Cap pyarrow<23 to fix QGIS 4 import failure

[sunt05]

Summary

• Cap pyarrow<23 on non-Linux and Python 3.14 dependency lines in pyproject.toml
• pyarrow 23.0.1 removed pyarrow.compute.match_substring_regex, breaking pandas.Series.str.contains() with regex on arrow-backed strings
• This caused supy to fail on import in OSGeo4W/QGIS 4 environments (Python 3.12 + pandas 3.0.1 + pyarrow 23)

Closes #1252

Test plan

• Verified fix on @biglimp's QGIS 4 machine via SSH (downgraded pyarrow 23 -> 22, supy.load_sample_data() works)
• CI passes with capped pyarrow versions

Notes

The docstring-parsing code in _load.py that triggered this has already been refactored out on master (replaced by static _var_metadata). A new release would also resolve this inherently for the released version. The cap is a safety net until pandas releases a version compatible with pyarrow 23.

🤖 Generated with Claude Code

[github-actions]

CI Build Plan

Changed Files

Python source (0 files)

Documentation (1 file)

• pyproject.toml

pyproject.toml (1 file)

• pyproject.toml

Build Configuration

	Configuration
Platforms	Linux x86_64
Python	3.9, 3.14
Test tier	standard (all except slow)
UMEP build	Skipped (no ABI changes)
PR status	Ready (standard matrix)

Rationale

• Python source changed -> single-platform build
• No compiled extension changes -> UMEP build skipped (nightly provides coverage)

---

_{Updated by CI on each push. See path-filters.yml for category definitions.}

[sunt05]

Closing in favour of UMEP-dev/umep-reqs#14 — the pyarrow cap belongs in umep-reqs (QGIS install layer), not in supy itself. The incompatible code path is already removed on master.