.gitattributes

# Language statistics overrides (GitHub Linguist).
#
# Nothing in this file deletes, ignores, or hides anything — every path below
# stays fully tracked, diffed and reviewed. The only effect is on the language
# bar GitHub prints at the top of the repo page.
#
# The reason it needs saying: this project gates every Rust port on a reference
# oracle, and every reference implementation worth gating against — Ultralytics,
# PaddleOCR, SVTR, CRNN, MobileDet, Piper, Whisper — is Python. So each port
# arrives with a Python twin that dumps the fixtures, produces the oracle
# digests the Rust tests read, and scores the external baseline. As of the
# Diana and Carmenta campaigns that is 65 files and 358 KB, which is ~11% of
# the bytes Linguist counts, which puts Python second on the bar of a repo
# whose whole premise is that it is Rust. The bar was measuring the evidence
# layer, not the deliverable.
#
# `linguist-detectable=false` is the honest attribute for that. `-vendored`
# would claim this is third-party code we merely carry (it is first-party, and
# written here); `-generated` would additionally collapse these files in diffs,
# which is exactly wrong for scripts whose review is how a fixture gets
# trusted. Detectable-false says the one true thing: not part of the shipped
# artifact, so not part of its language stats.
#
# SCOPED TO DIRECTORIES, NOT TO `*.py`. A blanket `*.py` rule would also
# silence any Python that ever lands inside a crate — and Python inside a crate
# is precisely the thing that should be loud. These two directories hold no
# shipped code by construction: tools/ is 44 .py + 3 .sh + 2 .ps1 harnesses and
# nothing else, corpora/refs/ is 21 .py alongside the fixtures they generate.
# If that ever stops being true, this rule is wrong and should be narrowed.

tools/**        linguist-detectable=false
corpora/refs/** linguist-detectable=false

# Fuzz corpora are BYTES, not text.
#
# Without this, git's autocrlf (true on the Windows box this was authored on)
# treats a corpus input with LFs and no NUL as a text file and rewrites the line
# endings on checkout. That silently corrupts the input: the fuzzer's filename is
# the SHA-1 of its content, so a mutated file no longer matches its own name, the
# regression replay stops testing what it was seeded with, and coverage-guided
# runs restart from inputs that no longer reach what they were minimised to reach.
#
# The stored blobs were verified intact when this was added; this keeps them that
# way through every future clone.
crates/*/fuzz/corpus/** -text -diff

# Binary oracle fixtures, same reasoning.
*.bin -text -diff
*.npy -text -diff
*.safetensors -text -diff