fix: declare pyarrow, pin worker interpreter, add CI + cleanup

[hamed]

The test suite had never been run. Doing so surfaced two defects that would
hit any user or contributor installing from a clean environment.

pyarrow was an undeclared runtime dependency

bootstrap.sample() calls applyInPandas, which requires Arrow
serialization. pyarrow appears nowhere in pyproject.toml, so a clean
pip install replicas produced a package whose core function raised:

PySparkImportError: [PACKAGE_NOT_INSTALLED] PyArrow >= 18.0.0 must be
installed; however, it was not found.

Declared as pyarrow>=4.

Spark workers ran the wrong interpreter

With pyarrow present the driver was fine, but every worker died:

File ".../pyspark/worker.py", line 2930, in read_udfs
    import pyarrow as pa
ModuleNotFoundError: No module named 'pyarrow'

Spark spawns workers as separate processes and picks the interpreter from
PYSPARK_PYTHON, falling back to python3 on PATH — not the virtualenv
running pytest. All 7 bootstrap tests failed. tests/conftest.py now pins
PYSPARK_PYTHON and PYSPARK_DRIVER_PYTHON to sys.executable.

pandas was uncapped

PySpark warns "does not yet fully support pandas >= 3.0.0". Uncapped, the
newest interpreter resolved to the least-supported pandas: 3.13 pulled pandas
3.0.5 and emitted 14 warnings per run, 3.9 resolved to 2.3.3 and emitted none.
Capped to <3; the suite is now warning-free.

One SparkSession fixture, not two

Both test modules defined a session-scoped spark fixture calling
getOrCreate(). That is not two sessions but one session with two teardowns —
whichever module finished first called .stop() while the other might still
need it. It passed on collection ordering alone. Now a single fixture in
conftest.py.

CI

The dependency bugs share a shape: they only appear where nothing is
preinstalled. .github/workflows/test.yml installs from pyproject.toml
alone into a fresh uv venv on Python 3.9-3.13, then runs pytest. 3.9 resolves
to pyspark 4.0.4 and 3.10+ to 4.2.x, so both resolution paths are covered. A
ruff lint/format gate runs alongside, pinned to the dev extra's version.

Housekeeping

ruff check --fix and ruff format across the package; Homepage/Issues
corrected from allaei/replicas to hamed/replicas.

README.md and examples/*.ipynb are excluded from the formatter only.
Ruff rewrites Python inside README code blocks, discarding alignment chosen
for readability, and reformatting every notebook cell produced a 420-line diff
on a file that still has to be executed. Both remain linted.

Verification

12 passed, zero warnings, on Python 3.13 (pyspark 4.2.0) and on a
from-scratch Python 3.9 env (pyspark 4.0.4). The CI lint and test commands
were each run locally as written. The 5-leg matrix went green on GitHub
runners before the cleanup commits were added.

Still open, not in this PR

• examples/precision_recall.ipynb has 27 code cells and no outputs, so the
  bootstrap-band plot the README is built around does not render on GitHub.
• docs/ is empty.
• README says pip install replicas; the package is not on PyPI.

🤖 Generated with Claude Code