Support source-evidence nonlinear dominance diagnostics

workspace/scripts/eval_nonlinear_dominance_selector.py

@@ -30,11 +30,15 @@ from scripts.eval_dominance_selector import (  # noqa: E402
     _rows,
 )
 from scripts.eval_learned_dominance_selector import (  # noqa: E402
+    FEATURE_SET_CHOICES,
     _candidate_dataset,
     _evaluate_predictions,
     _feature_names,
     _group_means,
+    _resolve_index_path,
     _simple_summary,
+    _source_evidence_map,
+    _uses_source_evidence,
 )
 
 
@@ -49,6 +53,15 @@ def main(argv: list[str] | None = None) -> int:
     parser.add_argument("--calibration-target-index", type=Path, required=True)
     parser.add_argument("--eval-input", type=Path, required=True)
     parser.add_argument("--eval-target-index", type=Path, required=True)
+    parser.add_argument(
+        "--source-index",
+        type=Path,
+        default=None,
+        help=(
+            "Train split chart index used for source-evidence features. "
+            "Defaults to --calibration-target-index."
+        ),
+    )
     parser.add_argument(
         "--checkpoint-template",
         default="runs/ctt_residual_full_seed{seed}/model.pt",
@@ -61,7 +74,7 @@ def main(argv: list[str] | None = None) -> int:
     parser.add_argument("--k", type=int, default=8)
     parser.add_argument(
         "--feature-set",
-        choices=("basic", "tangent", "context", "context_tangent"),
+        choices=FEATURE_SET_CHOICES,
         default="context_tangent",
     )
     parser.add_argument(
@@ -104,6 +117,10 @@ def main(argv: list[str] | None = None) -> int:
         args.eval_target_index,
         chart_feature_mode=chart_feature_mode,
     )
+    source_index_path = _resolve_index_path(args.source_index or args.calibration_target_index)
+    source_evidence, source_index = (
+        _source_evidence_map(source_index_path) if _uses_source_evidence(args.feature_set) else ({}, {})
+    )
     dataset_target = (
         "utility_margin" if args.target == "positive_margin" else args.target
     )
@@ -114,6 +131,7 @@ def main(argv: list[str] | None = None) -> int:
         k=args.k,
         feature_set=args.feature_set,
         target=dataset_target,
+        source_evidence=source_evidence,
     )
     eval_dataset = _candidate_dataset(
         eval_rows,
@@ -122,6 +140,7 @@ def main(argv: list[str] | None = None) -> int:
         k=args.k,
         feature_set=args.feature_set,
         target=dataset_target,
+        source_evidence=source_evidence,
     )
     fit_rows, select_rows = _split_rows(
         calibration_dataset,
@@ -182,12 +201,15 @@ def main(argv: list[str] | None = None) -> int:
         "chart_feature_mode": chart_feature_mode,
         "calibration_input": str(args.calibration_input),
         "eval_input": str(args.eval_input),
+        "source_index": str(source_index_path) if _uses_source_evidence(args.feature_set) else None,
         "data_hash": eval_index.get("content_hash"),
         "split_hash": eval_index.get("split_hash"),
         "calibration_target_content_hash": calibration_index.get("content_hash"),
         "calibration_target_split_hash": calibration_index.get("split_hash"),
         "eval_target_content_hash": eval_index.get("content_hash"),
         "eval_target_split_hash": eval_index.get("split_hash"),
+        "source_content_hash": source_index.get("content_hash"),
+        "source_split_hash": source_index.get("split_hash"),
         "num_calibration_rows": len(calibration_rows),
         "num_fit_rows": fit_dataset["num_rows"],
         "num_selection_rows": select_dataset["num_rows"],
@@ -508,10 +530,12 @@ def _write_provenance(out_dir: Path, args: argparse.Namespace) -> None:
     (out_dir / "git_hash.txt").write_text(_run(["git", "rev-parse", "HEAD"]) + "\n")
     hashes = {
         "calibration_input": _sha256(args.calibration_input),
-        "calibration_target_index": _sha256(args.calibration_target_index),
+        "calibration_target_index": _sha256(_resolve_index_path(args.calibration_target_index)),
         "eval_input": _sha256(args.eval_input),
-        "eval_target_index": _sha256(args.eval_target_index),
+        "eval_target_index": _sha256(_resolve_index_path(args.eval_target_index)),
     }
+    if getattr(args, "source_index", None) is not None:
+        hashes["source_index"] = _sha256(_resolve_index_path(args.source_index))
     (out_dir / "data_hash.txt").write_text(json.dumps(hashes, indent=2, sort_keys=True) + "\n")
     (out_dir / "split_hash.txt").write_text(
         json.dumps(
@@ -527,7 +551,7 @@ def _write_provenance(out_dir: Path, args: argparse.Namespace) -> None:
 
 
 def _index_hash(path: Path) -> dict[str, Any]:
-    payload = json.loads(path.read_text())
+    payload = json.loads(_resolve_index_path(path).read_text())
     return {
         "split": payload.get("split"),
         "content_hash": payload.get("content_hash"),