DF-2719 race-stress evidence (captured from guest sessions, untrimmed excerpts)

--- calibration ---
$ cc -O2 -o cal cal.c && ./cal
munmap(40000 entries) = 11.14 ms

--- run 1: 60 iterations, 20000 maps, delay sweep 0..3000us ---
parent: 20000 mappings at 0x10000000
iter 0 delay 0
iter 50 delay 750
kernel: pid 1478..1485 (racedump), uid 1001: exited on signal 11 (core dumped)
$ ls -l /tmp/racedump.core
-rw-------  1 maxx  wheel  1019904 ...    <- ~8.5k vn_hdrs of a 20k-entry map:
                                              entries were being unlinked while
                                              the dump ran (mid-teardown snapshot)

--- run 2: 2500 iterations, 40000 maps, sweep 500..10000us ---
iter 2400 delay 500 / iter 2450 delay 2875  (completed, guest up, no panic)

--- run 3: 3 parallel racers x 4000 iters, per-pid core files
    (sysctl kern.corefile="%N.%P.core") ---
8142 cores written; r1/r2/r3 reached iter ~2700 each; guest up; no panic

--- run 4: 2 racers x 1200 iters, 100000-entry maps (fork churn of the
    vm_map_entry objcache), sweep 1000..15000us ---
both racers completed iter 1150; 2406 cores; guest up; no panic
many children "exited on signal 11" without "(core dumped)" (dump write
failed on tmpfs capacity) - each_segment passes still executed before that

TOTAL: ~13,400 attempts, 0 panics, 0 wedges. Mid-teardown cores prove the
traversal/mutation overlap occurred; the torn-read needed for a crash did not.