Resolve blocking Python 3.15 dependency compatibility

[stefan-jansen]

Failure

The required Python 3.15 prerelease matrix is blocked by two core third-party dependencies.

Polars

Affected libraries: data, engineer.

Polars 1.43.1 and 1.43.2 fail common expression and datetime operations on CPython 3.15.0rc1
before ML4T code can process the result. The exact Linux, macOS, and Windows qualification jobs
fail.

Upstream evidence:

• pl.DataFrame fails on datetime.datetime in 3.15 Betas pola-rs/polars#28347 tracks the accepted CPython 3.15 failure.
• fix(python): Support Python 3.15 expression dispatch pola-rs/polars#28750 contains a focused fix but is not yet merged or released.

SciPy

Affected library: diagnostic.

SciPy 1.18.0 publishes no CPython 3.15 wheels on PyPI. No compatible prerelease wheels were
available from the configured scientific Python prerelease index at the 2026-08-11 evidence cutoff.
Linux, macOS, and Windows therefore attempt a source build and fail before the test suite starts
because the runner environment does not provide the required OpenBLAS build dependencies.

Resolved repository-owned failures

• NumPy 2.5.2 now provides CPython 3.15 wheels, including Windows. Affected locks were updated in
  data, backtest, live, and models.
• models now qualifies its hardware-independent core without making prerelease PyTorch a core
  requirement.
• Optional scientific and ML integrations were separated from core qualification only where the
  public core package does not require them.
• Required Polars behavior remains in the data and engineer suites.
• Required SciPy behavior remains in the diagnostic suite.
• specs passes Python 3.15 installation and tests on Linux, macOS, and Windows.

Release impact

The shared support policy makes Python 3.15 prerelease compatibility blocking before October 1,
2026. An affected library must not be released as qualified while its required job fails. Stable
Python 3.12 through 3.14 jobs completed so far pass.

Acceptance criteria

• An upstream Polars release containing the fix passes the complete data and engineer Python
  3.15 suites on Linux, macOS, and Windows.
• A SciPy release or prerelease publishes compatible wheels and the complete diagnostic Python
  3.15 suite passes on all three operating systems.
• Every other library's current integration commit passes its required Python 3.15 jobs.
• The merged default branches repeat the complete qualification successfully.
• Generated ecosystem status records every required lane as passing.

[stefan-jansen]

Additional Python 3.15 rc1 evidence from the exact qualification matrix:\n\n- ml4t-engineer originally failed on Numba, PyArrow, and unused scientific core dependencies. PR ml4t/engineer#34 now isolates those optional dependencies; local Python 3.15 core installation and fallback tests pass.\n- ml4t-diagnostic fails core installation because scikit-learn requires SciPy. SciPy 1.18.0 has no CPython 3.15 wheel on PyPI or the Scientific Python nightly index, and its source build fails without OpenBLAS. SciPy is used directly throughout diagnostic, so it cannot be removed or skipped without reducing supported functionality.\n- ml4t-models completed all 409 tests on macOS but the July 25 PyTorch nightly exited with signal 11 at shutdown. PR ml4t/models#36 now uses the August 10 nightly, the newest date with Python 3.15 wheels on Linux, macOS, and Windows; Linux passes and the cross-platform rerun is pending.\n\nThese are blocking checks under the approved policy. No test exclusions or source-build system dependency workarounds have been added.