Fix all findings from cross-model review (bugs + consistency + defense-in-depth)

[kimmingul]

Summary

Resolves every finding from the three-reviewer audit (Claude code-reviewer + Claude security-reviewer + Codex 2nd-opinion) plus a follow-up independent review pass that caught a HIGH that the first wave missed. 18 source files modified, +341/-52 lines.

Critical bugs

Area	Was	Now
report --kind power-curve	silent NaN curve for any two-sample method	derives sweep key from registry signature; real curve
report --kind r-code / sas-code	TypeError on CI methods (achieved_power=None)	emits N/A cleanly
report --kind protocol / grant	KeyError/crash; literal "None"/"α" placeholders for CI methods	localized "n/a" / "해당 없음" fallback for power/alpha/target_power
report --kind sensitivity	wrong solve mode when source audit was solve-for-power	drops n/n1/n2/power from base before sweeping

High — decision_tree.yaml no-fallback rule (per CLAUDE.md)

The prior fix that replaced logrank_kgroup with logrank_freedman violated the explicit "No fallback heuristics" rule in CLAUDE.md by silently downgrading k-group survival to a 2-group calculator. The same class of issue affected the count and ordinal branches (no group-count question; any group count silently routed to a single specific method).

Introduces an unimplemented: terminal marker that samplesize doctor accepts as valid. K-group survival, k-group Poisson, and one-sample ordinal now terminate with an explicit reason instead of routing to an approximate method. count and ordinal gained the missing group-count question; count → two routes to the (already-validated) tests_two_poisson_means; ordinal → three_or_more routes to kruskal_wallis_simulation.

High — Plugin defense-in-depth

samplesize-calculate and samplesize-validate SKILLs now ship explicit "shell-injection defense" preambles instructing the LLM to:

• registry-validate method_id against ^[a-z][a-z0-9_]*$
• prefer --json-args-file <tmpfile> (new CLI flag) over inline --json-args <json>
• never pass raw user text as a shell argument

Backed at the CLI by: --json-args-file, kwarg whitelist against the resolved methods signature, and _resolve_callable module-prefix allowlist (samplesize.tests.).

New samplesize doctor checks (11/11; was 9/9)

• registry.decision-tree-leaves-exist — every leaf: in decision_tree.yaml is a registered method id (or a deliberate unimplemented: marker).
• plugin.skill-kind-choices-exist — every --kind <name> referenced in plugin/*.md is in cmd_reports argparse choices.
• The existing plugin.skill-cli-flags-exist now also scans docs/**/*.md and README.md.

Other fixes

• reference_intervals_clinical_lab: linear N-search → binary search (5.5s → 0.004s worst case).
• audit.py: microsecond suffix + method_id filename sanitization.
• gen_method_coverage.py --check: substring → word-boundary regex.
• grant_aims now honors --lang.
• scipy>=1.17 floor (Anderson-Darling needs method="interpolate").
• randomized_block_anova / reference_intervals_clinical_lab inputs_echo adds solve_for.
• .github/workflows/release.yml: new verify job (doctor + coverage --check + pytest + ruff) gates build/publish so a v* tag on an unchecked commit cannot ship.

Verification

• Full pytest: 1303 passed, 232 skipped, 0 failed.
• samplesize doctor: 11/11.
• gen_method_coverage.py --check: up to date (234 methods, 819 fixtures).
• ruff check samplesize/ tests/ scripts/: clean.
• Independent code-review pass (oh-my-claudecode:code-reviewer, opus): all HIGH/MED findings from the first pass either addressed or surfaced as trade-offs.
• End-to-end repros confirmed: CI-method protocol text now reads "actual power of n/a" (not crash, not "0.0000", not literal "None"); two-sample power-curve now produces a real curve.

Process notes

• 4 parallel executor agents (opus) for the initial fix wave with non-overlapping file scopes.
• 1 follow-up executor (opus) for the reviewer-surfaced issues (the rec["result"] audit shape + None handling for protocol templates + unimplemented markers + audit docstring + doctor regex comment).
• Separate code-reviewer (opus) pass between the two waves (writer/reviewer separation per OMC).

🤖 Generated with Claude Code