fix: default RandomRedistributeSegGPU's retain_stats to True

[Hendrik-code]

The defect

RandomRedistributeSegGPU normalises each sample to [0, 1], adds a perturbation of up to 2.0 in that space — twice the input's full dynamic range — and, when retain_stats=False, never maps the result back.

So the output does not depend on the input's scale at all:

input std   1.0 -> output range (0.52, 2.85) mean 2.25
input std   1.5 -> output range (0.52, 2.85) mean 2.25
input std 100.0 -> output range (0.52, 2.85) mean 2.25

On z-scored nnU-Net patches that is finite, silent and badly out of distribution. It is the same failure the no-foreground branch of this very method carries a comment about ("turns a z-scored air patch … into [0, 1] … a perfectly finite, badly out-of-distribution patch that no NaN guard can see") — fixed there, not here.

What I tried, and why this PR is narrower than you might expect

The obvious fix is to undo the working normalisation before returning: x = x * denom + img_min. I implemented it and measured it, and it is worse. The perturbation is defined in normalised units, so rescaling it by the input range amplifies it:

before the "fix":  output mean  2.2
after  the "fix":  output mean 18.4

Bounding the amplitude in input units instead would be a redesign of an augmentation inherited from totalspineseg, not a bug fix, and I am not going to smuggle one into a bug-hunt PR.

So this PR does the two things that are defensible without redesigning it:

1. retain_stats defaults to True — the default was the broken setting.
2. The behaviour is written down in the class docstring, including the failed approach, so the next person does not repeat it.

Behaviour change

None for any shipped config. All six that use this transform set retain_stats explicitly. Both paper configs already use True; transform_params_hybrid.json and transform_params_hybrid_TAGE.json ask for False and keep it — they are still affected, and that is now documented rather than hidden.

all_augmentations.json is regenerated for the new default.

Tests

unit_tests/test_redistribute_scale.py: the default preserves the input's mean and std, still changes the image, and tracks the input scale across a 100× range. A separate class pins the known limitation — with retain_stats=False the output is identical for inputs differing only in scale — so it cannot be forgotten or silently change.

Full suite unchanged against baseline; ruff check . clean.

🤖 Generated with Claude Code