v0.27.0

Added

• LUCID (causalts.confounders) — regime-adaptive deconfounding for causal discovery
  under latent confounders. run_lucid(df, max_lag) diagnoses whether latent confounding
  is sparse or pervasive from the residual spectrum (against a Marchenko–Pastur no-factor
  null) and applies the matching correction, returning a LucidResult. Every
  CausalResult now also exposes .deconfound(), .tetrad_filter(), and .pds_filter()
  so LUCID (or a fixed-strategy comparator) can be applied to a graph already discovered
  with any algorithm, without re-running the skeleton search. See the new
  Unobserved Confounders (LUCID) tutorial and the
  causalts.confounders API page.
• DoWhy effects bridge. validate_transition_graph() and history_sufficiency()
  validate a lag-embedded graph (replacing the deprecated falsify_graph());
  refute_effect() and sensitivity_analysis() stress-test an estimate;
  build_identification_artifacts() / build_transition_artifacts() hand back a graph
  and the frame it belongs with, for driving DoWhy directly. All are on the result
  objects, and causal-ts dowhy validate --test exposes the validators from the CLI.
• corrplot(..., diag="glyph") — renders the diagonal as an ordinary cell,
  using the same method and colormap as the rest of the matrix. Intended for
  directed matrices (a cause→effect adjacency or an edge-stability matrix),
  where the diagonal is real data such as a self-loop rather than the trivial
  1.0 of a correlation matrix. Significance markers and confidence-interval
  overlays are still skipped on the diagonal.

Fixed

• Breaking. estimate_effect() reported an unadjusted regression as an identified
  causal effect — a lagged treatment had no parents, so no backdoor path was found. The
  graph is now unrolled and the adjustment set verified against the fitted estimator.
  ATEs change.
• The GES, LGES and TGES baselines were substantially understated. ges_discovery
  read causal-learn's adjacency matrix with the endpoints transposed, and the vendored
  GES search behind lges_discovery / tges_discovery had a broken CPDAG construction
  that stopped LGES far short of convergence. Both are fixed and now apply temporal
  background knowledge; F1 on baseline_comparison moves from 0.125–0.522 to 0.636–0.949.
  Any prior comparison against these baselines understates them.
• refute_effect() no longer reports an uncomputable refutation as a failure, and
  refute_structure() returns p_value and adjusted_p_value separately. DoWhy's
  renamed identification APIs are bound by capability, not version.
• causal-ts ci-test-info --test <name> now shows only the selected test's
  summary instead of the full guide. Registered tests without a guide section
  report a clear error; the default and --test all output are unchanged.
• The CI test selection guide now covers all registered tests. parcorr,
  cmiknn, cmiknn-mixed-gpu, fisherz, chisq and gsquared had no entry,
  so ci-test-info --test <name> failed for them. The cmiknn-gpu entry also
  now notes that its permutation null stops early.
• priority=3 and priority=4 (collider strength ordering) now raise ValueError
  at the entry point, instead of failing with an AttributeError from inside
  causal-learn after the skeleton search. They score each conflict over the full
  powerset of the endpoints' neighbours — exponential in node degree — and were
  never wired to a CI test. Use priority=1 (abstain) or 2 (keep first).
• detect_subsampling raises ValueError on a 1-D or single-variable input, instead of
  failing inside NumPy with LinAlgError: 0-dimensional array given.
• The generator in the Unobserved Confounders (LUCID) tutorial injected its true lag-1
  edges after the recursion instead of inside it, so the simulated data did not match the
  ground truth the notebook scored against.
• corrplot dropped the right and bottom edges of its grid border. All axes
  spines are hidden, and the border was drawn with axhline/axvline at
  exactly the axis limits, so half of each boundary line fell outside the clip
  box. The border is now an unclipped rectangle and all four edges render.
• corrplot(..., colorbar=False) was ignored for the colour-only glyph methods
  ("color", "shade"), which silently overrode an explicit argument and
  forced callers to use cl_pos="n" instead. colorbar=False now suppresses
  the colorbar for every method.
• Documentation: the "New in v0.26" banner linked to the example notebook with a
  path relative to the site root, but the banner renders on every page — from
  anything below the root (examples/, api/, getting_started/) it resolved to
  a nonexistent nested path and 404'd. The link is now resolved per page.

Changed

• run_cdnots_plus defaults changed — this can change results for existing
  callers that don't pin these parameters explicitly. run_cdnots_plus now
  follows PCMCI+'s conventions more closely: colliders that conflict during
  orientation are abstained on rather than tie-broken (priority=1, was 2),
  MCI conditioning excludes only the exact tested lag rather than every lag of
  a variable (legacy_mci_conds=False), the nonstationarity sink search stops
  when candidates aren't clearly separated rather than always committing
  (orient_margin=0.1, was unconditional), and the default significance level
  is alpha=0.01 (was 0.05), matching PCMCI+ and empirically better for
  CDNOTS+ (plain run_cdnots is unaffected and keeps alpha=0.05). In
  aggregate these changes track PCMCI+ much more closely on stationary and
  latently-confounded data while widening CDNOTS+'s advantage over plain
  CDNOTS where nonstationarity is real. Pin the old values explicitly
  (priority=2, legacy_mci_conds=True, orient_margin=0.0, alpha=0.05) to
  reproduce prior behavior.

• causalts.ci_tests.SigKCIGPU no longer uses the optional sigkernel
  package: it crashed with a buffer dtype mismatch whenever sigkernel was
  installed, and measured slower than the existing pure-torch fallback at the
  path lengths this test uses. SigKCIGPU now always uses that fallback.

• The guided_discovery example notebook is now agentic_discovery ("Agentic
  Causal Discovery"), naming it after the causal-ts-discovery agent skill whose
  workflow it walks through. The old examples/guided_discovery.html URL is gone;
  the page is at examples/agentic_discovery.html.