refactor(s5): share GGM/Mixed precision kernels as cholesky_helpers free functions

[MaartenMarsman]

S5 from the class-design review, in the free-function form the review-critique settled on (not a composed GaussianPrecisionBlock).

What

The rank-2/rank-1 determinant-lemma log-ratio and the Roverato proposal-constants extraction were near-identical between GGMModel and MixedMRFModel, differing only in where the current precision entries come from: GGM stores K directly; Mixed stores Kyy = -2 * pairwise_effects_continuous_. This is the cross-model divergence the review flagged (det-lemma block appearing 4×, and the drift-guard that was added to one model and not the other).

Promotes the shared math to stateless free functions in cholesky_helpers.h:

• log_det_ratio_edge_kernel / log_det_ratio_diag_kernel
• precision_proposal_constants (returns the 6 reparameterization constants)

Each takes the covariance (and Cholesky), the indices, and the resolved precision scalars — GGM passes precision_matrix_ entries, Mixed passes -2 * pairwise. The six member functions (GGM::{log_det_ratio_edge,log_det_ratio_diag,get_constants}, Mixed::{log_det_ratio_yy_edge,log_det_ratio_yy_diag,get_precision_constants}) become thin adapters.

Per the critique, these are free functions, not a composed block: only these leaf kernels are genuinely identical; the "owns Σ" framing leaks (Mixed fuses the Kyy update with the OMRF marginal recomputation).

The -2/-½ boundary accessor (the other half of the simpler-S5 plan) is not in scope here; this PR is the cross-model kernel dedup.

Why this is safe

The two models' originals were already identical up to int-vs-double literals (1 vs 1.0, 2 vs 2.0) that don't change the double result, and the precision-source difference now passed in as scalars (-2.0 * pairwise(i,j) is the same double as Mixed computed inline).

cholesky_helpers.h now includes explog_macros.h so the kernels use the same MY_LOG/MY_EXP as the originals — required for bitwise fidelity (these can be the OpenLibM __ieee754_* variants, not std::).

Verified bitwise across every path that hits these kernels: adaptive-Metropolis GGM (no-edge δ=0/1 + edge-sel), NUTS GGM (+ mass adaptation), and adaptive-Metropolis mixed-MRF (edge-sel + no-edge, δ=0/1). identical(main, branch) == TRUE for all three. Harnesses: dev/bitwise_ggm_mh_primitive.R, dev/bitwise_ggm_nuts.R, dev/bitwise_mixed_logdet_ratio.R.

[codecov]

Codecov Report

❌ Patch coverage is 94.64286% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 87.69%. Comparing base (9e21f06) to head (786d85b).

Files with missing lines	Patch %	Lines
src/models/mixed/mixed_mrf_metropolis.cpp	76.92%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #140      +/-   ##
==========================================
- Coverage   87.74%   87.69%   -0.06%     
==========================================
  Files          87       87              
  Lines       12860    12855       -5     
==========================================
- Hits        11284    11273      -11     
- Misses       1576     1582       +6     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.