fix(petab): undo petab 0.10.0's log(10) factor on a converted log10 observable's sigma - #1017

CI started installing petab 0.10.0 (released 2026-10-26), and 10 tests in tests/test_petab_convert.py fail on every leg. One of the causes is a silently wrong objective, which this PR fixes in pybnf/petab/convert.py.

What petab 0.10.0 changed, and what it did to PyBNF

A v1 log10 observable now converts with a rescaled noise formula. petab 0.10.0 writes log10-normal as log-normal with noise formula log(10) * (<v1 formula>). That is exact in v2's own terms. petab1to2_preserve_scale states the scale differently (ADR-0073): it re-injects observableTransformation = log10 and resets noiseDistribution to normal. Under that column the v1 σ is the right one, so together the two gave a log10 residual whose σ was ln 10 times the stated one. For a v1 observable with noiseFormula = 0.1:

A placeholder σ failed at import instead, which is the failure CI showed.

The priors petab 0.9.x refused now convert. These are logNormal, logLaplace, and log10 parameterScaleUniform / parameterScaleLaplace. The converter already rewrites every declared prior from its v1 row (#893), so only the test that expected a refusal failed.

The warning text changed. petab's warning for the log10 case is reworded, so it was no longer silenced.

Commits

1. fix(petab): undo petab 0.10.0's log(10) factor on a converted log10 observable's sigma.
   • restore_v1_noise_formulas puts each log10 observable's v1 noiseFormula back. It removes exactly the log(10) * (...) wrapper and leaves an unchanged formula alone. Any other rewrite is refused as a PybnfError naming both formulas, since the cell sets σ.
   • The new warning wording joins the silenced patterns.
   • On petab ≥ 0.10 the formerly refused priors run through the three existing prior oracles: hand-written v1/v2 densities, PyBNF's imported prior, and libpetab's own reading. That is what the test file's own comment said to do. The refusal test runs only where petab still refuses.
   • ADR-0073 gets a dated addendum.
2. fix(petab): read the converted observables table as text. This fixes a review finding. On petab 0.9 with pandas < 3.1, a 17-digit numeric σ was refused, because the converter re-read the v2 table with a float parse that moves about 40% of such values by an ulp. The table is now read as text. Commit 2 also narrows my "same under 0.9 and 0.10" claim to the noise formula, and the test reads petab's version from package metadata.

How it was checked

• Both petab releases: test_petab_convert.py plus the review file give 173 passed, 8 skipped on petab 0.10.0 / pandas 3.1.0, and 156 passed, 1 xfailed on petab 0.9.0 / pandas 3.0.5.
• Mutation: with the restore disabled, both log10 full-conversion tests fail on 0.10, and 14 of the reviewer's 17 fail.
• Independent oracle: the reviewer's tests compare the objective PyBNF builds from a converted, imported problem with libpetab's own v1 likelihood (calculate_llh) on the original tables. The cases cover numeric, placeholder, fixed-parameter, expression and prediction-dependent σ; log10, log and lin; and several observables across two files.
• Siblings:
  • A log (natural) observable: petab never rescales it.
  • Native v2 problems don't go through petab1to2, and the PEtab export doesn't call it.
  • The reviewer ran a 69-case conversion matrix. Wherever both releases convert, the converted observables.tsv and the imported .conf are byte-identical across releases.
• Independent review (a subagent that did not write the change) found nothing silently wrong in the change, and the one loud regression fixed in commit 2. Its tests are in tests/test_petab_convert_review.py.

Not in this PR (found in review)

• petab 0.9's petab1to2 silently drops a pre-equilibration whose condition changes nothing. PyBNF then imports an experiment that starts from the initial conditions, which is a wrong objective, and pyproject still allows petab>=0.9. petab 0.10 keeps it. This is upstream and pre-existing; a strict xfail on petab 0.9 records it. PyBNF could guard against it, or raise its floor to petab 0.10.
• A noise formula over non-estimated parameters fails at the first evaluation rather than at load: the job loop turns the error into a failed simulation.
• The silenced "Parameter scales are not supported" warning now carries a petab 0.10 sentence about the maximum a posteriori estimate changing, for converted parameter-scale priors. ADR-0073's mapping is exactly that conversion; worth a look.