feat: implement reproducible PROTEAN baseline pipeline

configs/datasets/datasets.yaml

@@ -13,6 +13,13 @@
 #   All preprocessed caches, partitions, manifests, checkpoints and analysis
 #   outputs must be written under `generated_root` (default: artifacts/), never
 #   back into the raw dataset directory.
+#
+# PREPROCESSING CONTRACT (frozen 2026-09-24, see docs/research/EXPERIMENT_PROTOCOL.md)
+#   * `categorical_columns` are one-hot encoded with categories fitted on the
+#     TRAIN split only; an explicit "__MISSING__" category is added.
+#   * every other feature column is coerced to numeric (missing tokens -> NaN),
+#     median-imputed and standardized using TRAIN statistics only.
+#   * `non_feature_columns`, the label column and its aliases are never features.
 
 generated_root: artifacts
 
@@ -30,7 +37,13 @@ datasets:
     drop_labels:
       - crypto-ransomware
     # '-' is the dataset's documented missing placeholder, not an empty string.
-    missing_tokens: ["-", "?"]
+    missing_tokens: ["-", "?", "#DIV/0!", "aza"]
+    drop_duplicate_rows: true
+    # Feature columns that must be treated as categorical (one-hot).
+    categorical_columns:
+      - Protocol
+      - Service
+      - anomaly_alert
     # Columns that are identifiers / timestamps / network endpoints and must NOT
     # be used as model features without an explicit, documented decision.
     non_feature_columns:
@@ -42,7 +55,7 @@ datasets:
       - Des_port
     notes: >
       Row-level missing placeholders ('-') and a small number of malformed tokens
-      (e.g. '?', 'aza') exist. Exact duplicate rows are present. See
+      (e.g. '?', 'aza', '#DIV/0!') exist. Exact duplicate rows are present. See
       docs/research/DATASET_INSPECTION.md.
 
   5gnidd:
@@ -57,9 +70,16 @@ datasets:
     benign_label: Benign
     drop_labels: []
     missing_tokens: [""]
-    # First column is an unnamed row index. Seq / Offset are flow-ordering fields.
+    drop_duplicate_rows: true
+    categorical_columns:
+      - Proto
+      - sDSb
+      - dDSb
+      - Cause
+      - State
+    # The first column is an unnamed row index. Seq / Offset are flow-ordering fields.
     non_feature_columns:
-      - ""
+      - "Unnamed: 0"
       - Seq
       - Offset
     notes: >