There was an attempt last year at a 7DRL based on Midsomer Murders. More recently, I've seen
Murder at Masquerade Manor, which is a lot more fun but quite easy, and
Noir Syndrome (which I haven't played).
MidsomerRL and MAMM both work on small, standalone mysteries with few actors. One person is declared dead, and you interview everyone else to work out who's responsible. In MidsomerRL they tell you where they claim to have seen another character, and in MAMM they offer a wider variety of information - who they were with, where, who they saw carrying improvised weapons, and who had a motive.
In MAMM it's quite easy to figure the mystery out with a notepad, because only the murderer lies. You can find the murder weapon, but the dialogue gives enough evidence anyway. Apart from mis-identifying the killer, it provides a lose condition in allowing the killer to stab you in the back. This has some pretty weird interactions with the FPP game mechanics.
If you want the game to be a direct representation of a fictional detective's role, rather than a more abstract representation like in Cluedo, I think the main way to improve on systems like MidsomerRL and MAMM will be to have more reasons for characters to lie and ways fr evidence to contradict statements. That way a single interview per character won't be enough to tell the player what happened - only who needs to be interviewed more thoroughly. This might make it more difficult to guarantee the mystery can be solved by perfect play, but not all roguelikes are winnable every time, right?
Noir Syndrome seems to be an open-world game, so I don't know how much inspiration it can contribute to design of small-scale procedural mysteries. There's another game I'm aware but can't remember the name of - it doesn't procedurally generate, but chooses one of 8 or so mysteries at random. You don't know which one until you solve it. It might have been an Infocom game?