Add KL bar plots and sample_frac to load_predicted_local; ignore root parquet scratch files

Adds plot_kl_bars/plot_kl_bars_pl (numpy/polars variants) for ranking which
target dimension or pdg/material stratum drives KL regressions, with the
same kl*n group capping as plot_marginals. Switches existing histogram
plots to step-type/log-scale. Adds sample_frac to load_predicted_local for
subsampling large predict parquets.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

tests/test_analysis.py

@@ -20,6 +20,8 @@ from giant.analysis import (
     plot_constraint_violations,
     plot_correlation_matrices,
     plot_direction_alignment,
+    plot_kl_bars,
+    plot_kl_bars_pl,
     plot_marginals,
     plot_pairwise,
 )
@@ -155,6 +157,24 @@ def test_plot_constraint_violations_runs_without_error():
     assert fig is not None
 
 
+def test_plot_kl_bars_runs_without_error():
+    fig = plot_kl_bars(_make_collection())
+    assert fig is not None
+
+
+def test_plot_kl_bars_grouped_runs_without_error():
+    fig = plot_kl_bars(_make_collection(), group_by="material")
+    assert fig is not None
+
+
+def test_plot_kl_bars_caps_groups_by_pdg():
+    collection = _make_collection(n=600)
+    collection.pdg = np.arange(600) % 8  # 8 distinct pdg values, > max_groups
+    fig = plot_kl_bars(collection, group_by="pdg", max_groups=3)
+    ax = fig.axes[0]
+    assert len({line.get_label() for line in ax.containers}) <= 3
+
+
 def _write_predicted_local_parquet(path, n=50, metadata=None, rng=None):
     """Mimic `giant predict --coord local`'s output schema for the loader tests."""
     rng = rng or np.random.default_rng(0)
@@ -300,6 +320,13 @@ def test_marginal_table_pl_matches_numpy_version(tmp_path, group_by):
         )
 
 
+@pytest.mark.parametrize("group_by", [None, "pdg", "material", "energy"])
+def test_plot_kl_bars_pl_runs_without_error(tmp_path, group_by):
+    path = _predicted_local_path(tmp_path)
+    fig = plot_kl_bars_pl(path, group_by=group_by)
+    assert fig is not None
+
+
 def test_marginal_table_pl_rejects_missing_metadata(tmp_path):
     path = tmp_path / "no_metadata.parquet"
     _write_predicted_local_parquet(path, metadata=None)