Audit and fill math-kernel benchmark coverage

[acgetchell]

Context

#436 adapts the release-performance workflow from la-stack: shared command names, temp-worktree release comparisons, durable GitHub Release benchmark assets, and curated performance documentation.

That workflow intentionally keeps the default release-signal suite focused and stable. The broader math-kernel coverage review called out in #436 should be handled separately so we can decide which kernels belong in release-signal benchmarking, profiling, allocation checks, or targeted microbenchmarks without making the default release comparison too noisy.

Scope

Review Delaunay's math-intensive benchmark coverage and fill important gaps for release-relevant kernels. In particular, audit coverage for:

• orientation predicates and exact fallback behavior;
• barycentric simplex-intersection checks;
• circumsphere / insphere predicate families;
• simplex quality, volume, radius, and related geometry helpers;
• ridge, Euler, manifold, and topology-validation work;
• locate, hull traversal, and visible-facet queries;
• simplex barycenter queries;
• allocation hot paths only when output is stable enough for release-to-release comparison.

Deliverables

• Decide which benchmarks belong in the release-signal suite versus broader profiling or targeted diagnostics.
• Add focused Criterion benchmarks where meaningful coverage is missing.
• Keep bench-latest curated: avoid adding noisy or expensive exploratory benchmarks by default.
• Update benches/README.md with the benchmark taxonomy and any new commands or suites.
• Run the matching benchmark compile/smoke validators, plus measured before/after evidence for any performance claims.

Acceptance Criteria

• Existing benchmark coverage is mapped against the math-kernel list above.
• Missing high-value kernels have focused benchmarks or an explicit rationale for deferral.
• Release-signal membership is documented for each added or reviewed benchmark.
• Benchmark docs explain which commands maintainers should use for release checks, regression checks, profiling, and kernel investigations.
• New benchmark code preserves Delaunay's numerical, topological, and validation-layer invariants.