Gotcha 39 is about non-KV headroom, and on WSL2 the shortfall is silent

MAX_SEQS and KV_MEM are two doors into the same room (mjungnickel18, #25):
walking it through the KV_MEM door at 8 seats lands on the same allocator
fingerprint as the seat ladder -- the identical four failed 20 MiB mappings,
byte-identical counters, both repeats. The floor at CTX=huge/8 seats sits
between 396 (dead on the first 8-way batch) and 436 MiB free (full speed
with 8 MiB left): the transient set is elastic, the first real batch's
per-seat bill is not, and nothing on either ladder was ever merely slow.

The platform note from the same thread: bare metal fails loud
(OutOfMemoryError naming the bytes), WSL2/WDDM host-backs the same miss and
fails silent -- no error, prefill 5-10x slower, and free-after-boot reads
hundreds of MiB rosier than Linux for the same pin (630 vs 98 MiB on the
pin that started this). The launcher now warns when KV_MEM exceeds the
profile default, which would have fired on that pin the day it was set.