Is it common to choose not best reflection coeff. just to ease matching network?

In oscillator design, if it gives enough power and reflection coefficients reside in unstable regions, is it good idea not to optimize it for better power if it makes matching more difficult?

It is very common. A good example are GaAs FET amplifiers. The gate of the FET may have a S11 of >.9 making it very difficult to get a good match at a single frequency due to component Qs and stability issues. This task becomes impossible over a broad frequency range. Also the optimal noise figure match does not occur with a S11=0 but something closer to .65 depending on device and frequency. Similarly the best output power will not be at S22=0.

A common approach is to place two identical amplifier between hybrid couplers. The coupler provides a good broadband match to the system. For a single amplifier ferrite isolators have also been used.
Another approach for broadband is using a distributed amplifier design.

The same applies to oscillators where tuning range and noise performance will drive the network design.