Describe your changes

novel_train drops every all-zero feature from the training matrix:

zero_row = input_train_mod1.X.sum(axis=0) == 0
rem_var = None
if True in zero_row:
  rem_var = input_train_mod1[:, zero_row].var_names
  input_train_mod1 = input_train_mod1[:, ~zero_row]

but only hands the first one to novel_predict:

adata.uns["removed_vars"] = [rem_var[0]]

so novel_predict removes exactly one column from the test matrix regardless of how many the model dropped. With two or more all-zero features the test matrix is wider than the model expects -- the TF-IDF idf vector no longer broadcasts, or nn.Linear gets the wrong input size on the ADT path:

all-zero training features: ['g1', 'g3']
  old: [rem_var[0]]      -> test has 4 features, model expects 3 <-- MISMATCH
  new: list(rem_var)     -> test has 3 features, model expects 3 OK

The comment says "Mostly only applicable for testing", but nothing stops a real GEX training split from having several all-zero genes.

Part of a series of PRs coming out of a pre-run review of the benchmark.

Checklist before requesting a review

• I have performed a self-review of my code

• Check the correct box. Does this PR contain:

  • Breaking changes
  • New functionality
  • Major changes
  • Minor changes
  • Bug fixes

• Proposed changes are described in the CHANGELOG.md

• CI Tests succeed and look good!