test(benchmark): calibrate again after the protocol and judge against the slower value

The relative gate calibrated once before the scenarios, so load that
started afterwards (the push and pull_request runs of one commit overlap
on the same runner, next to the test, build and image jobs) slowed the
scenarios but not the yardstick. A pull_request run failed `two years`
at 1653.1 ms against a 1648.7 ms budget this way.

The spec now repeats the calibration after the protocol and uses the
slower of the two values. The budgets can only loosen, never tighten,
and the log reprints the table with the budgets actually applied when
they change.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

docs/benchmark/README.md

@@ -24,6 +24,10 @@ Before the protocol, in the same process, the spec times the calibration engine
 a871f8f) on the `two years` scenario, best of three, and prints the result. This
 bundle never changes, so ratios stay comparable with the recorded history; the
 differential check's frozen bundle can be rebuilt independently.
+The spec repeats the calibration after the protocol and uses the slower of the two
+values, so load that starts during the protocol slows the yardstick too; the second
+value can only loosen the budgets. When it does, the log prints it and the summary
+table again with the budgets actually applied.
 Each scenario's budget is its multiplier times that calibration; the mean
 `planningMs` of its two measured repetitions must not exceed it. A scenario
 without a multiplier fails too. #286 tightened the multipliers once #302 met