fix(arrow): preserve nulls in validate autofix - #1598

Closes #867.

Summary

• Rewrites the fresh Add auto_coerce option to handle mixed-type columns during Arrow conversion #867 lane away from the reverted auto_coerce API.
• Keeps the existing public validate='autofix' / warn contract on plot(), upload(), and to_arrow().
• Improves pandas mixed-type Arrow autofix to use pandas nullable string dtype when available, preserving missing values as Arrow nulls.
• Keeps a compatibility fallback to standard str coercion for older pandas or nullable-string coercion failures.
• Leaves cuDF behavior on the existing string coercion path.

Design

master already has the right public API shape:

g.plot(validate='autofix', warn=True)
g.upload(validate='autofix', warn=True)
g.to_arrow(validate='autofix', warn=True)

This PR does not add auto_coerce. validate='strict' / validate='strict-fast' continue to raise ArrowConversionError, while validate='autofix' retries Arrow conversion by probing object columns and coercing only columns that fail individual Arrow conversion.

Pandas autofix now prefers nullable string dtype when available so None/missing values stay Arrow nulls. If that dtype is unavailable or fails, it falls back to the prior str behavior.

Review / Oversight

• Review skill converged after 2 waves with 0 blockers, 0 important findings, 0 suggestions.
• Local artifacts: plans/review-pr-1598/.
• Rebased over latest origin/master after it moved to 49e8dd02; final PR head is eee9d399ec4e193b34f283b55890bb8dd952cd48.

Local validation

• python3 -m pytest -q graphistry/tests/test_plotter.py::TestPlotterArrowConversions -> 28 passed, 6 skipped, 1 xfailed
• ./bin/ruff.sh graphistry/PlotterBase.py graphistry/tests/test_plotter.py
• ./bin/typecheck.sh graphistry/PlotterBase.py
• git diff --check

DGX / RAPIDS

• DGX checkout: /home/lmeyerov/repos/pygraphistry-issue-867-dgx
• RAPIDS 26.02: RAPIDS_VERSION=26.02 PROFILE=basic WITH_GPU=1 ... docker/test-rapids-official-local.sh
  • cudf DataFrame _table_to_arrow(..., validate_mode='autofix') smoke passed
  • focused Arrow autofix tests passed: 4 passed
• RAPIDS 25.02: RAPIDS_VERSION=25.02 PROFILE=basic WITH_GPU=1 ... docker/test-rapids-official-local.sh
  • RAPIDS import smoke passed (cudf 25.02.02, cuml 25.02.01)
  • focused Arrow autofix tests passed: 4 passed
  • actual cudf DataFrame construction smoke was attempted first and hit cudaErrorMemoryAllocation out of memory; nvidia-smi showed VLLM::EngineCore using 81870MiB. This PR leaves the cuDF coercion behavior on the existing str path, and 26.02 covers the actual cudf Arrow smoke.

CI

• GitHub Actions run 26262954992 completed success on head eee9d399ec4e193b34f283b55890bb8dd952cd48.
• Green coverage includes python-lint-types py3.8-3.14, test-pandas-compat legacy/latest/rapids-aligned, test-pandas-compat-gfql legacy/latest, test-minimal-python, test-gfql-core, test-docs, test-core-python, test-spark, test-neo4j, test-build, changed-line-coverage, and CodeQL.

Compiler-plan surface touched: no.