Scattered DRAM-missing stores against a 56-entry store buffer

On an x86 core with a 224-entry ROB, a 72-entry load buffer, and a 56-entry store buffer, this loop scatters 4-byte stores into a 4 GB table at random, precomputed indices. idx is read sequentially and prefetches well. Nearly every store targets a line not present in any cache, and nothing in the loop reads table. Profiling shows the ROB is rarely more than a quarter full, yet the allocation stage spends most cycles unable to issue new uops. What is happening?

main.cpp
// table: 4 GB of uint32_t; idx[i] is uniformly random
for (size_t i = 0; i < n; ++i)
    table[idx[i]] = static_cast<uint32_t>(i);

43 of 47 in Computer Architecture