[DanielThomas]

I was running some profiles against Images and Supermarket and found that Id normalisation was a hotspot. Avoiding this work should give a small but noticeable improvement to CPU utilisation from Spectator.

Benchmark

JDK 25, one thread, two forks:

Benchmark                                      Before       After       Change
CounterIncrement.lookupCost                    32.1 M/s     133.0 M/s   4.1x
CounterIncrement.taggedLookupCost               7.10 M/s    119.5 M/s  16.8x
CounterIncrement.invalidLookupCost              4.37 M/s      4.59 M/s  no regression

Allocation is unchanged: 40 B/op for valid lookups and 304 B/op for the invalid-id normalization path. The allocation is the SwapCounter returned from each registry lookup plus, for the invalid case, the normalized id.