There are several issues here.
- Note that what you call "LM-normal" and "GLM-normal" are the same thing. You may be confusing statistical concepts with the R codes that are used to instantiate them.
- The distributions / types of predictors is irrelevant. We are generally interested in the response distribution.
- Critically, however, we are interested in the conditional response distribution, not the marginal distribution (see: What if residuals are normally distributed, but y is not?)
The last point means it is often difficult to identify the correct distribution just by looking at the variable in isolation (e.g., in a histogram).
Typically, we identify a distribution / appropriate model by first thinking about the nature of our response data? Are they continuous? (If so, they could be normal, but they cannot be binomial or Poisson.) If they are counts, are they out of a known total (like the number of heads from a number of coin flips)? If so, they would be binomial. If there is no upper limit, they could be Poisson (but actual Poisson data are rare).