v0.26.0

Added

• causal-ts-discovery agent skill (Claude Code / OpenAI Codex) shipped in the package for model-driven causal discovery
• causal-ts inspect <file> — data-health facts plus an algorithm / CI-test / C-preset recommendation and a cost class (JSON)
• causal-ts discover --json — echoes the run summary, a named edge list, and a diagnostics block (density, self-loops, contemporaneous/lagged counts, hub in-degree, empty / saturated flags) to stdout
• causal-ts discover --validate (with --n-bootstrap / --window-frac) — temporal-bootstrap edge-stability check
• causal-ts discover --pvalues — opt-in p-value matrix in the output (off by default as a high-dimensional memory safeguard)
• causal-ts install-skill — install the skill into ~/.claude/skills and ~/.agents/skills
• .claude-plugin/ manifests — the skill can also be added as a Claude Code plugin from a checkout
• include_C and c_preset for GRACE — run_cdnots_gated, run_stability_selection, run_ci_skeleton, and causal-ts discover --algorithm grace/grace-ss now let you configure the C node of the skeleton GRACE refines, rather than always building one with the linear basis. include_C defaults to True (the previous behaviour); multi-column presets such as linear+sin are supported end to end.
• Public Python entry points inspect_df, recommend_config, discover_df
• Observed-data (non)stationarity detection (ADF + KPSS) and trend-form detection for C-preset selection
• Multi-format data reader (csv / parquet / feather) with an optional parquet extra
• guided_discovery example notebook — the inspect → recommend → discover → interpret workflow end to end
• Causal feature selection — select_features for O(d) single-target discovery (causalts.feature_selection)
• Correlation / association plots — corrplot and compute_association_matrix (causalts.plotting)
• Temporal-subsampling detection — detect_subsampling and DetectionResult (causalts.utils)

Fixed

• causal-ts discover --algorithm cdnots+ was an accepted choice with no dispatch branch — it produced no graph and exited successfully. CDNOTS+ now runs (and honours --impute).
• --c-preset was never forwarded to run_cdnots, so every CDNOTS run silently used the default linear basis regardless of the flag.
• --no-c had no effect on GRACE, whose skeleton always included a C node.
• discover_df silently dropped alpha for CEDAR and dropped alpha, include_C, c_preset, and ci_test for GRACE. Everything it can honour is now forwarded, and anything it cannot raises instead of being ignored. alpha now defaults to None, meaning "use the algorithm's own default", so CEDAR keeps its 0.01 thresholds.
• discover --validate ran its bootstrap windows without the main run's --impute settings, so persistence was measured under a different configuration than the graph it annotated.
• GRACE high-dimensional memory usage in run_cdnots_gated
• Documentation: the README CI-test table listed linsig, which the package does not ship — the path-space test is sigkci
• Documentation: the GRACE examples in the algorithms guide unpacked a tuple, but run_cdnots_gated / run_stability_selection return a GraceResult; copying them raised TypeError
• Documentation: absolute-value notation in two docstrings was parsed as an RST substitution reference, producing errors in the docs build

See CHANGELOG.md for the full history.