BASELINE (stock kernel #0, INVARIANTS):
  attempt 1 (lockstep churn, no persistent fillers):
    victim : 20971520 iters, 0 errors      <- table never full; negative evidence
    churn  : 20971520 cycles                  that the full-table precondition is required
  attempt 2 (persistent 25 fillers + dip/refill churn, staggered threads):
    table pre-filled with 25 filler entries; racing (expect panic: "semop - can't undo undos") ...
    [~3s later guest wedged in DDB]
    panic: semop - can't undo undos          (panic.txt, serial console)
    cpuid = 3
    sys_semop() at sys_semop+0x65d
    syscall2() at syscall2+0x11e

PATCHED (kernel #1, fix.diff): see run.fixed.log
    victim : 306184192 iters, 21247 errors
    dmesg  : 170 x "semop: SEM_UNDO rollback lost a race, pid 860"
    guest  : up (300+ seconds)