Background

LegacyObjectData is the fallback storage arm for columns that marrow can't represent. It is legitimately needed for object_ / datetime64[ns] / timedelta64[ns] columns (tracked separately in #622).

However it is also used for string columns that have nulls, with the comment:

"Marrow cannot build a string + null array"

This comment is no longer accurate. Marrow's StringBuilder has had append_null() support from the start, and the README benchmarks nullable string arrays explicitly.

Current demotion path

Column.set_null_mask (line 5277–5283) demotes a string AnyArray column to LegacyObjectData when nulls are applied, wrapping each String in a PythonObject:

if self.is_string():
 var src = self._str_list()
 var py_data = List[PythonObject](capacity=len(src))
 for i in range(len(src)):
 py_data.append(PythonObject(src[i]))
 self._storage = ColumnStorage(LegacyObjectData(py_data^, mask^))
 return

Fix

Replace the demotion with a proper nullable StringArray build via StringBuilder:

if self.is_string():
 var src = self._str_list()
 var b = StringBuilder(len(src))
 for i in range(len(src)):
 if mask.is_null(i):
 b.append_null()
 else:
 b.append(src[i])
 self._storage = ColumnStorage(AnyArray(b.finish()))
 return

Follow-on cleanup

Once no string columns can land in LegacyObjectData, remove the string-with-nulls special-case branches in:

• _str_list() (the LegacyObjectData string-extraction path)
• has_nulls() / null_mask_copy() (comments about string-with-nulls)
• Any other guards that branch on is_string() and _storage.isa[LegacyObjectData]()

LegacyObjectData itself is retained for object/datetime/timedelta columns.

See also #622 for the longer-term migration of datetime/timedelta away from LegacyObjectData.