Calibration derives what it needs from a proxy once

[bjmeetsfo]

A calibration round walks the proxy groups, and for each one it walked every
proxy three times over: once to collect what is attached to that group, once
to parse every proxy's location and test it against the group's, and once
more -- inside a filter -- to scan the attachments already planned this
round.

None of that varies with the group asking. A proxy's location is the same
string whichever group considers it, and parsing allocated a fresh one every
time. Which group a proxy is attached to is a property of the proxy, known
before the walk begins. And "has this round already promised this proxy" is
a membership question that was being answered by scanning the plan so far,
so the cost grew with each attachment made.

groups x proxies      before      after
    25 x    250     2 324 us     701 us
    50 x    500    10 479 us   3 238 us
   100 x  1 000    58 020 us  11 772 us

Five times faster at the largest, measured before and after back to back,
and what is left grows with groups times proxies rather than faster than it.

The rule that two groups never claim the same proxy is what the scan was
buying. It has a test, and removing the set that replaced the scan fails it.