fix(ml): parse datetime column and fix TA-Lib pattern names

- Add parse_dates parameter when loading enriched CSV
- Strip timezone from annotation timestamps to match data
- Fix pattern names: CDLTHREEWHITESOLDIERS -> CDL3WHITESOLDIERS
- Fix pattern names: CDLTHREEBLACKCROWS -> CDL3BLACKCROWS

services/ml/app/annotation_ingestion.py

@@ -188,8 +188,8 @@ class AnnotationIngestion:
         
         for ann in annotations:
             label = ann['label']
-            start_time = pd.Timestamp(ann['start_time'])
+            start_time = pd.Timestamp(ann['start_time']).tz_localize(None)
-            end_time = pd.Timestamp(ann['end_time'])
+            end_time = pd.Timestamp(ann['end_time']).tz_localize(None)
             
             # Find candles in span
             span_mask = (df['time'] >= start_time) & (df['time'] <= end_time)
@@ -504,7 +504,7 @@ def run_annotation_ingestion(
     logger.info("Running annotation ingestion stage")
     
     # Load enriched data
-    enriched_df = pd.read_csv(enriched_path)
+    enriched_df = pd.read_csv(enriched_path, parse_dates=['time'])
     logger.info(f"Loaded enriched data: {enriched_df.shape}")
     
     # Process annotations

services/ml/config/pipeline.yaml

@@ -86,8 +86,8 @@ stages:
         - "CDLHARAMI"
         - "CDLPIERCING"
         - "CDLDARKCLOUDCOVER"
-        - "CDLTHREEWHITESOLDIERS"
+        - "CDL3WHITESOLDIERS"
-        - "CDLTHREEBLACKCROWS"
+        - "CDL3BLACKCROWS"
     
     # Label merge strategy: "human_priority", "programmatic_priority", "both"
     merge_strategy: "human_priority"