Add fleiss_kappa: Fleiss' multi-rater kappa with exact (Conger) variant

[seonghobae]

Implements Fleiss' multi-rater kappa for nominal agreement, with the exact (Conger) chance-agreement variant ? iteration 67 of the autonomous paper-implementation loop. Stacked on #307.

Source basis

• READ and normative: CRAN irr 0.85, R/kappam.fleiss.R (full 113-line source) + man/kappam.fleiss.Rd. Algorithm source of truth; every formula verified line-by-line against it.
• NOT READ (cited as origins only): Fleiss (1971), Psychological Bulletin 76(5), 378-382; Conger (1980), Psychological Bulletin 88(2), 322-328.

What's in

• Rust core fleiss_kappa + FleissKappaResult in mlsirm_core::agreement: agreeP with the ?nr centering, classic chance ?p?, exact chance ?p? ? (1/nr)?s?? (sample variance over per-rater category proportions; algebraic reduction of R's sum(apply(rtab,2,var)*(nr-1)/nr)/(nr-1) verified), Fleiss' large-sample z/p (classic), category-wise kappas with NaN for empty categories (R's 0/0 preserved), listwise row drop for missing.
• PyO3 binding (plain name) + NumPy wrapper fleiss_kappa / FleissKappaResult dataclass in fast_mlsirm.validation (NaN = missing, k inferred or explicit, hardened input validation).
• Documented API deviations from R: index codes 0..k-1 (negative = missing, so negative numeric labels must be remapped), error on degenerate 1 ? chanceP = 0 (R returns NaN), size caps, always-computed classic detail, exact mode returns NaN z/p + empty detail (R returns neither).

Evidence

• Adversarial spec-verify: APPROVED (independent exact-rational recomputation of FK1/FK2/FK4 from the R source).
• Exact-Fraction oracle anchors FK1?FK5, executed: classic ? = 139/399 (var 181/10830, z = 2.6947398540854888), exact ? = 37/102 (?s?? = 3/50), missing-drop reproduces FK1 bitwise with subjects_used = 5, category ? = [1/21, 31/91, 43/63] (var? = 1/30), empty category NaN.
• 6 mutants EXECUTED and killed: MU1 agreeP drops ?nr; MU2 chanceP from row sums (1/5 ? 67/200 on the 5?4 fixture); MU3 exact ? classic; MU4 variance second-term sign; MU5 missing-as-category; MU6 pjk drops m?nr?p? centering.
• MC-500 #[ignore] permutation-invariance test (subject rows + rater columns, classic and exact, category detail), executed and passing. Disclosure: invariance cannot detect wrong-but-symmetric formulas; the FK value pins are the discriminating anchors. p-value pins use abs 5e-7 (crate erfc is the Numerical Recipes approximation, |err| < 1.2e-7).
• Suites: cargo mlsirm-core --lib 847 passed; pytest test_paper_features.py 336 passed.

Adversarial impl-review outcome

• Round 1: FINDINGS(2) -- MAJOR: uint64 > i64::MAX wrapped negative via astype(int64) and rows were silently dropped as missing; MINOR: explicit k lossy-coerced from 3.9/"3"/bool. Both fixed in 11bc1ed (unsigned-range guard before conversion; strict integer k check).
• Round 2: CLEAN -- fixes verified by execution (reproductions now raise ValueError; uint8/16/32/64 covered; k=None/int/np.integer paths intact; regression tests read wrapper raise paths).

[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: 5cc27061-5015-46e6-ab0d-a86a650a7ee3

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-fleiss

---

_{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.