These are wonderful logic puzzles and particularly good implementations of them. For instance, their Minesweeper guarantees that it is solvable - you will never have a 50/50 choice you cannot identify.
https://f-droid.org/de/packages/name.boyle.chris.sgtpuzzles/
Alas, I've backed myself into a corner a couple of times and triggered "just gotta guess" choices in it - while it's generally good at this, it's not a 100% implementation based on my playing.
I find "Net" way more fun though for quick casual gaming, a 7x11 grid (depends on exact screen size) with wraparound enabled is a favorite for easy to tap but enough squares to make it take some time to solve (about 5 minutes per game, give or take).
This should not happen because the current implementation [1] always tries to solve a randomly generated puzzle deductively, and never generates a puzzle that hasn't passed the check. (There are some shortcuts, including dynamically "perturbing" the current puzzle to make it uniquely solvable.) "Solvable" puzzles do not guarantee no backtracking though, so that's probably where you gave up. Also note that you should take account for the number of remaining mines, which can frequently be the sole information left for the very last mines.
[1] https://git.tartarus.org/?p=simon/puzzles.git;a=blob;f=mines...
I've also played it quite a bit and can't remember having any undecidable boards.
At least for me sometimes a situation that looked like "just gotta guess" was actually solvable by knowing the total number of remaining mines, as one choice in the "guess" would imply more mines than the other.
Sometimes in constraint puzzles, one clue is also that there is a unique solution. So if doing one choice implies that some other choice can be arbitrary, that's not the solution.