fix: D5 downsampling index mismatch + add 1M-row medium-scale test to docs

analysis/tool_registry.py

@@ -848,11 +848,14 @@ async def _handle_run_clustering(
             idx = rng.choice(len(data), 50000, replace=False)
             data = data[idx]
 
-        # ── D5: General downsampling for big data (>50K rows → 50K) ───────
+        # ── D5: Downsampling → extract matching indices ──────────────
+        # Downsample to 50K max, store indices so _handle_extract_features
+        # can filter the same rows.
+        _ds_idx = None
         if len(data) > 50000:
             rng = np.random.default_rng(42)
-            idx = rng.choice(len(data), 50000, replace=False)
-            data = data[idx]
+            _ds_idx = rng.choice(len(data), 50000, replace=False)
+            data = data[_ds_idx]
 
         # ── H17: Variance filtering ───────────────────────────────────────
         # Remove features with zero / near-zero variance (< 1e-10)
@@ -983,6 +986,7 @@ async def _handle_run_clustering(
             'params': p,
             'random_state': random_state,
             'feature_columns': feature_cols,
+            '_ds_idx': _ds_idx.tolist() if _ds_idx is not None else None,
         },
         parent_dataset_id=dataset_id,
     )
@@ -1092,9 +1096,23 @@ async def _handle_extract_features(
 
     try:
         lf: pl.LazyFrame = dataset_entry['lazyframe']
-        df = lf.collect(streaming=True)
         labels = cluster_entry['labels']
 
+        # D5: If clustering downsampled the data, filter matching rows
+        _ds_idx = cluster_entry.get('params', {}).get('_ds_idx')
+        if _ds_idx is not None:
+            df = lf.collect(streaming=True)
+            # Add row index, filter, then drop it
+            df = df.with_row_index('__ds_idx')
+            df = df.filter(pl.col('__ds_idx').is_in(_ds_idx))
+            df = df.drop('__ds_idx')
+        else:
+            df = lf.collect(streaming=True)
+
+        # Truncate labels if df was filtered but labels weren't
+        if len(labels) != len(df):
+            labels = labels[:len(df)]
+
         # Only use numeric columns for feature extraction
         numeric_cols = [c for c in df.columns
                         if df[c].dtype in (pl.Float32, pl.Float64,