[YPARK]

What

Speeds up both blockmodel presets without changing the moves.

• Release profile: fat LTO, one codegen unit.
• Finite strength (link preset):
  • The Gibbs relabel loops sites outer and blocks inner over contiguous memory.
  • The ln Gamma and prior ln Beta terms are cached per (site, block), and the caches are kept current on every update.
  • Whole-read sites multiply rising-factorial ratios into a running product per block, kept as a mantissa and a power of two, so each block takes one log instead of one per term.
• Infinite strength (blocks preset):
  • exp(-shift), exp(-level) and exp(-mu) are cached, so ln sigmoid becomes one log of a product per block.
  • Extreme logits fall back to ln_1p, so the product cannot overflow.
• Fast ln Gamma: an f32 approximation below 16, and the library at 16 and above. It is used at sites with fractional or large counts.
  • The target shifts slightly, by about 1e-3 per term.
  • The moves leave that target invariant exactly: the exactness tests' reference densities use the same functions.
• Memory estimate: counts the new caches.

Tests

• Golden test: fixed-seed chains are pinned bit for bit, recorded before any change. Every step was checked against it, and the changes in its recorded values are explained in the commit. It runs on Apple silicon only, because it depends on the platform's maths library.
• Exactness: the existing total-variation, swap-ratio, log-joint-definition and conditional tests all pass. A weighted fixture now exercises the fast-ln-Gamma sites.
• New checks, each with a mutation check that makes it fail:
  • the caches match the live reads, shifts and levels;
  • a long link's predictive equals its sites summed;
  • the mantissa splits rebuild exactly;
  • extreme logits score right;
  • ln Gamma differences stay near the library's.