Right now, the default interrupt handler eventually ends up
in restarting the application at address 0, so the user-
visible effect is almost similar to a reset, which has turned
out to cause minor surprises occasionally.

Other ideas have been discussed about the default ISR.  Just
using a RETI is not really better: some interrupts are not
auto-clearing so they would end up in an infinite loop, and
in all other cases a missing ISR is a programming error most
of the time which would go completely unnoticed for a long
time then.

The best idea so far came once from Peter Dannegger.  He
suggested to CALL/RCALL the default handler from within the
vector table, rather than JMP/RJMP to it.  That way, the
default handler can then pop the return address into a register
where it can be examined with a debugger (or inside the default
handler itself if it's a custom handler).  That way, programmers
will get an idea what interrupt triggered.

Finally, the default handler should best run into an infinite
loop with interrupts disabled.  abort() has been suggested a
possible implementation for that which is already there.

The difficulty of this idea is that the current scheme to
fill the vector table with JMP/RJMP instructions to locations
that are eventually inserted by the linker won't work anymore.
A new scheme would be needed where by default, a CALL/RCALL
is placed into the slot, and is then replaced by a jump
instructions for all vectors that are actually defined by
the application.