Hi,
When using a usb mic for input I have also noticed a growing latency like you describe it. The increase is rather fast: a few samples per second in the first minutes. This is probably caused by the fact that usb mic and internal soundcard (which is used for the output) have different clocks, and to keep the stream continuous, the buffer size needs to be constantly adapted. When running the same device configuration with jack, there is an option to 'compensate for clock drift'. The latency is then constant (at least, in my usb mic case). However, jack does not seem to solve clock drift in a very elegant manner. When I routed a pure sinusoid back from speaker via mic into a Pd patch, I noticed regular clicks, which were also visible as jumps on a scope trace.
All taken together, I suspect that the best solution is to use one and the same audio interface for all inputs and outputs, so they share one clock.
By the way how would you measure thru latency while music is playing? I use a Pd patch for sample-precise latency measuring but this works with a pulse so it can not be done while playing music. Oh maybe it can be done with correlation? Then still you would need a reserved input for measurement. Or not? Is there some clever method, if you know, please tell.
Katja
edit: sorry I forgot to mention, Pd-extended, OSX 10.5.8