perf(memmove): reject array-copy lowering gate

Baseline RCH vmi1227854 at 17e2a32: memmove_4096 FrankenLibC p50/mean 32.969/39.065 ns vs host 28.355/31.413 ns.

Tested one exact count==4096 safe-slice branch; behavior guard passed and golden payload sha256 was abaa13b6f9a5ccaa56e4f4b20d22a0421a0a780df1794353efb861ab66e548b4.

Rejected/restored: LLVM IR still lowered memmove to dynamic-length llvm.memcpy with no retained i64 4096 copy. Same-worker post was FL p50/mean 31.901/35.401 ns vs host 28.259/30.289 ns, but apparent win lacked a changed generated primitive. Score 0.0; source restored.

Author: Dicklesworthstone