Add kripp_alpha Krippendorff's alpha (irr 0.85 kripp.alpha, READ source)

[seonghobae]

Summary

Iteration 69 of the autonomous paper-implementation loop: kripp_alpha — Krippendorff's alpha for inter-rater agreement, a Rust reimplementation of CRAN irr 0.85 kripp.alpha() (R/kripp.alpha.R — READ and normative; Krippendorff 1980 NOT READ, cited as method origin only).

Stacked on #309 (icc).

What

• Rust core kripp_alpha(ratings, nraters, nsubjects, method) -> Result<KrippResult, String> in crates/mlsirm-core/src/reliability.rs: coincidence matrix over unordered rater pairs per subject, with the irr divisor quirk preserved verbatim (mc = #nonmissing − 1 per column ONLY when the matrix contains any missing value, else 1 — complete-data alpha intentionally differs from the m−1 convention), diagonal increment 2/mc, mirror by assignment, nmatchval = total cell mass.
• All four distance metrics: nominal, ordinal (half-endpoint coincidence-row-sum weights), interval, ratio; alpha = 1 − (nmatchval−1)·Σ(utcm·δ²)/Σ(nc_c·nc_k·δ²).
• Fewer than 2 observed levels → alpha = 1 (R line 45).
• Documented deviations from R: all-missing matrix, infinities, and ratio level pairs summing to zero are explicit errors (R would return alpha=1, propagate, or emit Inf/NaN).
• PyO3 binding + thin Python wrapper kripp_alpha / KrippResult with the hardened input policy (masked-array reject, object-dtype outright reject, complex/bool reject, dtype-kind fiu).

Evidence chain

• Spec-verify (adversarial, BEFORE implementation): APPROVED-WITH-CHANGES; all 5 mandatory changes adopted (all-missing → Err labeled intentional divergence; ratio zero-sum unconditional error; explicit Python dtype policy; MU2 mutant pin 11/18; tolerance wording).
• Exact-Fraction oracle (EXECUTED, all-pass): K1 = irr man-page 4×12 nmm matrix — nominal 113/152, ordinal 108577/133160, interval 951/1120, ratio 18222619/22852465, nmv = 40, levels = 5. K2 = 3×5 complete-data no-NA quirk pin: nominal 43/72 (the m−1 mutant would give 11/18). K3 = hand-derived 2×4 fixture (independently re-derived by the spec reviewer). K4 = single-level alpha = 1.
• Mutation kill map: 6/6 mutants EXECUTED and KILLED (MU1 diagonal 1/mc, MU2 mc always m−1, MU3 ordinal full-weight sum, MU4 interval |δ| not δ², MU5 nmv off-diagonal only, MU6 num × nc product).
• MC-500: rater/subject permutation invariance across all four metrics (#[ignore], passes).
• Full suites: cargo -p mlsirm-core --lib 858 passed / 0 failed; pytest tests/test_paper_features.py 344 passed.

Test discipline

Every assert reads crate/wrapper outputs (no test-local recomputation identities); exact == pins where dyadic-exact, rel 1e-15 otherwise; error-contract tests assert specific messages.

Adversarial implementation review outcome

• Round 1 (FINDINGS(2)): (MAJOR) integer rating labels beyond 2^53 silently collapsed to the same float64 level — complete disagreement returned alpha = 1 with a single level; fixed in 0332c7f by rejecting integer arrays with values outside [-2^53, 2^53] (boundary itself accepted), with a red-green-verified regression test executing the reviewer's int64/uint64 repros. (MINOR) spec/implementation contract mismatch on object dtype — resolved by updating the spec to the icc final policy (object dtype rejected outright; per-element vetting retired as an unwinnable arms race).
• Round 2 (CLEAN): all round-1 repros now raise ValueError; ±2^53 boundary and in-range integers verified against crate output (2-level disagreement alpha = −0.5); empty dims, uint64::MAX, int64::MIN, huge float values (by-design accepted as exact f64s), and NaN paths probed; cargo ka_ 5 passed / 1 ignored; pytest TestKripp 5 passed; spec dtype-policy paragraph confirmed to match the implementation.

Full review log: session evidence file kripp_impl_review.md (rounds 1–2, executed repros and outputs).

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 0b511ef5-dfc0-46d6-8b56-2f498c0f51f9

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch seonghobae-kripp

---

_{Comment @coderabbitai help to get the list of available commands.}

[seonghobae]

Note: this PR was squash-merged into a non-main stacked base, so its commits did not land on main via this PR. The feature set reached main via #374 (stack tip integration) after #290.