Widen DP lookback window and fix null-typed parquet columns

Increase LOOKBACK_WORDS from 15 to 20 for better alignment coverage
on segments that start further back from the expected position. Cast
null-typed Arrow columns to string in usage logger so all parquet
shards share a consistent schema, preventing HF dataset viewer errors.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

config.py

@@ -128,7 +128,7 @@ COST_INSERTION = 1.0                # Insert phoneme from reference (R)
 COST_DELETION = 0.8                 # Delete phoneme from ASR (P)
 
 # Alignment thresholds (normalized edit distance: 0 = identical, 1 = completely different)
-LOOKBACK_WORDS = 15                 # Window words to look back from pointer for starting positions
+LOOKBACK_WORDS = 20                 # Window words to look back from pointer for starting positions
 LOOKAHEAD_WORDS = 10                # Window words to look ahead after expected end position
 MAX_EDIT_DISTANCE = 0.25            # Max normalized edit distance for valid ayah match
 MAX_SPECIAL_EDIT_DISTANCE = 0.35    # Max normalized edit distance for Basmala/Isti'adha detection

src/core/usage_logger.py

@@ -165,6 +165,16 @@ if _HAS_DEPS:
                         row[feature] = None
 
             table = pa.Table.from_pylist(rows)
+
+            # Cast null-typed columns to string so all parquet shards share
+            # the same Arrow schema (prevents HF viewer concat errors).
+            for i, field in enumerate(table.schema):
+                if pa.types.is_null(field.type):
+                    table = table.set_column(
+                        i, field.name,
+                        pa.array([None] * len(table), type=pa.string()),
+                    )
+
             table = table.replace_schema_metadata(
                 {"huggingface": json.dumps({"info": {"features": schema}})}
             )