When we first introduced full time data capture, it was written to avoid damaging the parent process if it all possible. In the interest of achieving that aim, certain failures cause ftdc to halt itself, rather than retry or terminate the process.
However, over the years we've come to rely more and more on the diagnostics ftdc provides. We've also seen few to no instances of ftdc taking down an otherwise healthy node.
Making ftdc failures terminate the process will reduce the number of cases where a moment of resource exhaustion (say in file descriptors) causes ftdc to unload, after which we're left with a less than fully operational mongod/s