upsert_all() writes NULL to omitted columns in mixed-field batches

[glaziermag]

upsert_all() overwrites omitted columns with NULL when records in the same batch contain different fields. This reproduces on current main (6bc1d33d583c54bd69fbdd2071117e2d38c354a1): the batch-wide column list fills each record's missing fields with None, and the generated ON CONFLICT ... DO UPDATE assigns those values to the existing rows.

from sqlite_utils import Database

db = Database(memory=True)
table = db.table("dogs")
table.insert_all([
    {"id": 1, "name": "Cleo", "age": 5},
    {"id": 2, "name": "Nixie", "age": 6},
], pk="id")
table.upsert_all([
    {"id": 1, "age": 7},
    {"id": 2, "name": "Nixie II"},
])
print(list(table.rows_where(order_by="id")))

Expected:

[{'id': 1, 'name': 'Cleo', 'age': 7},
 {'id': 2, 'name': 'Nixie II', 'age': 6}]

Actual:

[{'id': 1, 'name': None, 'age': 7},
 {'id': 2, 'name': 'Nixie II', 'age': None}]

The upsert documentation says columns omitted from an upsert dictionary remain unchanged, and describes upsert_all() as performing upserts.

The relevant paths are missing-field expansion and the conflict-update assignments. The update columns need to reflect the fields supplied by each record while preserving record order.

Verified with Python 3.13.12 / SQLite 3.53.0. The existing tests/test_upsert.py suite passes (18 tests).

Found and reproduced using OpenAI Codex.