TESTING.md: replace-fixes: replace with discards (fstrim, discard=sync/async, no -K)
New rows for replace-discard-test.sh on both lanes, base zb64 and the
josef-trim candidate (lane zb64 + Josef Bacik's "btrfs: don't trim the
target of a device replace"), plus discard=async reruns of
replace-status and replace-cancel-persist. FITRIM during a replace
loses the target's copy on both lane bases (single: the fs no longer
mounts; raid1: target alone does not mount; raid5: silent, 169296 csum
errors on the target); the candidate PASSes every row. The bug is
upstream, not fixed by this series. Lists the checks that are invalid
in discard mode (the poison strip checks) and what is used instead.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>