RNTuple: top-level tuple fields do not round-trip as tuples

[ariostas]

🤖 AI text below 🤖

Split out of #1713, which now tracks only the optional-record bug.

A top-level tuple field written by uproot does not read back as a tuple. Reproduces on current main (a5f44fd) with plain uproot.recreate / uproot.open.

import awkward as ak, uproot

arr = ak.Array([(1.0, 2), (3.0, 4)])          # 2 * (float64, int64)
with uproot.recreate("tup.root") as f:
    f["t"] = arr

with uproot.open("tup.root") as f:
    form, _ = f["t"].to_akform()
    print(form.is_tuple)                       # False, expected True

The tuple comes back as a record whose field names are the strings '0' and '1'.

Why

Tuple recovery exists and works — the reader restores tuple-ness when every on-disk member name matches _[0-9]+:

https://github.com/scikit-hep/uproot5/blob/main/src/uproot/models/RNTuple.py#L693

if all(re.fullmatch(r"_[0-9]+", name) is not None for name in record_names):
    namelist = None      # namelist=None -> RecordForm is a tuple

(and the same idea at models/RNTuple.py#L635 for the collection role).

The writer applies that naming convention for nested records but not at the top level:

# _cascadentuple.py#L355 — nested
subfield_name = f"_{i}" if akform.is_tuple else akform.fields[i]

# _cascadentuple.py#L439 — top level (generate_field_col_records)
for field_name, topakform in zip(akform.fields, akform.contents, strict=True):

akform.fields on a tuple is ['0', '1'] — no underscore — so the reader's heuristic misses. Measured:

top-level  ak.Array([(1.0, 2), (3.0, 4)])
    on-disk names: ['0', '1']              is_tuple: written True -> read False
nested     ak.Array({"pair": [(1.0, 2), (3.0, 4)]})
    on-disk names: ['pair', '_0', '_1']    is_tuple: written True -> read True

Nested tuples round-trip correctly; only the top level is affected.

Possible direction

Apply the same f"_{i}" if akform.is_tuple rule in generate_field_col_records. Note this changes on-disk field names for top-level tuples, so it is a format-visible change rather than a pure bug fix.

Note

This surfaced while reviewing #1687, whose extend() form check is the first code to compare a form reloaded from disk against the original. That check refuses this shape cleanly before writing anything, so #1687 is not affected — the mismatch is pre-existing and independent of it.

Environment

• uproot 5.7.6.dev29+g425d57430 (main @ a5f44fd)
• awkward 2.11.0