Hugely complex question. It has been considered and reconsidered many times. There is the neutral theory, which states that the vast overwhelming majority of mutations have no effect on the phenotype. The nearly neutral theory and the classical theory. Neutral theory is a big subject.
On indels, these can be "neutral" and reversing depends how much selection pressure they are under. Microsatellites do "reverse".
Essentially, this is a massive question and its better to consider a single species as an example for a single type of mutation
See the comment by @MaximillianPress. That is a good answer, the step-wise mutation model is not a bad model for (microsatellite indels) and states indels can "grow" and "shrink" one repeat unit at a time. I know what the OP is saying alot of indels occur in protein genes and must have a specific function.
Indels under selection pressure are common in the surface antigens of many pathogens including HIV. The distribution for codon indels in env was modelled by Fletcher and Yang 2009 here. This is a strongly skewed distribution and maintenance of very long indel lengths of up to 100 could only be maintained by selection pressure, however the preponderance of small copy number indels is massively dominant.
The stepwise model (SMM) is not a skewed distribution it assumes that there is equal probability for increasing the indel size as it is reducing it. This was used for microsatellites which are assumed to be neutral (not under selection pressure).
In protein modeling under the old DUET model (now deprecated) by Pires and Tom Blundell 2014 showed maximum structural disruption for indels.