I'm not sure if this is a bug, but using ghc -prof -auto-all and manual SCC annotations for non-toplevel functions produces very different results from moving the same function to toplevel:

fib n = nfib' n
  where
    nfib' n = {-# SCC "nfib'" #-} nfib n -- %time: 0 (both individual and inherited)
      where
        nfib n = if n < 2 then 1 else nfib (n-1) + nfib (n-2)

fib n = nfib' n
nfib' n = nfib n -- %time: 100 (both individual and inherited)
  where
    nfib n = if n < 2 then 1 else nfib (n-1) + nfib (n-2)

This happens with both -O0 and -O2, even when these functions are not exported. Adding a NOINLINE for nfib' doesn't change anything.

Computing the correct inherited time for the SCC would be useful to pinpoint the cost centres inside a more complex function. Currently one needs to move them to toplevel to get the same effect. Given that the toplevel route works, it's also counterintuitive that manual SCC annotations don't.