Okay, round two:

Marc Horowitz pointed out that the Intel PCIC chip (which practically
everybody emulates) has a bit to tweak that causes an interrupt.  I
checked on this, and sure enough this is what Linux uses to get a IRQ
mask; it probes each IRQ that's not being used by anything else and
see's if it's valid.  (However, there's also a section in their
documentation that says, "Here's what you do if your machine hangs
during the IRQ scan").  The documentation I have for the Cirrus Logic
PCMCIA controller doesn't mention this bit, but it is marked preliminary,
so I don't know if the real Cirrus Logic chip has this bit.

Perhaps this is what Win95 and other OSs do?

Anyway ... if we did _this_, would people think it was reasonable
behavior?  (With the ability to turn this off, somehow ... maybe via
run-time configuration :-) ).

--Ken