step=0 or negative in --val_steps_to_log / --train_steps_to_log / --var_leads_metrics_watch silently logs the wrong step's loss

[RudraDudhat2509]

step > max_steps is the only bound checked at CLI parse time for --train_steps_to_log, --val_steps_to_log, and the per-variable lead times in --var_leads_metrics_watch. Nothing checks step >= 1.

Default is [1, 2, 3, 5, 10], so 1-indexing is clearly the intended scheme (there's no "step 0" in a 1-indexed unroll).

The actual bug is in ForecasterModule._log_step_loss:

for step in steps_to_log:
    if step <= len(time_step_loss):
        log_dict[f"{phase}_loss_unroll{step}"] = time_step_loss[step - 1]

step=0 passes step <= len(time_step_loss) and then indexes time_step_loss[-1], Python's negative-index wraparound, silently returning the last step's loss under the ..._unroll0 key. step=-1 returns time_step_loss[-2], and so on. Same pattern in test_step's spatial-loss indexing (spatial_loss[:, [step - 1 for step in val_steps_to_log if step <= spatial_loss.shape[1]]]).

So --val_steps_to_log 0 (an easy typo if someone is used to 0-indexing) doesn't error and doesn't get skipped, it logs real data under a misleading metric name.

Repro

from unittest.mock import MagicMock, patch
import loguru
loguru.logger.catch = lambda f: f
from neural_lam.train_model import main

mock_args = MagicMock()
mock_args.eval = None
mock_args.load = None
mock_args.config_path = "dummy.yaml"
mock_args.val_steps_to_log = [0]
mock_args.train_steps_to_log = []
mock_args.var_leads_metrics_watch = "{}"
mock_args.ar_steps_eval = 10
mock_args.ar_steps_train = 10

with patch("neural_lam.train_model.ArgumentParser.parse_args", return_value=mock_args):
    with patch("neural_lam.train_model.load_config_and_datastore", return_value=(MagicMock(), MagicMock())):
        main()  # does not raise, no error, no warning

Compare to --val_steps_to_log -1, -2, etc., same silent pass-through.

I have a fix ready (adds the missing step >= 1 check at CLI parse time, in the same two loops that already check the upper bound) and will open a PR shortly.