evidence(perf): the calibration of record, at n=15 on a clean tree

Recorded at ad4c302 with the tree clean, harness digest 2b36de7a7bc9 matching
the shipped instrument: 40 cells, all timed, all outcomes valid, admissible on
all three axes — outcomes_reproduced, timings_reproduced and environment_valid.

n=15 rather than n=5, and not because n=5 was red. A fresh n=5 pair reproduced
cleanly and is preserved alongside as sizing evidence, so both counts are
admissible on this machine. The record uses more samples because a report of
record should carry the better-estimated median, which was true before anything
went red. What the intermittent n=5 failure showed is that its admissibility is
luck-dependent here, not that the workload's spread is wide.

The first attempt at this report was discarded rather than committed: it carried
tree_dirty true, because writing the sizing evidence had modified the tree
before the run. A report of record with dirty provenance is not a report of
record.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CYLNQy6tLXqsV1CbuNqsSb

Author: claude