ci: Test against lowest supported direct dependencies

[cbueth]

This PR adds a lowest-direct row to the test matrix so CI verifies the suite against the minimum versions we claim to support, not just latest. Can't use uv sync --resolution lowest-direct directly. This somehow re-resolves all extras and dies building old 2008 matplotlib/pytest. The job compiles base deps with uv pip compile --resolution lowest-direct, installs those, then installs .[test] on top.

I had to add lower-bound pins, all the lowest versions that resolve/build/import/pass on Python 3.11:

• numpy>=2.0.1 (released 2024-07-21). 2.0.0 breaks infomeasure grassberger reduction (np.unique(axis=0, return_inverse=True) returns a column vector). 2.0.1 is the fix and still fits numba 0.59/0.60's ceiling
• scipy>=1.13.0 (2024-04-02). first 1.x without a numpy<2 upper bound
• statsmodels>=0.13.3 (2022-11-01). oldest build that installs on 3.11 and imports cleanly under numpy 2
• scikit-learn>=1.5.0 (2024-05-21). first version whose metadata allows numpy 2
• numba>=0.59.0 (2024-01-31). first with numpy 2 support. lowest-direct actually resolves 0.60.0 because of the numpy floor
• infomeasure>=0.6.2 (2026-07-06). 0.6.1 + sparse 0.19 dies with "Cannot convert a sparse array to dense" in the discrete TE path
• igraph>=0.11.5 (2024-05-07). 0.10.x–0.11.4 emit RuntimeWarning in eigenvector centrality on weighted directed graphs

Also:

• Added deterministic registry iteration in detrending_methods/__init__.py (set iteration made dict order differ per process, tripping pytest-xdist's cross-worker collection check). conftest.py fixture sorted too

[codspeed-hq]

Merging this PR will degrade performance by 22.72%

⚠️ Different runtime environments detected

Some benchmarks with significant performance changes were compared across different runtime environments,
which may affect the accuracy of the results.

Open the report in CodSpeed to investigate

❌ 1 regressed benchmark
✅ 27 untouched benchmarks

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
❌	test_detrend_method[z_score-p10]	1.4 ms	1.8 ms	-22.72%

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing ci/lowest-direct (a0169ed) with main (fe011f7)}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 100.00%. Comparing base (fe011f7) to head (a0169ed).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff            @@
##              main       #60   +/-   ##
=========================================
  Coverage   100.00%   100.00%           
=========================================
  Files           29        29           
  Lines          859       859           
  Branches       149       149           
=========================================
  Hits           859       859           

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.