Expand description
Batched topology scoring orchestration over request-local indexes.
The native bounded exact-induction path drives the kernels/ilp_exact.cu
kernels that score all (left, right) pairs for all four topologies in a
single batched GPU pass, avoiding per-candidate host transfers.