+Note [Case MFEs]
+~~~~~~~~~~~~~~~~
+We don't float a case expression as an MFE from a strict context. Why not?
+Because in doing so we share a tiny bit of computation (the switch) but
+in exchange we build a thunk, which is bad. This case reduces allocation
+by 7% in spectral/puzzle (a rather strange benchmark) and 1.2% in real/fem.
+Doesn't change any other allocation at all.
+