WeatherDataset.__len__ undercounts samples by num_future_forcing_steps when a datastore has no forcing

[AshNicolus]

Description

WeatherDataset.__len__ (non-forecast/analysis branch, neural_lam/weather_dataset.py) computes:

window = (
    max(2, self.num_past_forcing_steps)
    + self.ar_steps
    + self.num_future_forcing_steps
)
n_state_samples = len(self.da_state.time) - window + 1
if self.da_forcing is not None:
    n_forcing_samples = len(self.da_forcing.time) - window + 1
    base_len = max(0, min(n_state_samples, n_forcing_samples))
else:
    base_len = max(0, n_state_samples)

window unconditionally includes num_future_forcing_steps, even when computing n_state_samples in the self.da_forcing is None branch. But _slice_state_time never uses num_future_forcing_steps at all - verified against its actual index math (end_idx = idx + max(2, num_past_forcing_steps) + n_steps, no future-forcing term). Only _slice_forcing_time needs the extra num_future_forcing_steps steps.

This is inconsistent with the is_forecast branch just above it, which correctly computes required_state_steps = max(2, num_past_forcing_steps) + ar_steps (no future forcing) and only adds num_future_forcing_steps inside the if self.da_forcing is not None: block for the forcing-specific check.

get_dataarray(category="forcing", ...) is documented as optional (returns None when a datastore provides no forcing data), so da_forcing is None is a real, reachable, legitimate configuration - not hypothetical.

Failure scenario

For a no-forcing datastore with exactly T = max(2, num_past_forcing_steps) + ar_steps state time steps (just enough for one valid sample), __len__() incorrectly returns 0 instead of 1, and WeatherDataset.__init__ raises ValueError: ... too few time steps even though a sample is genuinely constructible. More generally, any no-forcing dataset undercounts available samples by up to num_future_forcing_steps, silently dropping valid samples at the end of the series.

Not covered by existing tests - all DummyDatastore/example configs provide forcing data.

Fix

Compute a separate state-only window (without num_future_forcing_steps) for n_state_samples, only adding num_future_forcing_steps for the forcing-window calculation, mirroring the pattern already used correctly in the is_forecast branch.