ci: install bngsim from PyPI so the simulation suites run in CI (#514)

[wshlavacek]

Closes #514.

tests.yml disabled the entire bngsim surface on the premise that bngsim is a private, macOS-only wheel. It is neither — bngsim 0.11.35, exactly the version pyproject.toml pins, is on public PyPI with manylinux_2_28_x86_64 wheels for cp310–cp313.

The published wheel is fully capable

Installed from PyPI into a clean venv, every capability PyBNF gates on comes back available:

feature	published wheel
nfsim, rulemonkey	✅
libsbml, sbml_import, sbml_ssa, sbml_psa	✅
codegen, output_sensitivities	✅
SteadyStateResult.output_sensitivities	✅
antimony_import	needs the companion antimony PyPI package (manylinux wheel exists)

So the action installs bngsim[antimony] — the bngsim_antimony-marked tests unskip along with everything else.

Changes

• setup-pybnf grows a bngsim input (default true). A leg that must stay bngsim-less passes 'false' and keeps PYBNF_NO_BNGSIM=1: py3.14 (no cp314 wheel — it would fall back to an sdist build), the docs job (autodoc never simulates), the native-BNGL petab leg, and integration.yml's analytical slow tier.
• The pytest matrix moves to include so 3.14 can carry no-bngsim: '1'. That leg is no longer an accident of packaging: it is now the deliberate guard that the bngsim-free code paths still work, which is worth a leg regardless. On the other legs PYBNF_NO_BNGSIM is empty, which _bngsim_caps.py reads as unset.
• The marker expression becomes -m "not slow and not recovery". Naming -m on the command line replaces the addopts default, so the old command was also selecting the recovery tier — harmless only because 129 of its 130 tests skip without bngsim. With bngsim present they are real parameter-recovery fits and the job would run for over an hour. The cost of deselecting is the 130th (a roadrunner-backed Boehm oracle, bngsim-free); noted in the comment.
• One real bug the new coverage exposed. test_fd_acceptance_gate_estimated_scale is @pytest.mark.bngsim and fits a loguniform_var parameter, whose log-space chain rule autodiffs through jax — an extra CI does not install. Its twelve siblings in the file declare pytest.importorskip('jax'); this one did not, so it failed rather than skipped the moment bngsim arrived. Declared.
• Stale "private / mac-only" rationale corrected in .pre-commit-config.yaml, integration.yml, docs.yml and tests/README_integration.md. ADRs left alone as historical records.

Verification

The CI install was mirrored exactly — bngsim from PyPI, not the local dev wheel — and the job's own command run against it:

3221 passed, 78 skipped, 0 failed in 4m28s

No bngsim-family test skips remain. The 78 residual skips are all optional extras CI has never installed (jax/blackjax, arviz, matplotlib) — test_hmc.py alone loses ~40 tests to that, which is worth a separate issue.

Landing as a hard gate rather than behind continue-on-error: the local run is 0-failure against the published wheel, and a non-blocking gate is the failure mode #514 is about.

Follow-ups (not in this PR)

• The composite action installs a hand-mirrored dep list that omits the tests extra's jax / arviz / matplotlib, so those oracles skip in CI.
• The recovery tier still runs nowhere automatically.

🤖 Generated with Claude Code