fix: 5 bug fixes on tabular executor

- top_n: remove fallback to sort_col (must use value_col only)
- coercion warning: log NaN introduced by numeric coercion in filters
- source_type filter: restrict tabular results to source_type=document
- raw fallback: cap columns to 20 to avoid oversized response
- add comment clarifying cross-file JOIN is out of scope for v1

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

src/query/executors/tabular.py

@@ -82,6 +82,9 @@ class TabularOperation(BaseModel):
 
 def _get_filter_mask(df: pd.DataFrame, col: str, value: str, operator: str) -> pd.Series:
     numeric = pd.to_numeric(df[col], errors="coerce")
+    coerced_nulls = numeric.isnull() & df[col].notna()
+    if coerced_nulls.any():
+        logger.warning("numeric coercion introduced NaN", col=col, count=int(coerced_nulls.sum()))
     if operator == "eq":
         return df[col].astype(str) == str(value)
     elif operator == "ne":
@@ -99,6 +102,9 @@ def _get_filter_mask(df: pd.DataFrame, col: str, value: str, operator: str) -> p
 
 def _apply_single_filter(df: pd.DataFrame, col: str, value: str, operator: str) -> pd.DataFrame:
     numeric = pd.to_numeric(df[col], errors="coerce")
+    coerced_nulls = numeric.isnull() & df[col].notna()
+    if coerced_nulls.any():
+        logger.warning("numeric coercion introduced NaN", col=col, count=int(coerced_nulls.sum()))
     if operator == "eq":
         return df[df[col].astype(str) == str(value)]
     elif operator == "ne":
@@ -162,7 +168,7 @@ def _apply_operation(df: pd.DataFrame, op: TabularOperation, limit: int) -> pd.D
             raise ValueError(f"filter requires filter_col/filter_value or filters or or_filters, got {op}")
         return result.head(limit)
     elif op.operation == "top_n":
-        col = op.value_col or op.sort_col
+        col = op.value_col
         if not col:
             raise ValueError(f"top_n requires value_col, got {op}")
         n = op.n or limit
@@ -206,7 +212,8 @@ class TabularExecutor(BaseExecutor):
     ) -> list[QueryResult]:
         tabular = [
             r for r in results
-            if r.metadata.get("data", {}).get("file_type") in _TABULAR_FILE_TYPES
+            if r.source_type == "document"
+            and r.metadata.get("data", {}).get("file_type") in _TABULAR_FILE_TYPES
         ]
 
         if not tabular:
@@ -270,6 +277,7 @@ class TabularExecutor(BaseExecutor):
                 )
                 return None
 
+        # Each group runs independently — cross-file JOIN is out of scope for v1.
         gathered = await asyncio.gather(*[
             _process_group(doc_id, sheet_name, info)
             for (doc_id, sheet_name), info in groups.items()
@@ -305,7 +313,7 @@ class TabularExecutor(BaseExecutor):
 
         # Fallback: return raw rows
         logger.warning("tabular agent failed after retries, returning raw rows")
-        return df.head(limit)
+        return df.head(limit)[df.columns[:20]]
 
 
 tabular_executor = TabularExecutor()