feat(smart): Sprint-B coverage in assumption_audit + pub_ready + fail-loud tests

Two Smart-tool gaps for the Sprint-B causal modules, closed after a
fresh-eyes code review of this session's work.

1. sp.pub_ready — single-result acceptance + array/Series guard
   Pre-fix, sp.pub_ready(my_causal_result) crashed with
   "'CausalResult' object is not iterable". Now auto-wraps bare
   scalars; numpy.ndarray and pandas.Series pass through the iterable
   guard (so pub_ready(np.array([r1, r2])) iterates elements rather
   than wrapping the whole array). The truthy check was replaced by
   an explicit len() to avoid the "ambiguous truth value" error on
   ndarray.

2. sp.assumption_audit — Sprint-B method-specific audits
   Detection now reads ``.method`` + ``model_info['model_type']`` +
   ``model_info['estimator']``; plus class-name check for
   PrincipalStratResult. Seven new _audit_* functions fire method-
   specific assumption checks:
     - proximal: bridge form, proxy order condition, first-stage F
     - msm: sw_mean drift, positivity (sw_max), exposure target,
       sequential exchangeability reminder
     - principal_strat: monotonicity violation fraction, stratum
       proportions, principal ignorability (only under the
       principal_score method)
     - g_computation: outcome-model correctness reminder, bootstrap
       health
     - front_door: mediator family, no-direct-path reminder, no
       unmeasured M-Y confounder reminder
     - mediate_interventional: decomposition identity IIE+IDE=Total
       with explicit fallback when keys are missing (no vacuous pass),
       treatment-induced-confounder listing
     - mediate (natural): cross-world independence reminder +
       bootstrap failure-rate tracking

   BUG FIX — keyword-substring leak: pre-fix, naive substring match
   triggered Panel ('fe'/'re') and IV ('iv'/'2sls') checks inside
   "Proximal Causal Inference (linear 2SLS)" because the short tokens
   appear as substrings of "inference" and "2sls". Now uses word-
   boundary regex + Sprint-B-first dispatch, so classical branches
   only run when NO Sprint-B branch matched.

   BUG FIX — dispatch collision: method strings containing BOTH
   'interventional mediation' and 'causal mediation' would double-fire
   both branches. Natural-mediation branch is now guarded by
   ``not sprint_b_matched``.

3. Test coverage
   - tests/test_smart_tools_sprint_b.py (15 tests): pub_ready scalar
     equivalence, ndarray support, None/empty; assumption_audit for
     each Sprint-B method with explicit forbidden-names leak check;
     review-round-2 regression tests (no vacuous pass, no double
     firing, bootstrap failure-rate check, ndarray auto-wrap).
   - tests/test_diagnose_batteries_sprint_b.py:
     - strengthened test_mediate_wald_pvalue to use scipy.stats.norm.sf
       as independent ground truth plus a null-effect cross-check
       (catches sign errors the self-circular version missed);
     - new test_battery_proximal_order_condition_check_directly
       constructs a hand-crafted CausalResult with k_z<k_w so the
       previously-dead battery code path is actually tested.

Regression: 180/180 Sprint-B tests pass; 50/50 broader smart-workflow
regression tests pass.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: brycewang-stanford