evidence(perf): re-record both pairs under the D7 §6 and reference-identity repair

The digest moved with 637a619, so the previous artifacts certify an instrument
that no longer exists. Both halves of both pairs re-measured on that one clean
commit, written outside the repository under their final names, verified on all
ten PAIR_IDENTITY_FIELDS, then copied in and committed together.

The tenth axis is new and is the point: python_reference_tree. The reference now
reads 20879db while the tree reads 637a619 — a reference identity and a
repository position are finally two different things, so an evidence commit no
longer moves the reference and a D7 freeze no longer invalidates the binding it
just wrote.

Both pairs passed this time. Running tally: n=5 has failed twice and passed five
times; n=15 has failed twice and passed three times. The note records that every
re-record was forced by a provenance repair and never sought for a better
answer, because the raw counts drift toward passing simply because repairs keep
requiring fresh runs, and that drift is not evidence about the workload.

No calibration of record is restored. Round 6 has not run, and a passing pair
does not change what the record requires.

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

Author: claude