chore: declare matplotlib and apply three behavior-preserving nits

- pyproject.toml: add a figures extra with matplotlib (used by the two figure
  scripts) so the environment is reproducible from pyproject alone.
- partition.py: correct the partition_eval_split docstring (simulated per-client
  eval sets from the global split, not client-owned held-out data).
- protean_cnn.py: extract_embeddings now moves batches to the model device.
- posthoc_statistical_audit.py: refuse to write an empty table.

pyproject.toml

@@ -20,6 +20,10 @@ research = [
     "pandas>=2.0",
     "torch>=2.0",
 ]
+figures = [
+    # Used by scripts/make_final_figures.py and scripts/make_comparison_figures.py.
+    "matplotlib>=3.8",
+]
 dev = [
     "pytest>=7.0",
 ]

scripts/posthoc_statistical_audit.py

@@ -55,6 +55,8 @@ def write_csv(name: str, rows: list[dict]) -> None:
     # never touch any pre-existing analysis artifact.
     if not name.startswith("posthoc__"):
         raise SystemExit(f"refusing to write non-posthoc artifact: {name}")
+    if not rows:
+        raise SystemExit(f"refusing to write empty table: {name}")
     path = ANALYSIS / name
     with path.open("w", newline="") as fh:
         w = csv.DictWriter(fh, fieldnames=list(rows[0].keys()))

src/protean_research/data/partition.py

@@ -103,11 +103,13 @@ def partition_eval_split(
     seed: int,
     split_name: str,
 ):
-    """Partition a validation/test split into per-client local eval sets.
+    """Partition a validation/test split into per-client simulated eval sets.
 
-    The baseline evaluation is per client (each client scores its local held-out
-    data). This uses the same frozen Dirichlet algorithm with a split-specific
-    seed, so it is deterministic and paired across variants.
+    Clients do not own held-out data in this design; instead the *global*
+    validation/test split is divided into per-client eval sets with the same
+    frozen Dirichlet algorithm, using a split-specific seed. The partition is
+    therefore deterministic and identical across every aggregation variant of a
+    condition, which keeps per-client comparisons paired.
     """
     import numpy as np
 

src/protean_research/models/protean_cnn.py

@@ -105,11 +105,14 @@ def extract_embeddings(model, features, batch_size: int = 1024):
     import numpy as np
     import torch
 
+    device = next(model.parameters()).device
     model.eval()
     outputs: list[np.ndarray] = []
     with torch.no_grad():
         for start in range(0, len(features), batch_size):
-            batch = torch.as_tensor(features[start : start + batch_size], dtype=torch.float32)
+            batch = torch.as_tensor(
+                features[start : start + batch_size], dtype=torch.float32, device=device
+            )
             outputs.append(model.embedding(batch).cpu().numpy())
     if not outputs:
         return np.empty((0, model.cfg.dense_units), dtype=np.float32)