test_dunedin_pace_normalization fails on pandas 3.0 — stale pace_normalized.pkl fixture orientation (normalization output is correct) #209

Summary

On pandas 3.0, biolearn/test/test_model.py::test_dunedin_pace_normalization fails with KeyError: 'cg00000029'. I traced it, and the good news is the DunedinPACE normalization output is correct — the failure is a stale .pkl reference fixture whose axis orientation predates the current loader, plus a diagnostic loop that assumes label-based lookups.

This is separate from and downstream of the read-only-array bug in #195 / PR #206. On master this same test aborts earlier with the #195 ValueError: assignment destination is read-only; with that fix applied it advances past normalization and reveals this KeyError.

CI doesn't currently catch it because CI runs Python 3.10 and resolves an older pandas (pyproject.toml pins pandas>=2.1.0, no upper bound), so pandas 3.0 is never installed there.

Reproduce

# Python 3.11.13, pandas 3.0.3, numpy 2.4.6
pytest biolearn/test/test_model.py::test_dunedin_pace_normalization

biolearn/test/test_model.py:125: in test_dunedin_pace_normalization
    print(f"Location: {value[0]}, Actual: {actual.at[value[0]]}, Expected: {expected.at[value[0]]}")
E   KeyError: 'cg00000029'

(On unpatched master the same command instead fails earlier with ValueError: assignment destination is read-only at dunedin_pace.py:61 — the #195 bug. This report assumes that fix is applied so normalization runs to completion.)

Root cause

The test compares the live normalization output against the reference fixture testset/pace_normalized.pkl. The two are in different orientations:

The fixture was evidently saved before the loader promoted the probe-ID column (ID_REF) to the DataFrame index; the current loader returns CpG-indexed data.

Because the axes don't align:

1. np.abs(actual - expected) > 1e-6 flags every cell as a mismatch (200000/200000).
2. The diagnostic print loop then does expected.at[value[0]] where value[0] is a (CpG, sample) tuple, e.g. ('cg00000029', 'GSM1009660'). Looking up the CpG label in expected's integer RangeIndex raises KeyError.

The normalization math is not broken

Realigning the fixture (promoting its ID_REF column to the index) and comparing values directly:

max abs diff (realigned): 5.55e-16
values > 1e-6: 0 of 200000
NaNs: none

So the live output reproduces the reference to floating-point precision. The problem is purely the fixture's stored orientation plus the test's comparison logic — not the DunedinPACE computation.

Suggested fix (happy to open a PR)

Two independent things:

1. Fixture orientation — regenerate testset/pace_normalized.pkl in the current loader orientation (CpG-indexed, no ID_REF column), or have the test normalize the reference's orientation on load (e.g. set_index('ID_REF') when that column is present) so it stays robust to how the fixture was saved.
2. Diagnostic loop — the mismatch-reporting block uses DataFrame.at[...] with (row, col) tuples, which is fragile and, here, masks the real comparison result. Comparing on aligned axes avoids the KeyError and gives a meaningful diff if values ever do drift.

Happy to put up a small PR — let me know if you'd prefer the fixture regenerated vs. the test normalized on load.

Environment

• Python 3.11.13, pandas 3.0.3, numpy 2.4.6
• Reproduced with PR fix: handle read-only arrays in DunedinPACE quantile normalization (#195) #206 applied (so normalization runs); root cause and value equivalence verified as above.