You're right that proving that a program halts doesn't tell us whether or not it's practical to run (it could take 1 trillion years), but it can be useful for other reasons.
For example, if a function claims to return a value of type T, and we can prove that the function halts, then we know that a value of type T must exist (it's whatever the function will return, however long that takes).
This is how dependently typed languages can implement pre and post conditions: we can represent conditions as types, and put preconditions as extra argument types, and postconditions as extra return values (e.g. in a tuple).
For example, we might have a function 'f1' which turns 'A' values into 'B' values:
f1 : A -> B
f1 myA = ...
If we require some precondition 'PRE' to hold for the argument, and postcondition 'POST' to hold for the return value, we can use dependent types to represent them:
f2 : (a : A) -> PRE a -> (b : B, POST b)
f2 myA preProof = ...
We can 'f1' by just giving it an 'A' argument, but the precondition won't be checked. Likewise, we learn nothing about whether the output satisfies the postcondition.
With 'f2', we must give it an extra argument, representing a proof that the precondition holds for the given 'A' value. The function must also do more work: not only must it return a 'B' value, but it must also give us a proof that that 'B' value satisfies the postcondition (we can pass that proof around, e.g. if another function requires it as a precondition).
If we can prove that these calculations will halt, then we don't actually need to run them at all! Knowing that 'PRE a' can be proved in a finite amount of time (however long) is itself a proof of 'PRE a'! (And likewise for 'POST b'). Hence we can "erase" this proof stuff after type checking, to generate code equivalent to 'f1'.
If we don't know whether or not those proof calculations will halt, we can't use this shortcut, and we're forced to run the proof calculations in full at runtime (which may not halt, or might take a trillion years!).