Fix guanlab_dengkw_pm

[rcannood]

Describe your changes

Three problems in guanlab_dengkw_pm, checked against the original submission.

1. The consensus scheme had been replaced. The original averages ten kernel ridge models -- five reshuffles of the batches into two halves, seeded with 1000:

np.random.seed(1000)
for _ in range(5):
    np.random.shuffle(batches)
    for batch in [batches[:batch_len//2], batches[batch_len//2:]]:
        ...
y_pred /= 10

The port did something else entirely: one fixed two-way split when either modality is ADT, leave-one-batch-out otherwise, with no reshuffling and no seed. So the result depended on the order the batches happened to arrive in, and for ADT pairs it averaged two models where the original averaged ten. Restored the original, with --n_repeats and --seed as arguments.

Effect on the test resources -- 23 of 32 metric/dataset combinations improve:

dataset	metric	old	new	delta
bmmc_multiome/swap	overall_pearson	0.3474	0.3648	+0.0175
bmmc_multiome/swap	mean_pearson_per_cell	0.3416	0.3590	+0.0173
bmmc_multiome/swap	rmse	1.0748	1.0555	-0.0194
bmmc_multiome/swap	mae	0.5507	0.5400	-0.0107
bmmc_cite/swap	(all 8)			all improve
bmmc_cite/normal	(6 of 8)			small
bmmc_multiome/normal	(7 of 8)			small regressions, mae -0.0100

bmmc_multiome/normal is the one that gets slightly worse, which makes sense -- that is the pair where the old code's leave-one-batch-out trained each model on more batches. But the point is fidelity to the published method rather than picking whichever variant scores best, and the differences here are muted anyway: these fixtures have few batches, while the real datasets have nine or more, where "half the batches" and "all but one batch" differ a lot more.

2. embedder_mod2 is defined conditionally but used unconditionally.

if n_mod2 is not None and n_mod2 < input_train_mod2.n_vars:
    embedder_mod2 = TruncatedSVD(n_components=n_mod2)
    train_gs = embedder_mod2.fit_transform(...)
else:
    train_gs = input_train_mod2.to_df(layer="normalized").values...   # no embedder_mod2
...
    y_pred += (krr.predict(test_norm) @ embedder_mod2.components_)    # NameError

Present in the original too. The else fires when the target has <= n_mod2 features (50 or 70), which no dataset we currently ship does -- ADT is 134, GEX ~14k, ATAC 10k -- so it has never been hit. And the re-mapping is wrong there anyway: in that branch the KRR already predicts in the target space. Now tracked with embedder_mod2 = None and the re-mapping skipped:

n_mod2=5,  n_vars=30, svd=yes -> y_pred (10, 30)
n_mod2=50, n_vars=30, svd=no  -> y_pred (10, 30)

3. n_comp_dict covers only the four NeurIPS pairs. Anything else is a bare KeyError, which reads as "the method is broken" rather than "the method doesn't apply here". Switched to exit_non_applicable() (exit 99), which also means adding src/utils/exit_codes.py, matching task_batch_integration.

Also dropped --distance_method and --n_pcs, which were declared but never read.

One thing deliberately not ported

The original clips ATAC predictions to [0, 1]:

if mod2_type == "ATAC":
    np.clip(y_pred, a_min=0, a_max=1, out=y_pred)

That assumed the starter kit's binarised ATAC. Here the ATAC target is log_cp10k, reaching 8.8 with 2.9% of values above 1, so the clip would cap most of the signal. Left out.

All four pairs run green:

bmmc_cite/normal      GEX->ADT   EXIT=0  5s  10 fits
bmmc_cite/swap        ADT->GEX   EXIT=0  6s  10 fits
bmmc_multiome/normal  GEX->ATAC  EXIT=0  7s  10 fits
bmmc_multiome/swap    ATAC->GEX  EXIT=0  7s  10 fits

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!