If UA (server side) is unknown, return feature detection kit that sets features object client side and sends results back to server so the server can cache it.
Any consequtive request for that specific UA will return a much smaller server cached features object script.
Lets say (theoretically) the full feature test script is 50kb and a cached feature set script 10kb. The first ever request for an unique UA will be a hit for 50kb and every other request after that (for that specific UA) for just 10kb.
For a high traffic website this can offer substantial savings.
Also, after saving say 20 unique UA feature tests, the service will also become rather static as 80 to 90% of the audience is covered. No biggy, the service will be useful still to feature test future browsers.