cuda-ab: vector_add bad-region localization + h_c marker scan

Prints first/last bad index, per-128K chunk bad counts, the marker
scan of h_c (0xCC-filled before the D2H: 'VADD marker-untouched=N'
names exactly which regions the readback never wrote) and the a/b/c
sample at the first bad index. This localized the large-buffer
failure: engine outputs match the readback pages (content loss is
upstream of the read), and in dirty boots the client's H2D ops do
not decode at all (zero CEAUDIT for them).

0 files changed, 0 insertions, 0 deletions