Unblock CI: install the real test deps and run a python 3.10-3.14 matrix

[outofculture]

Why

CI has been red since at least October 2025, and never once ran a test. Every run died at collection in ~30s:

ERROR neuroanalysis/tests/test_release_model.py  -> ModuleNotFoundError: No module named 'lmfit'
ERROR neuroanalysis/tests/test_test_pulse.py     -> ModuleNotFoundError: No module named 'neuron'
!!!! Interrupted: 2 errors during collection !!!!  -> exit code 2

[project.optional-dependencies] test declared only pytest, but the suite imports lmfit (via synaptic_release.py) and neuron (top-level in test_test_pulse.py). Nothing that pip install -e .[test] produced could collect the suite.

This PR makes the suite installable and collectable, and puts CI on a real interpreter matrix. It deliberately does not try to fix the underlying test failures — see "What this does not fix".

What changed

Commit	Change
e8fd395	np.product -> np.prod in fitting/searchfit.py. np.product was removed in numpy 2.0, so building the parameter-space index raised AttributeError for every search fit.
8dd233b	Move tools/conftest.py back to the repo root, and defer the pyqtgraph import into the audit branch of test_psp_fitting.
9bc19ed	Declare requires-python = ">=3.10"; add lmfit and neuron to the test extra.
1188914	Mark the two unimplemented noise tests as strict xfail.
3770c24	Run CI against python 3.10 through 3.14 with fail-fast: false; drop the $CONDA/bin step; bump setup-python to v5.

The conftest move

conftest.py was relocated under tools/ in 82f058e (Dec 2021) during the packaging reorg, where pytest never loads it. It is the only thing defining --audit, so all 17 test_psp_fitting cases died on ValueError: no option named 'audit'. test_spike_detection survived only because it wraps the lookup in try/except.

The pyqtgraph import in test_psp_fitting sat at the top of the test body and ran unconditionally, so the tests needed a Qt binding even for non-audit runs. It now loads inside if audit and test_ui is None:, matching how test_spike_detection already handles its UI.

Why the python version pin was wrong

git show 14e07db shows the workflow arrived as GitHub's "Python Package using Conda" starter template verbatim. Set up Python 3.10 shipped paired with the $CONDA/bin line because the install step was conda env update --file environment.yml. That file is long gone, but the conda path entry stayed and shadowed setup-python — so CI has actually been running on miniconda's interpreter, not the one it selected. The most recent run reports python 3.14.

Dropping that step and running a real matrix instead. 3.10 is a genuine floor: both neuron and pyqtgraph declare requires_python >=3.10. Both ship wheels through cp314, as does numba, so every leg installs.

fail-fast: false so one version's failures do not hide another's.

timeout-minutes: 30

Added as a guard rail. Once collection is unblocked, test_psp_fitting actually executes, and several of its brute-force parameter searches run for minutes each (details below). Across five matrix legs that would otherwise sit on the 6h default. This bounds the damage without silencing anything.

What this does not fix

Unblocking collection reveals 134 genuine failures that have accumulated over the years. None are addressed here.

Locally, with every optional dependency present, the suite minus test_psp_fitting is 117 failed / 1444 passed / 2 xfailed in 2m10s. Expect the matrix to be red on every leg. That is the point — the failures are now visible instead of hidden behind an exit-2.

Count	Test	Cause
98	test_spike_detection	Stale recorded expectations in the test_data submodule: 86 .pkl files have expected_result = None (predating the switch to returning a list), and 5 use pre-rename keys max_dvdt/max_dvdt_time/peak. The other 7 are real count mismatches.
17	test_psp_fitting	Value drift at rtol=0.01, plus pathological runtime — 4 cases fail in ~5s, at least 9 more run past 60s each.
15	test_exp_fit	Accuracy assertions at degenerate corners (tau far below the sample interval, 1e-9-scale signals).
1	test_model_spike_detection	A 100 pA sub-threshold pulse into ModelCell now yields 1 spike where 0 is expected. Reproduces 5/5, so not a noise flake.
1	test_analyzers	Calls stim_params(), deleted from GenericStimPulseAnalyzer by 89be28c ("Remove dead code", Sep 2020).
1	test_trace_timing	Expects TSeries(data=np.zeros((10,10))) to raise; 3fe3da4 deliberately made TSeries n-dimensional and the check is commented out.
1	test_release_model	ReleaseModel.__init__ was replaced by b0ac6c8 with a new signature. The class is unconstructable regardless: self.ode_variable=={...} is a comparison against a never-assigned attribute.

Worth noting: every spike-detection case whose counts and key sets line up passes at rtol=0.01. The detector is largely intact; it is the baselines that rotted.

Reviewer notes

• Adding neuron to the test extra makes it a hard test dependency. It ships linux wheels for cp310-cp314, so this installs cleanly, but it is a heavy dependency for a test run. Say the word if you would rather guard the import in test_test_pulse.py and keep it in neuron_sim.
• The xfails are strict=True, so if someone implements those noise tests the marker has to come off.
• If you have a local submodule.test_data.url pointing at aiephys/neuroanalysis-test-data, it no longer has the pinned commit — git submodule sync test_data picks the right URL back up from .gitmodules. CI is unaffected.