broken three ways — w // stride produced Rational dims which reshape rejects; the batch dim was dropped; and the constructor didn't multiply channels by stride², so every conv after the reorg read weights with the wrong input-channel count. For yolov2-COCO the resulting deficit is exactly 3*3*(1280-1088)*1024 floats = 7,077,888 bytes — precisely the gap in the old skip note ("Read 196856372 bytes. Filesize 203934260"). The weights were never bad. The new implementation replicates darknet's legacy [reorg] (reorg_cpu with forward=0, dispatched via REORG_OLD in AlexeyAB) as reshape/permutedims, was verified element-for-element against a literal transliteration of the C loop, and asserts input divisibility so non-conforming models get a clear error.
2. BatchNorm read order — weight files with darknet headers < 0.2 (both v2 models, header 0.1.0) took a branch that read scales before biases. Neither PJReddie nor AlexeyAB darknet ever wrote that order (load_convolutional_weights: biases, scales, means, vars in every version), so BN scales/biases were swapped, blowing up activations. The version header still governs the seen field width, which is the only real format difference.
3. Region-layer class scores need softmax when the cfg enables it — darknet's region layer (yolov2) applies logistic to objectness, and softmax over class scores when the cfg sets softmax=1 (as the bundled v2 cfgs do; with softmax=0, class scores are left linear, matching forward_region_layer). Sigmoid was applied to both unconditionally, skewing v2 confidences by up to ~0.15.

With all three fixed, both v2 models match Darknet.jl (AlexeyAB) exactly to 3 decimals on both test images (classes, boxes, confidences; verified locally on macOS). They're removed from the skip list, their artifacts download in tests again, and reference images/detection refs are regenerated (the old v2_tiny refs were produced under the BN bug).

Other fixes

• overridecfg! indexed cfgvec with indices computed from cfgchanges — only (:net, 1, ...) worked, by coincidence. The documented (:yolo, 1, :classes, n) use case errored. One-line fix + regression test.
• GPU keepdetections always dropped the last detection — col < cols (strict) in kern_genbools meant the final column's flag was never set.
• GPU/CPU numeric mismatches — kern_clipdetect used > where CPU keeps >=; kern_findmax! seeded its max at 0/index 0, reporting an invalid class index when all scores are ≤ 0.
• Class-score max scanned one row too far (CPU and GPU) — the scan included the first zero-filled scratch attribute appended after the classes, so all-negative class scores (possible with region softmax=0) selected a phantom class one past the last real one.
• Soft-NMS performed no suppression — decayed scores were never written back and the candidate window never advanced, so every box was kept at its original score and the already-kept box could be re-kept. Now decays are written into the results (nms is renamed nms! accordingly), survivors are selected allocation-free, and kept boxes whose decayed score falls below detect_thresh are pruned. Unit-tested, plus fuzz-verified against a reference implementation.
• Dead outnr write with latent BoundsError — weights[:, :, a+3, outnr, :] indexed the anchor dim with the output number (BoundsError for models with more heads than anchors), and the value was overwritten by findmax! anyway. Removed; output is byte-identical.
• prepare_image! matching-size 2D Float32 branch returned a bare 1-channel array instead of the (arr, padding) tuple, breaking the documented destructuring. Fixed + test.
• Dead broken maxpool branch removed — the @static if flux_maxpool else-branch called darknet_maxpool_layer, which doesn't exist anywhere; flipping the flag could never have worked (−77 lines).
• README — disable_bumper → disallow_bumper (the documented kwarg didn't exist), v2 no longer marked broken, typo fixes.
• CHANGELOG — added an Unreleased section covering the above.

All bundled models (v2, v2_tiny, v3, v3_tiny) were re-verified against the stored reference results (1e-5) after the follow-up commits.

Caveats

• The CUDA commits are review-verified but untested on GPU hardware (no NVIDIA GPU available locally); CI has no GPU runners either.
• The v2 tests run on CI for the first time here — worth watching the first run on Linux/Windows.

🤖 Generated with Claude Code

[IanButterworth]

Remove this and the direct dep before merge

[IanButterworth]

I want to keep iterating, so I'm going to merge and remove this before release