Dump of file fixture: the negative twin of scheduled-jump-own-code.*. Byte
for byte the same function, the same isa-flag compare and the same
scheduled mov, with ONE line taken out: the jl that jumped over the two
vpmullq into the scalar fallback. Without it the compare's result is never
branched on and execution falls straight into the AVX-512 block, which is
the real K1 crash shape.

This must stay UNPROTECTED. It is what keeps the tolerance added for form
(i) honest: the guard learned to follow a conditional jump across
instructions that cannot write EFLAGS, not to accept "there is a cmp on the
isa flag somewhere above".

  000000018007A970: 48 83 EC 18        sub         rsp,18h
  000000018007A974: 83 3D 15 10 19 00  cmp         dword ptr [000000018020B990h],6
                    06
  000000018007A97B: 4C 8B 19           mov         r11,qword ptr [rcx]
  000000018007AA0A: 0F 54 C4           andps       xmm0,xmm4
  000000018007AA0D: 62 F2 FD 08 40 D5  vpmullq     xmm2,xmm0,xmm5
  000000018007AA25: 0F 54 C4           andps       xmm0,xmm4
  000000018007AA28: 62 F2 FD 08 40 CD  vpmullq     xmm1,xmm0,xmm5
  000000018007AA4D: C3                 ret
  000000018007AA4E: 48 89 5C 24 20     mov         qword ptr [rsp+20h],rbx
  000000018007AA53: 45 33 C0           xor         r8d,r8d
