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

.github/actions/setup-pybnf/action.yml

@@ -1,16 +1,29 @@
-name: Set up pybnf (bngsim-less)
+name: Set up pybnf
 description: >-
-  Install BioNetGen (BNG2.pl) plus pybnf and its public runtime/test
-  dependencies into a uv venv, deliberately omitting the private, macOS-only
-  bngsim wheel. Tests run under PYBNF_NO_BNGSIM=1 so bngsim-marked tests skip
-  (see tests/conftest.py). BNG2.pl is still required: Configuration validation
-  execs `BNG2.pl -v` for any BNGLModel regardless of bngsim, and the analytical
-  tests aside, most of the suite builds such configs.
+  Install BioNetGen (BNG2.pl) plus pybnf and its runtime/test dependencies into
+  a uv venv. bngsim comes from public PyPI by default (`bngsim` input), so the
+  bngsim-marked suites actually run on hosted runners; pass `bngsim: 'false'`
+  for a leg that must stay bngsim-less (a python with no cp wheel, or a job
+  that does not simulate at all), and run it under PYBNF_NO_BNGSIM=1 so the
+  bngsim-marked tests skip cleanly (see tests/conftest.py). BNG2.pl is required
+  either way: Configuration validation execs `BNG2.pl -v` for any BNGLModel
+  regardless of bngsim, and the analytical tests aside, most of the suite
+  builds such configs.
 
 inputs:
   python-version:
     description: Python version for the venv (e.g. "3.11", "3.12").
     required: true
+  bngsim:
+    description: >-
+      Install bngsim (from public PyPI) alongside the rest of the deps.
+      Default "true". Set "false" on a leg where bngsim must be absent:
+      bngsim publishes manylinux_2_28_x86_64 wheels for cp310-cp313 only, so a
+      newer interpreter would fall back to building the sdist (C++17 toolchain,
+      CMake, Ninja) -- and jobs that never simulate (docs) do not need it. A
+      "false" leg must also set PYBNF_NO_BNGSIM=1 when it runs pytest.
+    required: false
+    default: 'true'
   petab-spec:
     description: >-
       The petab requirement to install. Defaults to stock PyPI petab. A CI leg
@@ -70,24 +83,39 @@ runs:
         tar -xzf "$RUNNER_TEMP/bionetgen.tar.gz" -C "$RUNNER_TEMP"
         echo "BNGPATH=$RUNNER_TEMP/BioNetGen-2.9.3" >> "$GITHUB_ENV"
 
-    - name: Create venv and install pybnf without bngsim
+    - name: Create venv and install pybnf
       shell: bash
+      env:
+        BNGSIM: ${{ inputs.bngsim }}
+        PETAB_SPEC: ${{ inputs.petab-spec }}
       run: |
         # Build the venv from the setup-python interpreter (see note above), not
         # a uv-managed one.
         uv venv --python "${{ steps.python.outputs.python-path }}"
-        # Mirror pyproject.toml's runtime deps MINUS bngsim (private, mac-only
-        # wheel, not on PyPI), plus the test stack (pytest, hypothesis,
-        # pytest-xdist) and petab. petab is BOTH the PEtab v2 export/BnglModel
-        # oracle (ADR-0025/0026 -- BNG2.pl is installed above, so model-level
-        # validation runs here too) AND the runtime math translator backing the
-        # expression observableFormula path (pybnf[petab], ADR-0035, petab.v2.math);
-        # stock PyPI petab carries the math grammar, so one install covers both.
-        # The exact petab requirement is the `petab-spec` input (default stock
-        # PyPI; a CI leg overrides it with the Step B fork to run the NATIVE
-        # BNGL loader). Keep these version bounds in sync with pyproject.toml;
-        # bngsim is the only intentional omission.
+        # bngsim is on public PyPI (manylinux_2_28_x86_64 + macOS + win wheels,
+        # cp310-cp313), so hosted runners get the real simulation backend --
+        # NFsim, RuleMonkey, libsbml, codegen and output sensitivities are all
+        # compiled into the published wheel. The [antimony] extra is the one
+        # capability that needs a companion PyPI package (antimony>=3.1.1, pure
+        # python wheel), and it is what unskips the bngsim_antimony-marked
+        # tests, so install the extra rather than bare bngsim. Keep the pin in
+        # sync with pyproject.toml.
+        BNGSIM_SPEC=()
+        if [ "$BNGSIM" = 'true' ]; then
+          BNGSIM_SPEC=('bngsim[antimony]>=0.11.35,<1')
+        fi
+        # Mirror pyproject.toml's runtime deps, plus the test stack (pytest,
+        # hypothesis, pytest-xdist) and petab. petab is BOTH the PEtab v2
+        # export/BnglModel oracle (ADR-0025/0026 -- BNG2.pl is installed above,
+        # so model-level validation runs here too) AND the runtime math
+        # translator backing the expression observableFormula path
+        # (pybnf[petab], ADR-0035, petab.v2.math); stock PyPI petab carries the
+        # math grammar, so one install covers both. The exact petab requirement
+        # is the `petab-spec` input (default stock PyPI; a CI leg overrides it
+        # with the Step B fork to run the NATIVE BNGL loader). Keep these
+        # version bounds in sync with pyproject.toml.
         uv pip install \
+          "${BNGSIM_SPEC[@]}" \
           'dask>=2024.1.0' \
           'distributed>=2024.1.0' \
           'libroadrunner>=1.6.0' \
@@ -101,7 +129,9 @@ runs:
           'pytest>=7,<10' \
           'hypothesis>=6,<7' \
           'pytest-xdist' \
-          '${{ inputs.petab-spec }}'
-        # Install pybnf itself without re-resolving deps (which would pull
-        # bngsim and fail).
+          "$PETAB_SPEC"
+        # Install pybnf itself without re-resolving deps: on a bngsim: 'false'
+        # leg a re-resolve would pull bngsim in through pyproject.toml and
+        # defeat the point, and everywhere it would re-derive a dependency set
+        # this step just pinned.
         uv pip install --no-deps .

.github/workflows/docs.yml

@@ -49,14 +49,18 @@ jobs:
     steps:
       - uses: actions/checkout@v6
 
-      # Reuse the tests' composite action: it installs pybnf + its PUBLIC runtime
-      # deps into a uv venv WITHOUT the private, mac-only bngsim wheel -- exactly
-      # the environment autodoc needs to import pybnf on Linux. (It also fetches
-      # BNG2.pl, unused here but harmless; the win is one source of truth for the
-      # dependency list.)
+      # Reuse the tests' composite action: it installs pybnf + its runtime deps
+      # into a uv venv -- exactly the environment autodoc needs to import pybnf
+      # on Linux. (It also fetches BNG2.pl, unused here but harmless; the win is
+      # one source of truth for the dependency list.)
       - uses: ./.github/actions/setup-pybnf
         with:
           python-version: '3.12'
+          # autodoc only imports pybnf; it never simulates. Skip the bngsim
+          # install and build under PYBNF_NO_BNGSIM=1 (below) so the docs job
+          # stays lean -- and so a docs build keeps proving that pybnf imports
+          # cleanly with no backend present.
+          bngsim: 'false'
           # Distinct uv cache namespace so this job does not race the tests /
           # lint jobs' content-addressed cache saves on the same push.
           cache-suffix: docs

.github/workflows/integration.yml

@@ -3,10 +3,10 @@ name: integration
 # Manual only. The slow tier is the 4 statistical-recovery tests against
 # analytical targets (AnalyticalModel) -- bngsim-free and BNG2.pl-free, ~8 min.
 #
-# NOTE: the bngsim *simulation* suites (true NFsim/RuleMonkey/SBML runs) are
-# deliberately NOT in hosted CI -- they need the private, macOS-only bngsim
-# wheel. They remain covered by the local pre-push hook (.pre-commit-config.yaml)
-# on a dev machine until bngsim ships Linux wheels.
+# NOTE: the fast bngsim *simulation* suites (true NFsim/RuleMonkey/SBML runs)
+# DO run in hosted CI -- tests.yml installs bngsim from public PyPI (#514). Not
+# covered anywhere hosted yet: the `recovery` tier (real parameter-recovery fits
+# through the backend, over an hour), which still runs on a dev machine.
 #
 # Runs nightly and on manual dispatch.
 on:
@@ -27,6 +27,10 @@ jobs:
       - uses: ./.github/actions/setup-pybnf
         with:
           python-version: '3.12'
+          # The slow tier is analytical (AnalyticalModel): no simulation
+          # backend, so no bngsim install, and PYBNF_NO_BNGSIM=1 below keeps it
+          # honest about that.
+          bngsim: 'false'
 
       - name: Run slow analytical recovery tests
         env:

.github/workflows/tests.yml

@@ -14,26 +14,49 @@ jobs:
         # Every supported version (requires-python >= 3.11; classifiers list
         # through 3.14). CI is free on this public repo, so we test each claimed
         # version rather than just the endpoints.
-        python-version: ['3.11', '3.12', '3.13', '3.14']
-    name: pytest (py${{ matrix.python-version }})
+        include:
+          - python-version: '3.11'
+          - python-version: '3.12'
+          - python-version: '3.13'
+          # bngsim publishes cp310-cp313 wheels, so 3.14 would fall back to
+          # building the sdist (C++17 toolchain, CMake >= 3.20, Ninja) inside
+          # the pytest job. Keep this leg deliberately bngsim-less instead: it
+          # is the guard that the bngsim-free code paths still work (a bare
+          # `pip install pybnf` on a machine with no wheel), which is worth a
+          # leg of its own regardless. Drop `no-bngsim` here the day a cp314
+          # wheel ships.
+          - python-version: '3.14'
+            no-bngsim: '1'
+    name: pytest (py${{ matrix.python-version }}${{ matrix.no-bngsim == '1' && ', bngsim-less' || '' }})
     steps:
       - uses: actions/checkout@v6
 
       - uses: ./.github/actions/setup-pybnf
         with:
           python-version: ${{ matrix.python-version }}
+          bngsim: ${{ matrix.no-bngsim == '1' && 'false' || 'true' }}
 
-      - name: Run unit tests (bngsim-less, parallel)
+      - name: Run unit tests (parallel)
         env:
-          # bngsim is a private mac-only wheel; with it absent, conftest.py
-          # auto-skips bngsim/nfsim/sbml/etc.-marked tests. The remaining suite
-          # runs against the bngsim-less code paths (and BNG2.pl from the
-          # composite action). -n auto parallelizes across the runner's cores.
-          PYBNF_NO_BNGSIM: '1'
+          # Empty except on the bngsim-less leg, where it is '1'. An empty value
+          # reads as unset (pybnf/_bngsim_caps.py tests truthiness), so bngsim
+          # is imported and the bngsim/nfsim/rulemonkey/sbml/antimony-marked
+          # tests run for real; with '1' conftest.py auto-skips them.
+          PYBNF_NO_BNGSIM: ${{ matrix.no-bngsim }}
+        # -m: the default addopts expression is `not slow and not recovery`, and
+        # naming -m here REPLACES it -- so both tiers must be restated. Omitting
+        # `not recovery` used to be harmless because 129 of the 130 recovery
+        # tests skip without bngsim; with bngsim installed they are real
+        # parameter-recovery fits through the backend and the job would run for
+        # over an hour. (The cost of deselecting: the 130th, a roadrunner-backed
+        # Boehm oracle in test_petab_sbml_layer.py, is bngsim-free and no longer
+        # runs here.) -n auto parallelizes across the runner's cores.
+        #
         # --no-sync: do NOT let `uv run` re-resolve the project env from
-        # pyproject.toml (that would pull the private bngsim and fail). The
-        # composite action already populated .venv.
-        run: uv run --no-sync pytest -m "not slow" -n auto
+        # pyproject.toml -- the composite action already populated .venv with an
+        # exact dependency set (and on the bngsim-less leg a re-resolve would
+        # pull bngsim right back in).
+        run: uv run --no-sync pytest -m "not slow and not recovery" -n auto
 
   # The PEtab v2 export oracle (ADR-0025/0026) validates emitted `language: bngl`
   # problems at MODEL level. On the pytest job above (stock PyPI petab) that runs
@@ -51,6 +74,11 @@ jobs:
       - uses: ./.github/actions/setup-pybnf
         with:
           python-version: '3.12'
+          # This leg is about petab's loader, not simulation: the two oracle
+          # files it runs are backend-free, so skip the bngsim install (and run
+          # under PYBNF_NO_BNGSIM=1 below) rather than pay for a wheel nothing
+          # here exercises. The matrix legs above cover bngsim.
+          bngsim: 'false'
           petab-spec: 'petab @ git+https://github.com/wshlavacek/libpetab-python.git@bngl_model_support'
           # This job also runs 3.12 through setup-pybnf, so without its own
           # suffix it shares the matrix py3.12 job's cache key and one save loses
@@ -71,4 +99,4 @@ jobs:
         # `language: bngl` through petab's own native BnglModel here (the
         # register_bngl() monkeypatch collapses to a no-op). This is where the
         # TestRegisterBngl native-no-op test actually runs instead of skipping.
-        run: uv run --no-sync pytest -m "not slow" tests/test_petab_export.py tests/test_petab_import.py
+        run: uv run --no-sync pytest -m "not slow and not recovery" tests/test_petab_export.py tests/test_petab_import.py

.pre-commit-config.yaml

@@ -8,18 +8,23 @@
 #      `ruff@0.15.14 check .`). Runs FIRST and takes seconds, so a lint failure
 #      aborts the push immediately instead of after the pytest gate (the pytest
 #      hook does NOT run ruff -- that gap let an F401 reach CI once).
-#   2. pytest-bngsim -- the bngsim (mac-only) tests CI *cannot* run, since the
-#      private bngsim wheel is absent on hosted runners. This is the gate's
-#      unique value: every other test already runs in CI. Run with `-n auto`
-#      so the whole gate finishes in well under a minute.
+#   2. pytest-bngsim -- the simulation-backend tests, as a fast local smoke.
+#      These are no longer CI-exclusive: bngsim is on public PyPI and tests.yml
+#      installs it, so hosted runners now run these files too (#514). The hook
+#      stays because it is the shortest path from "I broke the backend seam" to
+#      knowing it -- seconds locally instead of a CI round-trip. Run with
+#      `-n auto` so the whole gate finishes in well under a minute.
 #
-# Deliberately NOT here: the bngsim-less suite (`PYBNF_NO_BNGSIM=1 pytest
-# -m "not slow" -n auto`). CI's tests.yml runs that exact command on every push,
-# so mirroring it locally added minutes to the gate for zero extra coverage --
-# and a multi-minute pre-push hook holds the push's SSH connection idle long
-# enough for GitHub to reap it ("Connection to github.com closed by remote
-# host"), so the push silently never lands. Keeping the gate fast is what makes
-# `git push` actually complete. Bngsim-less regressions still surface in CI.
+# Deliberately NOT here: the rest of the suite (`pytest -m "not slow and not
+# recovery" -n auto`, ~3300 tests). CI's tests.yml runs that on every push, so
+# mirroring it locally added minutes to the gate -- and a multi-minute pre-push
+# hook holds the push's SSH connection idle long enough for GitHub to reap it
+# ("Connection to github.com closed by remote host"), so the push silently never
+# lands. Keeping the gate fast is what makes `git push` actually complete.
+#
+# Also NOT here: the `recovery` tier (real parameter-recovery fits through the
+# backend, over an hour). That one runs nowhere automatically -- run it by hand
+# when a change touches the fitting loop.
 #
 # `uv run --no-sync` uses the project's interpreter without touching the
 # environment -- the .venv is already populated and uv.lock isn't fully
@@ -40,9 +45,9 @@ repos:
         stages: [pre-push]
       - id: pytest-bngsim
         name: pytest (bngsim test suite)
-        # Every bngsim-area test file: bngsim is mac-only and absent in CI, so
-        # these run nowhere else. `-n auto` parallelizes across cores to keep the
-        # gate fast. Add files here as bngsim coverage expands.
+        # Every bngsim-area test file, as a pre-push smoke (CI runs them too --
+        # this is the fast local copy). `-n auto` parallelizes across cores to
+        # keep the gate fast. Add files here as bngsim coverage expands.
         entry: >-
           uv run --no-sync pytest -m "not bionetgen" -n auto
           tests/test_bngsim_bridge.py

tests/README_integration.md

@@ -60,8 +60,11 @@ pytest -m recovery -k m03 # just one model
 
 Requirements / behaviour:
 
-- **bngsim** must be installed (the macOS-only C++ wheel) — auto-skips via the
-  `bngsim` marker otherwise, so this never runs in hosted CI.
+- **bngsim** must be installed (public PyPI: `pip install bngsim`) — auto-skips
+  via the `bngsim` marker otherwise. Hosted CI installs bngsim and runs the
+  bngsim-marked suites, but explicitly deselects this tier
+  (`-m "not slow and not recovery"`): the fits across all 16 recovery-marked
+  files run well over an hour. It stays a dev-machine tier.
 - **BNG2.pl** must be resolvable (via `BNGPATH`): bngsim is a simulation engine,
   not a network generator, so rules→`.net` expansion runs once per fit at setup.
   `recovery_harness.require_bng2pl` skips the tier if it isn't found.

tests/test_gradient_assembly.py

@@ -2980,6 +2980,9 @@ def test_fd_acceptance_gate_estimated_scale(model_kind):
     chain rule end-to-end on the real bngsim decay net. The scale never enters the data fit, so its
     whole gradient comes from the constraint's scalar column -- the estimated-noise pattern, applied
     to a qualitative constraint."""
+    # ``s_q`` is loguniform, so the scale column's log-space chain rule autodiffs through jax
+    # (the optional pybnf[jax] extra) -- same declaration as every other log-scale leg here.
+    pytest.importorskip('jax')
     obj = ChiSquareObjective()
     sigma = 5.0
     model_name = 'e2e_ode_decay'