You can actually do a soft transition from the (opaque) Arduino programming style to a programming style closer to hardware. After all, Arduino programs are just C++ and you can plug in your own functions and assembler and you can even compile and upload your own ASM/C/C++ program written from scratch and compiled with the gnu-avr-toolchain to the UNO.

After you have done all of that it might be a good idea to start soldering the AVR on your own boards.

Isn't it better to start with the mbed though(if you intend for a commercial product, at least at kickstarter/~20K-units level) ?

The quality of the libraries is higher(written by professionals), you have a wider selection of supported mcu's if you'd need to port, some mcu's support the mbed, easy to-use low-power api(an event driven framework that automatically puts the mcu to sleep when it isn't needed), ARM has put a lot of work in the security of the device and that's something that's very hard to replicate.

And like the arduino, if something isn't optimal - you can always write your own.

The only major drawback is the mbed is a bit more complex. For example, you need to use pointers, unlike the Arduino, where it's not a necessity.

Not really, I generally find vendor libraries to be middling in quality, I think the core Arduino libraries are actually far more battle tested (if much less featureful). If you don't need more than the Arduino provides, it's probably not worth the extra complexity.

The Atmega8 (and entire series of attiny/atmegas) is totally suitable for commercial applications - just stick the chip directly on your PCB. AVRs are pretty ubiquitous in consumer products.