v0.1.0

First release. The metrics were extracted from the ISP preprocessing benchmark
(task_isp_preprocessing, itself derived from the OpenProblems
task_ist_preprocessing) and from the scverse hackathon sprot prototypes, and
the extraction is verified rather than asserted: every one of the 336 metric
values the benchmark produced across six families and four exemplar-001 runs is
reproduced bit for bit by scripts/check_against_benchmark.py.

Added

• A metric registry. Each metric declares what it measures, its range, which
  pipeline stage it scores, what it needs beyond the sample, and whether a
  larger value is better. list_metrics() returns it as a DataFrame and
  to_openproblems_info() emits the info.metrics[] block the OpenProblems
  schema requires. Direction previously lived only in a hand-maintained
  spreadsheet, so nothing could read it at runtime.
• Six metric families as single calls: basic_statistics,
  marker_distribution, normalization_diagnostics, covariate_dependence,
  segmentation_quality and signal_background, plus cluster_separability
  and segmentation_agreement.
• Per-quantity functions underneath each family, vector-valued where the
  quantity is per cell or per marker -- so marker_coverage says which marker
  is dragging the panel down, not only that the panel is down.
• CellPixelStats: one pass over an image, reused by every pixel-level metric.
  Its .subset() restricts the cells while holding thresholds, ring statistics
  and background curves fixed, which is what makes a QC-filtered value
  comparable with an unfiltered one.
• _qc_pass variants on every family via qc_variants=, and arbitrary cell
  selections via subset=.
• Metrics that were designed but never built: marker_bimodality (including the
  published bimodality index), tile_cv, nuclear_signal_containment,
  signal_drop_on_aggregation, clustering_stability, nuclear_enrichment and
  localisation_agreement.
• sprotmetrics.sd, thin SpatialData wrappers, under the spatialdata extra.
• A record of metrics this package deliberately does not compute, each with its
  reason and replacement, queryable through
  list_metrics(include_deprecated=True).

Changed from the benchmark implementations

• One Otsu helper replaces two that disagreed on which scale they returned.
  Positivity is compared in log space, avoiding an expm1 round trip that can
  move a value sitting exactly on the cut.
• object_f1 returns 0.0 when nothing matches, rather than NaN. The script it
  came from returned 0.0 for one no-match case and NaN for the other, which made
  an over-expanded segmentation look unmeasurable instead of bad.
• nuclear_signal_containment is marked weak evidence: growing every mask raises
  it, so it must be read against spillover_ratio.
• Missing QC annotation now warns instead of silently producing NaN columns.
• A missing layer raises and names the layers that do exist, instead of quietly
  falling back to .X.

Known issues

• median_intracell_coverage reads exactly 1.0 on real data. It averages the
  most-covered 5% of cells, which are above their own Otsu threshold everywhere,
  so it saturates by construction. Registered experimental pending redesign.
• median_cell_marker_rank_correlation is exactly 1.0 whenever normalization is
  monotone per cell, which all the benchmarked methods are. Kept for
  compatibility and registered as descriptive, so it cannot be ranked on.