Pad 1310nm cross-connects?

  LR usually needs padding in that scenario, IMHO.

Usually does not.

1G parts are so cheap that measurement, test and the attenuators
(unless you are wrapping the fibre round a pencil) will cost more
than each device is worth. How many fail?

If in doubt check the device spec such as

                 Transmit Power (dBm) Receive Power (dBm)
                  Maximum Minimum Maximum Minimum
Cisco SFP-10G-LR 0.5 -8.2 0.5 -14.4

the max levels are the same so for that part they are designed to be
fine without even the loss of a minimal interconnect.

This also
applies to MMR interconnects or other "premises" / "campus" situations. 5
or 10dB depending on patching quality -- sometimes up to 15

15 would take the above device out of spec, even 5 is marginal with
typical MMR cross connects. Over its lfe the tx output may drop
and still be within spec so you can't base this on measured when
installed levels.

As you write, the damage from overload is gradual, so simply
trusting "it works" is quite bad for longevity reasons.

If it's operating within spec then return the parts if there is a
significant failure rate attributable (what temperature are you running
them at, that would need to be considered in any statistically
significant study)

If you have evidence of level affecting longevity as described please
post the data or paper demonstrating it which should then let us
choose the optimum signal level for maximum life with a reasonable
operating margin