Filter heatmap rows without data for selected metric

A variant/head that never reports the selected metric (e.g. ridge is
regression-only, so it has no auprc_test/mcc_test/etc.) previously left
an all-blank row once its only columns got filtered out. Now such rows
are dropped, and multi-model best-variant selection skips families with
no non-null metric values instead of crashing on idxmax over all-NaN.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

src/plots.py

@@ -343,20 +343,28 @@ def heatmap_variants(
             family_df = mdf[mdf["model_alias"] == m]
             if family_df.empty:
                 continue
-            # Pick the variant with highest mean metric for this family
-            vkey_means = family_df.groupby("_vkey")[metric].mean()
+            # Pick the variant with highest mean metric for this family. A variant/head
+            # (e.g. ridge, regression-only) that never reports this metric has an all-NaN
+            # group mean — idxmax would raise, and it'd render as an all-blank row anyway.
+            vkey_means = family_df.groupby("_vkey")[metric].mean().dropna()
+            if vkey_means.empty:
+                continue
             best_vkey = vkey_means.idxmax()
             best_df = family_df[family_df["_vkey"] == best_vkey]
             rows_iter.append((_variant_label(best_df.iloc[0]), best_df))
     else:
-        configs = sorted(mdf["_vkey"].unique())
+        configs = sorted(
+            c for c in mdf["_vkey"].unique() if mdf[mdf["_vkey"] == c][metric].notna().any()
+        )
         if len(configs) > 1:
             rows_iter = [
                 (_variant_label(mdf[mdf["_vkey"] == c].iloc[0]), mdf[mdf["_vkey"] == c])
                 for c in configs
             ]
+        elif len(configs) == 1:
+            rows_iter = [(_variant_label(mdf[mdf["_vkey"] == configs[0]].iloc[0]), mdf[mdf["_vkey"] == configs[0]])]
         else:
-            rows_iter = [(_variant_label(mdf.iloc[0]), mdf)]
+            rows_iter = []
 
     # A task/dataset only gets a column if the selected metric is ever reported for it —
     # e.g. bulk_rna_expression is regression-only (r2_test), so it never has mcc_test/
@@ -386,6 +394,9 @@ def heatmap_variants(
         def _cell_vals(vdf: pd.DataFrame, col_key: str) -> pd.Series:
             return vdf[vdf["task_name"] == col_key][metric].dropna()
 
+    if not rows_iter:
+        return go.Figure()
+
     col_names = col_names + ["Overall"]
 
     rows_z, rows_text, rows_hover, row_labels = [], [], [], []