replicas 0.2.0

Second alpha. Three breaking changes — read the first section before upgrading.

Changed

• Breaking. box_plot and plot_pr take a pandas, Polars, or Spark DataFrame, like the rest of the package. The plot extra declares pandas and is sufficient on top of any backend: a Polars frame is collected without PyArrow when PyArrow is absent, since replicas[polars,plot] pulls none and polars.DataFrame.to_pandas goes through Arrow. box_plot was pandas-only and plot_pr was Spark-only, so the two plot helpers demanded opposite backends and a pandas user could not call plot_pr at all. Each one now reduces the data on its own backend and collects only the result.
• Breaking. The spark extra requires PySpark 3.5, up from 3.3. pr_band uses F.percentile, added in 3.5. percentile_approx returns a different order statistic — on 100 replicas it put the 5th-percentile band edge 2.4% off the pandas and Polars value, which is a visible shift in a plotted band, not a rounding difference. Spark 3.3 and 3.4 are both past end of life.
• Breaking. confusion_table, calculate_pr, and at take by instead of group_by. Every public function that partitions a computation by columns now uses the same name, matching sample and bootstrap and pandas' own groupby(by=...). Positional calls are unaffected; a group_by= keyword call raises TypeError. The type alias replicas.metrics.GroupBy is now ByColumns.
• The replica column is a 32-bit integer on all three backends. pandas previously produced int64 and Polars Int64.
• Every public function documents its arguments, its return value, and the errors a caller can act on. replicas.plotting uses the same docstring convention as the rest of the package and names seaborn.FacetGrid as what both helpers return. sample documents the round(group_size * fraction) draw size, including fractions above 1, and both samplers document that the uniqueness order_by needs for a reproducible seeded Spark draw is the caller's responsibility and is not verified.
• at documents that its "lowest qualifying threshold" rule suits a metric that does not increase as the threshold falls, such as precision. A recall target degenerates to the group's minimum threshold.
• confusion_table documents that the non-null and mutually-exclusive conditions on its indicator columns are the caller's responsibility.
• The metric functions and bootstrap document the Spark cost of an empty by.
• The README defines "replica" once at first use, and states the limitation of a percentile-bootstrap band as asymptotic coverage with a finite-sample shortfall, instead of claiming the true interval is always the wider one. bootstrap points at the README for the NaN exception to the cross-backend parity guarantee.

Added

• pr_band(df, by=..., ci=..., recall_round=...), the reduction behind plot_pr, as a public metric function on all three backends. It returns the original curve and a pointwise quantile band across the replicas. A null grouping value is one group, as it is everywhere else in the module: the three backends reach that through one grouped aggregation rather than a join, because they disagree on whether a join matches null keys.
• A notebook extra and a CI job that executes examples/quickstart.ipynb and compares its outputs with the committed ones.

Fixed

• bootstrap restores the Spark checkpoint directory the caller had configured instead of leaving its own in place for the rest of the session. Spark offers no per-call checkpoint path, so the read/set/checkpoint/restore sequence runs under a process-wide lock: two concurrent bootstrap calls can otherwise checkpoint into each other's directory. A restore that fails raises a RuntimeWarning rather than passing silently. Known limitation: Spark reports a checkpoint directory as <root>/<generated-id> and setting that value back appends another id, so a directory configured outside replicas ends up one generated level deeper the first time it is restored. Later calls reuse the remembered root and the depth holds.
• The local checkpoint fallback is scoped to the current user rather than a fixed /tmp/replicas/ that the first user on a machine takes ownership of.
• box_plot explains that it needs a replica column instead of raising a bare pandas KeyError.
• plot_pr rejects a ci outside (0, 1] up front, instead of failing later inside percentile_approx.

Full changelog: v0.1.0...v0.2.0