0.0.9

• License change – crispyx 0.0.9 and later is distributed under a Modified
  MIT License, which adds two attribution conditions for commercial use. All MIT
  freedoms are retained and no fee or royalty is imposed. Versions up to and
  including 0.0.8 remain under the unmodified MIT License; that grant is
  perpetual and is not withdrawn. See LICENSE for the terms.

• Unified normalized effects – compute_normalized_effects /
  cx.pb.normalized_effects replaces the two earlier one-command estimators with a
  single function selected by method. method="mean_log1p" averages per-cell
  log1p values (mean of logs); method="log_mean" averages normalised counts and
  then applies log1p(baseline_count * mean) (log of mean). Both normalise library
  size themselves, and both return the effect in X with
  layers['perturbation_profile'], plus
  layers['control_profile_matched'] when batch_column is given, so that
  X == perturbation_profile - control_profile_matched exactly. Supplying
  batch_column is itself the request for batch correction; there is no flag.

  compute_average_log_expression and compute_pseudobulk_expression remain as
  deprecated aliases with their original layer and uns names, and now emit a
  DeprecationWarning. They will be removed in 0.1.0.

  Note that cx.pb.effects deliberately does not normalise: it computes a contrast
  on whatever scale its input already carries. Normalise beforehand with
  cx.pp.normalize_total_log1p, or use cx.pb.normalized_effects to have it done in
  one pass.

• Generic streaming batch statistics – batch_process /
  cx.tl.batch_process applies a user-supplied mergeable reducer within
  experimental batches without loading the complete cell-by-gene matrix. A
  BatchReducer provides initialize / update / finalize callbacks
  for per-group statistics, plus an optional compare callback for
  group-versus-reference contrasts in mode="comparison". Finalized batch
  statistics are combined as sum(weight * values) / sum(weight), and only
  batches containing both the group and the reference contribute to a contrast.
  Argument names follow the differential-expression API (groupby aliases
  perturbation_column; reference aliases control_label). Cached
  results are keyed on the input path and modification time, so regenerating a
  source file invalidates its cached statistic; force=True remains necessary
  when a reducer's implementation changes without changing statistic_name.

• Batch-level absolute pseudo-bulk profiles – aggregate_pseudobulk /
  cx.pb.aggregate groups by one or more observation columns and retains one
  profile for every observed combination. It supports strict raw-count sums,
  mean log1p expression, a five-cell default threshold, deterministic
  one-resample bootstrapping, source-layer selection, and versioned provenance
  metadata. perturbations keeps a profile when any of its grouping values
  matches, so it selects on whichever column holds the labels regardless of its
  position in groupby and preserves every combination of the others.

• Explicit pseudo-bulk effects – compute_pseudobulk_effects /
  cx.pb.effects consumes a saved crispyx pseudo-bulk artifact directly or
  aggregates cell-level input first. It returns within-batch target-minus-
  reference effects by default and can explicitly combine batches using the
  existing harmonic-count weighting.

• Tuple-level differential-expression results were intentionally not added;
  wilcoxon_test(batch_column=...) remains the batch-stratified test over all
  cells and batches.