Add Livingston-Lewis (1995) classification accuracy and consistency

[seonghobae]

Implements Livingston & Lewis (1995) single-administration classification accuracy and consistency, as implemented in CRAN betafunctions 1.9.0 LL.CA (read line by line; the original article was not consulted directly). Stacked on #227.

What

• mlsirm_core::classification::livingston_lewis: effective test length -> true-score raw moments (factorial-moment identity, unrounded ETL scale) -> Hanson (1991, as cited in Haakstad, 2022) four-parameter beta moment fit with two-parameter [0,1] fail-safe -> binomial observed-score model with N = round_ties_even(ETL) -> accuracy cells (tp/fp/tf/ff), sensitivity/specificity, consistency cells, Cohen's kappa.
• Singularity-safe composite Gauss-Legendre quadrature: power substitution when a beta shape < 1, endpoint-graded panels otherwise (a plain fixed rule loses ~2e-5 mass on high-shape cases; caught during development).
• PyO3 binding + thin Python wrapper fast_mlsirm.livingston_lewis returning a dataclass.

Documented divergences from the oracle

• Single k = round_ties_even(N*c) threshold in BOTH accuracy and consistency blocks. betafunctions mixes round (accuracy) with floor (consistency), making its consistency cells asymmetric (measured p.ij-p.ji = 0.0345 on a test fixture); here p_ij == p_ji by construction.
• Pass = observed score >= cut is the positive class (the oracle labels fail as positive; its sensitivity is our specificity).
• Fail-safe also engages on numerically invalid 4P fits (oracle only checks out-of-bounds support).
• Conditional ratios (sensitivity/specificity/kappa) are explicit NaN when their margin or chance denominator vanishes (found by adversarial impl-review; oracle propagates NaN silently).

Evidence

Check	Result
Spec-verify (adversarial, pre-implementation)	GO-WITH-FIXES; all 5 fixes applied (singular-safe quadrature, asymmetry wording, failsafe divergence, independent-replication literals, non-integer-ETL fixture)
Independent replication	scipy.integrate.quad pipeline, 4 fixtures (2P fail-safe, 2P skewed, genuine 4P path, extreme-cut round/floor discriminator), literals pinned at 1e-7/1e-9
Mutation spot-checks	5/5 killed: integrand swap, unrounded-ETL k (required the extreme-cut fixture; base fixtures could not discriminate), k-1 -> k shift, ddof n-1 -> n, fail-safe drop
cargo test -p mlsirm-core	498 passed, 0 failed
500-rep Monte Carlo (#[ignore])	consistency recovers empirical two-form agreement
Python tests	2 passed (pinned literals + trust-boundary rejections)
Adversarial impl-review	2 confirmed NaN defects found -> fixed with explicit ratio guard + documented semantics + regression test

Known limitation (disclosed in-test): at exactly-zero denominators IEEE 0/0 already yields NaN, so a dropped ratio guard is not killed by the NaN test; the guard's value is for tiny nonzero denominators.

[Copilot]

Pull request overview

Adds Livingston–Lewis (1995) single-administration classification accuracy/consistency to the Rust core and exposes it through the PyO3 extension and Python API, with pinned-literal tests validating the numerical pipeline against an independent SciPy quadrature replication.

Changes:

• Implement mlsirm_core::classification::livingston_lewis, including effective test length → beta moment fit → binomial observed-score integrals (accuracy/consistency/kappa).
• Expose the new functionality via PyO3 (fast_mlsirm._core) and a Python wrapper returning a LivingstonLewisResult dataclass.
• Add Rust and Python test fixtures covering 2P fail-safe, 4P path, threshold rounding behavior, and trust-boundary input validation; update the changelog.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
crates/mlsirm-core/src/classification.rs	Core Livingston–Lewis implementation, quadrature/integral machinery, and result struct.
crates/mlsirm-core/src/reliability.rs	Makes inc_beta pub(crate) so classification can reuse it for binomial tail computation.
crates/fast-mlsirm-py/src/lib.rs	Adds PyO3 binding that returns a dict of Livingston–Lewis outputs.
python/fast_mlsirm/classification.py	Adds LivingstonLewisResult dataclass + livingston_lewis() wrapper calling the compiled core.
python/fast_mlsirm/__init__.py	Re-exports livingston_lewis at the package top level.
tests/unit/classification_tests.rs	Adds Rust-side fixtures and invariants for Livingston–Lewis outputs.
tests/test_paper_features.py	Adds Python-side pinned-literal and rejection tests for the wrapper/core boundary.
CHANGELOG.md	Documents the newly added Livingston–Lewis feature and its documented divergences from the oracle.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[seonghobae]

@copilot Fix the code for all comments in this review thread.

When a review comment includes a suggested change, apply the suggestion exactly.

Do not make changes beyond what is described in the linked review thread.

[Copilot]

@copilot Fix the code for all comments in this review thread.

When a review comment includes a suggested ch...

All four review items addressed in commit Fix review comments: rename p_tf/p_ff→p_tn/p_fn, cache GL64, explicit docstring, standardize citation year:

1. Field naming (p_tf/p_ff → p_tn/p_fn): Renamed to unambiguous p_tn (true negative: P(true fail, observed fail)) and p_fn (false negative: P(true pass, observed fail)) in the Rust struct, PyO3 binding, Python dataclass, and all Rust/Python tests. Derived field formulas (accuracy, sensitivity, specificity) updated accordingly.
2. GL64 caching: Added static GL64: OnceLock<(Vec<f64>, Vec<f64>)> with an inline fn gl64() accessor. beta_weighted_integral now uses the cached rule instead of recomputing Newton iterations on every call.
3. Docstring: LivingstonLewisResult now lists each accuracy cell with its full definition (true positive, false positive, true negative, false negative), the four-cell sum invariant, and derived-field formulas.
4. Citation year: Standardized to Haakstad (2022) in classification.rs (was 2023), matching the Python docstring and CHANGELOG.

[Copilot]

Pull request overview

Copilot reviewed 8 out of 8 changed files in this pull request and generated 3 comments.