toon: 5 correctness tests fail against the current toon-format reference (CI installs it unpinned from git) #183

What

Five toon correctness tests fail in CI on Python 3.11, 3.12 and 3.13. The failures do not involve any repo change: the CI installs the reference library from an unpinned git ref, and that reference has moved.

FAILED toon/test_toon_correctness.py::TestEncodeArrays::test_empty_array
FAILED toon/test_toon_correctness.py::TestEncodeOptions::test_tab_delimiter
FAILED toon/test_toon_correctness.py::TestEncodeOptions::test_pipe_delimiter
FAILED toon/test_toon_correctness.py::TestComplexStructures::test_mixed_nested_arrays
FAILED toon/test_toon_correctness.py::TestCrossImplementation::test_decode_ref_encoded[empty-array]
5 failed, 7523 passed, 34 skipped, 7 xfailed in 55.45s

Why this is environment drift, not a regression

toon/ is byte-identical to master in the run where this appeared (git diff --stat 46a0426 <head> -- toon/ is empty). The same five tests fail locally with only the reference library installed the way CI installs it:

pip install "toon_format @ git+https://github.com/toon-format/toon-python.git"   # -> toon_format 1.1.0
pytest toon/test_toon_correctness.py -q
# 5 failed, 110 passed

The two shapes of failure both point at the reference, not at our module:

• TypeError: dumps() takes 1 positional argument but 2 were given — the tests call _tf.encode(obj, {"delimiter": "\t"}), and the current reference exposes a one-argument dumps.
• Encoding mismatches, e.g. empty array: ours '[0]:' vs theirs '[]', and a nested array: ours '[2]:\n - name: A\n...' vs theirs '[2]{name,data{x}}:\n A,1\n B,2' — the reference changed its tabular encoding.

pyproject.toml declares bench-toon = ["toon-format>=0.9.0b1"], and .github/workflows/lint-test.yml deliberately excludes bench-toon from the extras install and pulls it from git main instead, so the reference floats with upstream.

Impact

Master's last green CI run was 2026-09-28 (36380174333). Every PR opened since is red for this reason, including #182 and #171. The 3.10 job passes only because most reference extras do not resolve there: it collects 2638 items against 7529 on 3.11, so the toon comparison tests never run.

Options

Both are a call on which side is authoritative, so this is left open:

1. Pin the reference in the CI step (and/or the bench-toon extra) to the version the module was written against, and raise it deliberately when the module is updated.
2. Update toon/toon.py and toon/test_toon_correctness.py to the current reference API and encoding, which looks like a real spec change rather than a rename.

Environment: run 38358964727, Python 3.11.17, toon_format 1.1.0 from git main, 2026-10-13.