> Converting them to a cleaner ReaderT led to a 10% total run time regression, so I had to revert it. It makes me wonder about the speed penalty of code I designed better to begin with.

I have heard from other Haskellers that you can sometimes get good performance by hand-rolling an application monad, and then writing out the MonadFoo instances so you get good ergonomics:

    -- Instead of this:
    newtype App a = App { runApp :: ReaderT AppEnv (ExceptT AppError IO) a }
      deriving (Functor, Applicative, Monad, MonadIO, MonadReader AppEnv, MonadError AppError)
    
    -- Try this:
    newtype App a = App { runApp :: AppEnv -> IO (Either AppError a) } deriving Functor
    
    instance Applicative App where ...
    instance Monad App where ...
    instance MonadIO App where ...
    instance MonadReader AppEnv App where ...
    instance MonadError AppError App where ...