docs(handoff): a timed benchmark does not distinguish Windows power schemes

Tried the obvious way round the missing Windows metric: if the reported CPU
clock is nominal, measure achieved throughput instead. It does not work,
for two independent reasons.

Measured, fixed single-threaded integer workload, nine samples each:

  Balanced          median 236.8 ms, MAD 2.2
  High performance  median 239.2 ms, MAD 2.3

A 2.4 ms delta against a 4.5 ms noise band, sign backwards. By the
threshold rule in tuning::verify that is Unchanged. A lightly loaded
desktop boosts a single thread to the same place under both schemes.

The second reason closes the avenue whatever a sustained all-core run would
have shown: a metric that requires running a benchmark competes with the
workload it is tuning for. Burning every core to find out whether the
machine got faster is self-defeating when it was already busy and
meaningless when it was not.

So the metric must be passive observation of the real workload's own rate —
tokens per second off an inference server simon already detects, not a
synthetic load. Recorded as the direction, with the caveat that it is a
larger piece of work than a registry entry.

The machine's power scheme was recorded with powercfg before the
experiment, switched, and restored; confirmed Balanced afterwards from
powercfg rather than from simon.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>