.gitattributes

# Signed-envelope fixtures are byte-exact inputs to signature verification:
# CRLF checkout conversion on Windows changes the signed bytes and breaks
# the generator-match and signature tests.
crates/aft/tests/fixtures/gh_shim/** -text
crates/aft/tests/fixtures/pty_render/*.raw binary
# Golden render outputs are compared byte-exact against `\n`-joined screen
# output, so they must stay LF on every platform (Windows autocrlf would
# otherwise check them out CRLF and break the deterministic comparison).
crates/aft/tests/fixtures/pty_render/*.xterm.txt eol=lf
# Truncation-contract goldens are compared byte-exact against rendered
# text (`\n` line ends) on every platform; the same Windows autocrlf rule.
crates/aft/tests/fixtures/callgraph/truncation/*.txt text eol=lf
# Bash-surface fixtures: input.txt is the compressor input and is compared
# line-for-line against the rendered output; keep LF on every platform.
crates/aft/tests/fixtures/bash/** text eol=lf
# The real-query vector pack and the digest-pinned JSON beside it are hashed
# byte-for-byte against pins captured from LF bytes (`embedding_pack_sha256`
# in the manifest, `manifest.sha256`, and the sibling definition digests
# imports_resolve.py checks). A default Windows checkout with
# `core.autocrlf=true` writes CRLF and fails those pins before the gate can
# run, so the checked-out bytes have to be the pinned bytes on every autocrlf
# setting.
benchmarks/aft-search/*.json text eol=lf
# The real-query vector pack is float16 data read byte-for-byte by its digest.
benchmarks/aft-search/*.bin binary
benchmarks/aft-search/manifest.sha256 text eol=lf