Code:
But I see that for each file we have to apply a manual correction factor to avoid saturation of the amplitude or volume of the final sound.
You can use a floating-point format and a normalization after addition.

Code:
It also loses the modifications made in each individual sound like volume, pan, frequency, effects, etc. In other words, I could only work with each original sound, load without play them and to make the mixing and later recording.
You can't access to the data without playing the sound. AFAIK DS mixes the buffers in the kernel mixer and passes it to the driver.