PEtab v1->v2 converter emits log10 over log-normal under petab 0.9.0, so every converted log10 observable is refused at import

[wshlavacek]

PyBNF 1.8.1, petab 0.9.0 (released 2026-09-07), Python >= 3.12

Summary

The scale-preserving PEtab v1 -> v2 converter (pybnf.petab.convert.petab1to2_preserve_scale) produces a v2 problem the importer then refuses, for every v1 observable with observableTransformation = log10:

pybnf.printing.PybnfError: Observable 'obs_V': observableTransformation 'log10' contradicts the scale of
noiseDistribution 'log-normal'. Give the residual scale in one place -- a log observableTransformation over
a linear noiseDistribution (normal / laplace).

tests/test_petab_convert.py::TestFullConversion::test_converted_log10_problem_imports_as_lognormal fails on an unmodified main (212b11a) under petab 0.9.0. CI shows it on the py3.12 / 3.13 / 3.14 legs of every PR since 2026-09-07 (first seen on #678). The py3.11 leg stays green only because petab 0.9.0 requires Python >= 3.12, so that leg resolves 0.8.2.

Root cause

This is PyBNF's problem, not petab's.

petab 0.9.0 fixed a missing return in petab1to2's noise-distribution merge (PEtab-dev/libpetab-python#502). Before the fix the converted noiseDistribution column was always blank. After it the converter does what it was designed to do: fold the v1 observableTransformation into the v2 noiseDistribution. For log10 + normal that means log-normal, with a warning that log10-normal does not exist in PEtab v2 and the natural-log family is being substituted.

PyBNF's wrapper then re-injects observableTransformation = log10 as a preserved extra column (issue #499, ADR-0073), precisely to avoid that lossy substitution. The importer's _native_noise_family sees a log10 transformation over a natural-log distribution and refuses the contradiction. Under petab 0.8.2 that never fired because the blank cell defaulted to normal. The converter was silently relying on a petab bug.

The natural-log log case survives because both sides agree on ln. Linear observables are untouched.

Fix

When step 4 of petab1to2_preserve_scale re-injects a log / log10 transformation for an observable, it should also reset that row's noiseDistribution to the v1 base family (normal / laplace). That restores the importer's "give the scale in one place" contract regardless of which petab is installed, and the existing test is the regression pin.

Not a petab bug

Anyone running plain petab1to2 on a log10 problem now gets sigma reinterpreted on the ln scale (off by a factor of ln 10) with only a warning. That is a PEtab v2 specification gap, since v2 has no log10-normal, rather than a library bug, and it is exactly what the PyBNF wrapper exists to work around.