[sumitsaraswat362]

Fixes #1782.

This PR addresses two related issues with RNTuple array conversion:

1. arrays(how="zip") failed because HasFields.arrays passed None as the expression context, causing Awkward.group to raise a TypeError when checking context["is_jagged"]. We now pass {"is_jagged": ...} in expression_context, checking for ListOffsetArray and ListArray while excluding strings and fixed-size arrays.
2. numpy.asarray(ntuple) raised ValueError: object __array__ method not producing an array because HasFields.__array__ returned a dictionary from arrays(library="np"). Following the TTree and TBranch design, __array__ is moved to RField so fields produce NumPy arrays directly, while RNTuple delegates to sequence iteration so numpy.asarray(ntuple) returns an array of fields (one row per field) matching TTree.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 83.95%. Comparing base (8503a78) to head (0c57f25).
⚠️ Report is 6 commits behind head on main.

Additional details and impacted files

Files with missing lines	Coverage Δ
src/uproot/behaviors/RNTuple.py	84.12% <100.00%> (+1.03%)	⬆️
src/uproot/models/RNTuple.py	93.92% <100.00%> (+0.02%)	⬆️

... and 25 files with indirect coverage changes

[ariostas]

🤖 AI text below 🤖

Thanks for this! The np.asarray part matches TTree. I checked that np.asarray(tree) and np.asarray(ntuple) both return (n_branches/fields, n_entries), and both raise the same "inhomogeneous shape" error when field shapes differ. Small nit on the description: "returns an array of rows" reads like one row per entry; "one row per field" would be clearer.

The main issue is in the how="zip" context: string and fixed-size array fields get treated as jagged (details and reproducers inline).

[ariostas]

🤖 AI text below 🤖

layout.is_list is also True for strings (a list of chars) and for fixed-size arrays (RegularArray). TTree treats neither as jagged (is_jagged there is only isinstance(interpretation, AsJagged)), so how="zip" restructures these fields in ways TTree doesn't.

The most common case is a single string field next to a scalar:

import awkward as ak, numpy as np, uproot

with uproot.recreate("t.root") as f:
    f["ntuple"] = {"x": np.arange(2), "dataset": ["abc", "de"]}
with uproot.open("t.root") as f:
    print(f["ntuple"].arrays(how="zip").type)
# 2 * {x: int64, jagged0: {dataset: string}}
# expected: 2 * {x: int64, dataset: string}

Other inputs (TTree output shown where uproot can write the same data to a TTree):

fields	RNTuple how="zip"	TTree how="zip"
name_a, name_b (strings)	{name: {a: string}, jagged1: {name_b: string}}	n/a
x, pos shape (n, 3)	{x, jagged0: 3 * {pos}}	{x, pos: 3 * float64}
x, arr_a, arr_b shape (n, 2)	{x, arr: 2 * {a, b}}	{x, arr_a: 2 * int64, arr_b: 2 * int64}
jet_pt=[[1.,2.],[3.]], jet_name=["ab","c"]	jet: var * {name, pt} (event string copied onto every jet)	n/a

The last row only happens when every string's length equals that entry's jet count, but then the string gets zipped into the jet collection.

A check on the layout type that excludes strings fixes all of the above. I tried it locally: vector<string> still zips into jet.tag correctly, and the tests in this PR still pass:

expression_context = []
for f in arrays.fields:
    layout = arrays[f].layout
    is_jagged = isinstance(
        layout, (ak.contents.ListOffsetArray, ak.contents.ListArray)
    ) and layout.parameter("__array__") not in ("string", "bytestring")
    expression_context.append((f, {"is_jagged": is_jagged}))

Could you add a test with a string field and a fixed-size array field as well?

[ariostas]

🤖 AI text below 🤖

This fallback has the same is_list problem as the RNTuple context (strings and fixed-size arrays count as jagged). Once HasFields.arrays passes a real context, nothing in the codebase calls group with None. I'd drop the fallback and keep context["is_jagged"], so a caller that builds a bad context fails loudly instead of zipping with a guess. The same applies to test_library_awkward_group_defensive_none_context, which locks this behaviour in.

[ariostas]

🤖 AI text below 🤖

Nit: the other test files start with the license line # BSD 3-Clause License; see https://github.com/scikit-hep/uproot/blob/main/LICENSE.

[sumitsaraswat362]

Good catch on the string and RegularArray list handling. I updated the layout check, dropped the fallback in library.py, added the license header, and added tests covering strings, fixed-size arrays, and vector of strings.