pytest.ini

[pytest]
# Pytest configuration for EROTICA
minversion = 8.0
addopts =
    -ra
    --strict-markers
testpaths = tests
python_files = test_*.py
python_classes = Test*
python_functions = test_*
markers =
    slow: marks tests as slow (deselect with '-m "not slow"')
    integration: marks tests as integration tests
    unit: marks tests as unit tests
    plotting: marks tests that generate plots
filterwarnings =
    ignore::UserWarning
    ignore::DeprecationWarning
    ignore::FutureWarning
    # RuntimeWarning is an ERROR here, not noise. In a numerics package
    # "divide by zero", "invalid value in sqrt" and "overflow" are how a silent
    # NaN enters a published number. This was a blanket `ignore::RuntimeWarning`
    # until 2026-07-27; auditing it showed every occurrence came from PyTensor's
    # own graph rewriter, not from erotica -- so nothing of ours was being
    # hidden, and nothing of ours may be hidden from now on.
    error::RuntimeWarning
    # Third-party, unfixable from here: constant folding inside PyTensor's own
    # rewriter arithmetic (mul_calculate) divides by zero while simplifying a
    # graph. Scoped narrowly rather than ignored wholesale. NOTE: an earlier
    # version of this comment claimed "the rewritten graph is never the one that
    # runs" -- that is false, rewritten graphs are exactly what run. What is
    # actually checked is that the isochrone models recover injected truth; the
    # affected models use no Cauchy and so are unrelated to the numba CauchyRV
    # bug documented in tools/validation/pytensor_cauchy_bug_report.md.
    # NUTS's kinetic-energy overflow IS its divergence detector, not a silent NaN, so the
    # policy above does not apply to it. Traced through the source rather than assumed:
    # integration.py computes `kinetic = pot.velocity_energy(p_new, v_new)`, which is
    # `0.5 * np.dot(x, v_out)` in quadpotential.py. A momentum blow-up overflows that dot to
    # +inf, so `energy_change = right.energy - self.start_energy` is inf, the guard
    # `if energy_change < self.Emax` fails, and nuts.py returns a DivergenceInfo. The
    # trajectory is REJECTED and the value never reaches the posterior.
    #
    # Surfaced 2026-08-04, when CI resolved pytensor 3.2.4 + numba 0.66 + jax 0.11 + pymc
    # 6.2 -- a stack the working environment does not have -- and four test_inference tests
    # began erroring. Escalating it ABORTS sampling at the first divergence, which is
    # strictly worse than letting NUTS do its job: it hides the divergence COUNT behind a
    # crash, so you cannot tell one benign divergence from a broken geometry.
    #
    # The gate is not removed, it is moved to where it belongs: every affected test asserts
    # zero divergent transitions on its own trace. A real problem still fails, with a number
    # instead of a warning.
    ignore:overflow encountered in dot:RuntimeWarning:pymc.step_methods.hmc.quadpotential
    ignore:overflow encountered in dot:RuntimeWarning:pymc.step_methods.hmc.nuts
    ignore:divide by zero encountered:RuntimeWarning:pytensor.tensor.rewriting.math
    ignore:overflow encountered:RuntimeWarning:pytensor.tensor.rewriting.math
    ignore:invalid value encountered:RuntimeWarning:pytensor.tensor.rewriting.math