Summary

• M-081 (done): promote a shared validate_n_bootstrap to diff_diff/utils.py (verbatim from ChangesInChanges' local validator: non-negative integer, numpy integers accepted, bool/None/float/negative rejected) and apply it at __init__ across the previously-unvalidated roster - CallawaySantAnna, SunAbraham, EfficientDiD, ImputationDiD, TwoStageDiD, WooldridgeDiD, ContinuousDiD, StaggeredTripleDifference, and the DiD family (DiD's __init__; MultiPeriodDiD/TwoWayFixedEffects inherit); CiC/QDiD re-point to the shared helper with a byte-identical message. 0 stays legal and still means bootstrap off on every > 0-gated analytical lane; no numeric defaults change. Named exception recorded in the ledger + design doc: HAD keeps its >= 1 floor (its n_bootstrap powers only the optional sup-t band, whose off-switch is fit(cband=False)).
• M-096 (done): the inference selector fails closed. Accepted set is exactly {"analytical", "wild_bootstrap"}, string-typed (an isinstance guard closes the one-element-ndarray hole in bare tuple membership), at __init__ and transactional set_params. At fit - placed after the survey and Conley front doors so their NotImplementedError rejections keep precedence - DiD with wild_bootstrap and no cluster= raises ValueError (previously a silent analytical fallback; test_did_wild_bootstrap_requires_cluster flips BY DESIGN, disclosed in CHANGELOG), and DiD/TWFE with n_bootstrap < 2 raise (n_bootstrap ∈ {0, 1} ran WCR with too few draws and returned a wild-labeled all-NaN inference tuple; the < 2 floor amends the originally-locked < 1, recorded in docs/v4-design.md §7 + the M-096 ledger notes in this diff). TWFE's unit auto-cluster stays; MPD's warn-and-analytical-fallback stays (n_bootstrap-independent).
• Fixed: DiD/TWFE never cleared per-fit bootstrap state, so a wild fit followed by set_params(inference="analytical") + refit reported stale inference_method="wild_bootstrap" + n_bootstrap/n_clusters/p_val_type metadata. Both now reset _bootstrap_results at the top of fit().
• Docs sweep: REGISTRY WCR fail-closed Note + MPD wild-availability corrections (SE-summary table + section bullet); troubleshooting.rst/choosing_estimator.rst examples gain the cluster= prerequisite (plus the adjacent weight_type= → bootstrap_weights= kwarg fix); README/llms.txt/llms-practitioner.txt normative wild recommendations qualified DiD-vs-TWFE and the two CS-adjacent advice strings reworded; llms-full.txt DiD block annotated; the wild-bootstrap cluster-count guidance harmonized to the single 50-cluster convention (choosing_estimator.rst was the <30 outlier). A systematic repo-wide wild-mention disposition sweep was run: normative recommendations qualified, capability/method prose untouched.
• Bookkeeping: docs/doc-deps.yaml gains troubleshooting.rst + llms-practitioner.txt under estimators.py (dependencies this diff proves); TODO row for the type-blind n_bootstrap holes in the already-validated estimators (HAD bool, dCDH bool+float, TROP float, SyntheticDiD float/jackknife); DEFERRED MPD anchor corrected; ContinuousDiD enrolled in the BAD_VALUES rollback lane (pre-existing catalog gap).

Methodology references (required if estimator / math changes)

• Method name(s): Wild Cluster Restricted (WCR) bootstrap - inference selection policy only; the WCR kernel, weighting, CR1 SEs, p-values, and test-inverted CIs are unchanged. Bit-identical results for every previously-legal, non-degenerate configuration.
• Paper / source link(s): Cameron, Gelbach & Miller (2008); Roodman, MacKinnon, Nielsen & Webb (2019) fwildclusterboot::boottest - see docs/methodology/REGISTRY.md "Wild cluster bootstrap (WCR)".
• Any intentional deviations from the source (and why): None to the procedure. The fail-closed selector contract (cluster required; n_bootstrap >= 2 under wild) is documented as a labeled Note in the REGISTRY WCR section; the < 2 floor amendment to the locked v4-design decision is recorded in docs/v4-design.md §7 and the M-096 ledger notes in this same diff.

Validation

• Tests added/updated: tests/test_v4_inference_policy.py (new, 126 tests - both rows' test_ref: validation sweep with set_params rollback atomicity, selector value-set incl. a non-string probe, cluster/floor coherence with full-message pins, boundary n=2 acceptance with a finite inference quintet, survey/Conley precedence at sub-floor counts on both estimators, MPD carve-out at n=0/1/999, refit metadata transitions, dynamic roster guard pinning inference exposure to exactly {DiD, MPD, TWFE}); tests/test_wild_bootstrap.py (the by-design flip); tests/test_estimators_vcov_type.py (cluster added to the wild/vcov coexistence pin); tests/test_base_estimator.py (ContinuousDiD BAD_VALUES enrollment).
• Backtest / simulation / notebook evidence (if applicable): N/A - no numerical changes; both tutorial notebooks using wild bootstrap already pass cluster=. Sphinx -W build + the four post-build HTML guards pass; test_doc_snippets.py executes the corrected RST examples.

Security / privacy

• Confirm no secrets/PII in this PR: Yes