Any sufficiently simple technology (like recording sound) is bound to be made more complex through the application of programming complexity.

I suggest that you keep the simple things, simple.

The tool you need is already written and available for Linux/Unix, Windows, and Mac OS 9 or OS X. See Audacity.

As a special bonus, the source code is available under GNU|GPL.

I realize this is not a perlish solution, but the requirements you mention (1. recording 2. modification via rewinding and 3. Overwriting recent recorded information and 4. cross platform issues) are all covered.

As a special bonus, you can change the speed of a recording with Audacity. This greatly assists the transcriptionst.

Your project may also benefit from text to speech technology but perhaps only after you get the recording requirements nailed down first.

No suggestion will be useful unless the software actually works. I have found Audacity extremely stable on Linux where editing single sound files the size of seven standard CDs is a simple task.