Regime-C Cholesky reportedly host-synchronizes per front through its info check (blocks stream overlap of wide fronts); verify #110

Found while working on: PR #107 (performance prototypes; stream-pool experiment in bench/fact_fused.jl)

What: the prototype reports that _factor_panel_c! (src/numeric/factorize.jl, regime-C Cholesky) synchronizes the host once per front through its info check, so independent wide fronts cannot overlap on a stream pool (measured neutral for that reason) and the numeric phase is not host-synchronization-free as PLAN §3.9 states. The code calls _potrf_info_impl!, which was introduced in T09 to keep the status on the device, so either the vendor binding (vendor_potrf_info!, CUDACore.with_workspace plus the raw cusolverDn?potrf) synchronizes internally, or the report refers to another read. To verify with a CUDA profile of one refactorization on a matrix with several regime-C fronts (bench/profile_phases.jl): count the host synchronizations per factorization.

Why it is out of scope: #107 is bench-level; the vendor path is T09 code and overlap of wide fronts is T25 work. Blocks nothing; it caps experiment 5 if confirmed.

Suggested fix or plan change: if confirmed, an asynchronous vendor potrf whose info stays on the device without a workspace query per call (query the workspace size once at analysis), and the same for the LDLᵀ/LU regime-C paths. Owner note under T25.

Generated by Claude Code