# Byte-exact benchmark inputs.
#
# `bench/rot/select.py` pins the corpus frame by sha256, and that hash is the
# whole reproducibility argument: a third party with these bytes derives the
# same repository ordering, with no API access and no trust in us. Git's EOL
# translation breaks that — a Windows checkout rewrote the file's single LF to
# CRLF and the hash check correctly refused it. Marking the frame binary keeps
# the committed bytes verbatim on every platform, so the anchor is genuinely
# platform-independent rather than platform-independent-on-Linux.
bench/rot/frame/*.json -text

# DELIBERATELY ABSENT: any `*.py diff=python` rule.
#
# Git's default funcname heuristic — not the Python driver — is what produces
# the `@@ ... @@` section headings this repository's history contains, and
# `bench/rot`'s claim-level detector is documented against that behaviour. It
# does not READ those headings (they label a method-body edit inside `Store`
# as `class Store:`, which would make a heading-keyed detector a body-churn
# amplifier), but the measurements and the reasoning in bench/rot/README.md
# describe the default heuristic's output. Turning on the Python driver would
# silently change what a future reader sees when they check that reasoning
# against `git log -p`.
