Fix novel silently saving no model on multiome/swap

[rcannood]

novel scored 12/16 in the run_2026-08-01_11-48-20 benchmark. The four missing scores are both multiome/swap datasets (2021 + 2022), where novel_predict died on:

FileNotFoundError: '_viash_par/input_model_1/log_cp10k.novel.novel_train.output/tensor.pt'

What actually happened

novel_train exited 0 and published a model directory with no model in it.

The chain, straight from the run logs:

• An empty ATAC cell produces a constant lsi row, so the per-cell z-score in lsiTransformer.transform() divides 0 by 0 -- helper_functions.py:78: RuntimeWarning: invalid value encountered in divide.

• That nan reaches rmse(), and if best_score > rmse(...) is always False for nan, so torch.save() never runs.

• train_and_valid() then prints its untouched sentinel and returns normally:

  bmmc_multiome/swap    best rmse:  100000     <- never saved
  pbmc_multiome/swap    best rmse:  100000     <- never saved
  bmmc_multiome/normal  best rmse:  0.47737852 <- fine
  

So the train step looked green, and the failure only surfaced one step later in predict.

Changes

• Guard the zero-variance divide in lsiTransformer.transform().
• Track the best score as None/nan-aware rather than a 100000 sentinel, and raise if it never became finite -- a training step that saves no model should fail, not exit 0.

Verification

• viash test src/methods/novel/novel_train/config.vsh.yaml: 1/1 passed

• viash test src/methods/novel/novel_predict/config.vsh.yaml: 1/1 passed

• The fixture is bmmc_cite (ADT), so it does not exercise the lsi path that failed. Checked that separately against the real lsiTransformer with an all-zero cell present:

  LSI output (60, 20): 0 NaN entries
  empty cell row all-finite: True
  

The helper_functions.py:33 "divide by zero" warning also present in the run is left alone: empty rows have no stored values, so the inf never multiplies anything and no nan reaches the output (0 NaN above, with that warning still firing).