Case in point: putting if _ then _ else on separate lines is fine, but inside a do-notation it causes problems because the layout stage inserts semicolons between the parts, and the production rules listed at the bottom of the page do not include optional semicolons in the "if" alternation. You need to indent the then/else to overcome it. This problem is fixed with a hack[0], which fortunately has no side-effects for the rest of the language, but it just shows that throwing together ad-hoc rules for parsing code, rather than well understood theory like LR parsing, is inviting room for problems.
With CFG subsets like LR/LL, every production rule in these grammars is also a valid LR/LL grammar - in other words, it's not only the entire grammar of a language that is context-free, but all the parts are too. An indentation-sensitive language is only context-free when looked at for the entire language grammar, but the individual production rules of are context-sensitive. You can't just copy and paste some code into the syntax without providing this context (by means of manually indenting the code).
[0]:https://ghc.haskell.org/trac/haskell-prime/wiki/DoAndIfThenE...
Anyway, Haskell is a very instructive example about how semantic indentation creates problems.
[0] https://www.haskell.org/onlinereport/syntax-iso.html#sect9.3