v0.28.1

Added

• Score-based methods in causal-ts discover and discover_df: lges, tges, ges,
  lasso-var, granger, varlingam (returns a BaselineResult). run_grace and
  run_reap accept a result as candidate_graph.
• plot_lag_dependence / compute_lag_dependence: per-lag dependence of every pair from
  raw data, optionally parent-adjusted (adjust="parents") and marked against a
  true_graph.

Changed

• compare_graphs places its legend at "best"; lag-function grids get clearer axis labels
  and wrap_labels=.
• CI also runs the core tests on macOS and Windows and reports coverage.

Fixed

• With statsmodels 0.15, CausalForecaster's fallback model fits its AR candidates again; a removed
  AutoReg argument made them fail silently, so only exponential smoothing was chosen.
  CausalForecaster.fit now emits a RuntimeWarning when every AR candidate fails to fit.
• CausalForecaster.predict keeps each prediction on its own row after a missing value
  (later predictions were shifted one row).
• HMMRegimeDetector stays within its switch budget; LeveneRegimeDetector raises on an
  unknown method.
• Plotting: plot_time_series_graph(order="auto") handles lag-0-only and 4-D graphs,
  node_layout="random" works, and setup_matrix(lag_array=...) works with the default
  x_base.
• Cedar with include_C=True keeps lagged edges whose strongest lag is 0, and MCI pruning
  no longer conditions on the edge's own target (lag-0 pairs can now be o-o).
• CPU parcorr gives p = 1 for a constant column instead of NaN (#111).
• The CPU cmiknn cache key includes every setting that changes the p-value.
• Clear CI-test caches (cache_dir) written before this release: they may hold the NaN
  parcorr results or cmiknn results under the wrong settings.
• discover --json / deconfound --json print only the JSON to stdout.
• extract_subgraph reads lag-0 <-- / <-o marks as reversed edges;
  tigramite_graph_to_binary reads bytes mark graphs.
• compare_graphs, plot_lag_metrics and plot_pvalue_distribution no longer count <->
  as an edge and reject a ground truth with any mark other than --> / <--.
• No statsmodels adfuller FutureWarning from inspect; no blank lines in CI-test caches
  on Windows; __version__ is 0+unknown (not 0.14.0-dev) in an uninstalled checkout.