zdb_modified was just a modified zdb to cause it to always print holes
with >= 5 -v flags - the diff for zdb can be found in zdb_print_holes.patch
(and a version of the patch that applies to ZoL head can be found in PR
#5369, or zdb_print_holes_rebased.patch).

sound_alarms.py currently has hardcoded out_A1 and out_A2 reads, which are
simple to fix - it expects the output from running (modified) zdb on the 
same file across two snapshots in sequence, e.g. 
zdb -ddddd -bbbbb -vvvvv pool/dataset/foo@snap1 $(stat -c '%i' filename)
zdb -ddddd -bbbbb -vvvvv pool/dataset/foo@snap2 $(stat -c '%i' filename)

It might make sense to run something like a log(n) search - check the first
snapshot after hole_birth was enabled for birth_time 0 holes, and the most
recent snapshot, and if any are in the latter that aren't in the former,
DING DING DING WE HAVE A HOLE_BIRTH BUG.

I verified that this successfully detects known hole_birth bugs in the test
datasets I generated using the scripts in #4809 - patches to make this more
robust and/or generally useful welcome.