Graphlet NaN at W > trained_wl

The symptom

For any graphlet / transformer_graphlet row with evaluation window greater than the training window, the output tables contained NaNs in the last W - trained_wl timestamps.

The root cause

merge_graphlet_dfs used to join the target station's test dataframe and the neighbours' predicted dataframes with how="outer". At W > trained_wl, the neighbours simply don't have predictions for the extra horizon — they were trained at a shorter window. An outer join fabricates NaN-filled rows for those dates, which then silently leaked into the sweep metrics.

The fix

merge_graphlet_dfs now uses how="inner". A graphlet method can only produce an honest prediction on days where every neighbour already has one. If the training window is 90 and the eval window is 180, the graphlet sweep produces 90 valid predictions and no NaNs.

Knock-on effects

• Fewer rows per station in sweep CSVs at W > 90. That's correct — don't scale metrics back up.
• Downstream notebooks aggregate over valid rows only.

See the merge_graphlet_dfs docstring in swissrivernetwork/benchmark/util.py for the inline rationale.