So how should I structure my chunks of SQL so that I can have "safe" (all-or-nothing) blocks,
AND use them from within one another?
While there are more advanced constructs that may aid here I would suggest just following two rules:
top-level scripts handle transactions - usually just a single BEGIN/COMMIT at the top/bottom of the script respectively. These are scripts you are allowed to name on the psql command line.
support scripts don't handle transactions - these are scripts you include into the main script with \i or \ir
