Add LUCID to the CLI, inspect, skill, notebook and llms.txt

[fesanghary]

Summary

LUCID was reachable only from Python. Adds the four surfaces every other algorithm
already has.

• CLI — causal-ts deconfound GRAPH --data CSV [--strategy adaptive|tetrad|pds],
  post-processing a graph saved by discover. Not registered in discover --algorithm:
  LUCID corrects a graph, it doesn't discover one.
• inspect — facts["latent_factor"] {detected, spectral_ratio, tau}, taken from
  LUCID's own router so the two can't drift. recommend_config suggests .deconfound()
  only when detected is True, and never changes the algorithm/CI-test choice.
• skill — SKILL.md surfaces latent_factor, plus a "many contemporaneous edges →
  deconfound" step in the interpretation playbook.
• notebook — new section 5 (5–9 renumbered). CDNOTS reports 23 spurious lag-0 edges;
  .deconfound() removes all 23 and keeps every true edge (F1 0.20 → 0.86). One lag-1
  false positive survives, which the notebook points out rather than hides.
• llms.txt — adds LUCID.

detected is tri-state on purpose: False means no factor structure this test can see
— not "no confounding", since a 2–3 variable confounder leaves no dominant
eigenvalue and stays invisible. None means the check couldn't run (NaN, d<2,
T<d+3), kept distinct from False so "didn't check" is never read as "checked and
found nothing".

Drive-by fixes

Both were blocking or actively wrong, and are in files this PR already touches:

• api_reference.ipynb wouldn't execute: run_cdnots_gated unpacked as a 3-tuple, the
  p-value demo needed return_pvals=True, and a wrap_graph table predated
  GraceResult inheriting the DoWhy methods.
• llms.txt's GRACE snippets unpacked run_cdnots_gated / run_stability_selection
  as 3-tuples — both return a single GraceResult, so the documented code raised
  TypeError for anyone following it.
• Two LucidResult docstrings conflated the router's "pervasive" label with any
  detected factor: R > tau selects sf or pervasive (only regime says which),
  and n_factors counts eigenvalues above the MP bulk regardless of regime
  (verified: sf → 1, pervasive → 2, sparse → 0).

Test plan

• 315/315 tests — 7 new, covering detection on pervasive vs full-rank data, all
  three guard paths (returning None, not False), nudge-only-on-positive, and
  that algorithm choice is unaffected
• api_reference.ipynb executes end to end: 0 error cells
• CLI exercised end to end (generate → discover → deconfound) for all three
  strategies plus --json and the ignored-flag warning
• python-lint job: black(24 pin) / isort / flake8, exact CI invocation
• super-linter: markdownlint clean on SKILL.md, codespell clean on all
  changed files; no YAML/shell/workflow files touched
• sphinx-build warning set identical to baseline

[github-actions]

⚠️ Deprecation Warning: The deny-licenses option is deprecated for possible removal in the next major release. For more information, see issue 997.

Dependency Review

✅ No vulnerabilities or license issues or OpenSSF Scorecard issues found.

Scanned Files

None