Add Luce Spectral Ranking (LSR / I-LSR) paired-comparison estimators

[seonghobae]

Implements Luce Spectral Ranking (LSR) and iterative LSR (I-LSR) for pairwise comparison data, following choix 0.4.1's lsr.py dense path (source READ; Maystre & Grossglauser, 2015, Fast and Accurate Inference of Plackett-Luce Models, NIPS — NOT READ, cited as described by choix).

Stacked on #291 (seonghobae-bradley-terry) — review only the last commit.

What

• Rust core mlsirm_core::scaling::{lsr_pairwise, ilsr_pairwise} + LsrResult {params, weights, iterations}: LSR Markov chain (rate c/(w_i+w_j) on each loser->winner edge, alpha everywhere as regularizer), stationary distribution via Gaussian elimination with partial pivoting (sum-constraint row replaces the redundant stationarity equation), scaled to sum n; params = centered logs.
• Post-solve guards (spec-review mandate): positivity/finiteness, sum ~ n, and pi . chain ~ 0 residual check — a custom elimination must prove stationarity, not just return something positive. Disconnected graphs at alpha=0 and overflow from huge counts/alpha raise errors instead of NaN (choix can emit NaN there).
• I-LSR: choix NormOfDifferenceTest semantics (L1 <= tol*n, first check never fires, default max_iter=100). At alpha=0 the fixed point IS the Bradley-Terry MLE — verified against bradley_terry_mm to 4e-19 on the oracle fixtures (cross-algorithm anchor). For alpha>0 the two regularization paths deliberately differ (chain-rate vs Dirichlet-MAP; both follow their sources, documented).
• PyO3 bindings (plain names) + thin Python wrappers fast_mlsirm.lsr_pairwise / ilsr_pairwise with full input validation before casts.

Spec-verify (BEFORE implementation)

Adversarial review of hand-derived spec + EXECUTED oracle: APPROVED-WITH-CHANGES; all 4 mandatory changes implemented (overflow tests for huge counts/alpha, statdist residual guard, alpha-diagonal equivalence documented, no-initial_params exposure documented).

Oracle (EXECUTED)

Exact-Fraction one-shot statdist + mpmath 50-digit I-LSR fixed points, cross-checked against pip choix 0.4.1 (<= 2.2e-13) and the shipped BT oracle (<= 4e-19). Exact pins e.g. fixture A statdist = [12/17, 45/34, 33/34]; iteration pins A=15 / C=17 at tol=1e-8.

Mutation kills (all EXECUTED, restore-verified)

Mutant	Killed by
MU1 chain transpose	one-shot A pins (dev 0.685)
MU2 drop diagonal subtraction	positivity guard errs (exact statdist [9/2, 3/4, -9/4])
MU3 drop centering	A params (dev 0.033)
MU4 statdist sum n -> 1	weights pins (params PROVEN invariant — documented limitation)
MU5 denominator (w_i+w_j) -> w_i	I-LSR pins (one-shot UNOBSERVABLE at uniform weights — documented limitation; dev 0.331 from pass 2)
MU6 tol*n -> tol	fixture C iteration pin 17 vs 18 (A does not separate: 15/15 — measured)

Tests

• 5 Rust tests + MC-500 #[ignore] (all asserts read crate outputs); cargo 745 pass / 92 ignored.
• 3 pytest tests (exact pins, BT cross-anchor, error contract incl. overflow); pytest 273 pass.

Adversarial impl-review outcome

• Round 1 (files/lsr_impl_review.md): 1 HIGH finding — valid connected win matrices with globally huge counts (base*1e20, asymmetric 1e150) were falsely rejected: the Gaussian-elimination pivot threshold compared an O(1) sum-constraint pivot against a global scale dominated by O(count) stationarity rows.
• Fix (commit 343fc2f): normalize the generator to unit max magnitude after the overflow guard (stationary distribution is invariant under global rescaling of transition rates). Regression asserts: base vs base*1e20 params/weights equal to 1e-12; asymmetric 1e150 ILSR finite. Overflow fixture replaced with a genuinely overflowing n=4 all-1.7e308 matrix (the old n=3 all-1e308 fixture never overflowed and is correctly accepted post-fix).
• Targeted adversarial re-review (files/lsr_rereview.md): VERDICT CLEAN — invariance re-derived, guard ordering verified, regression tests confirmed to fail without the fix (revert probe), exact-fraction pins unaffected, pip choix cross-check on scaled inputs matches.

[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: ce1e2286-f8b9-463c-a7e3-2cf05bf8cccc

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-lsr

---

_{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.