# .gitattributes for the Hugging Face Space ONLY.
#
# The sync workflow copies this over the repository's own .gitattributes, the
# same way it swaps README_SPACE.md in for README.md. The two files cannot be
# merged, because the LFS rules below are true on the Space and false on GitHub.
#
# Why they differ
# ---------------
# Hugging Face stores any file above roughly 10 MB in LFS automatically, plus
# some extensions such as .pt regardless of size. GitHub stores those same files
# as ordinary blobs, because they were committed without LFS and fit inside the
# 100 MB per-file limit.
#
# A file is only reconstructed from LFS on checkout if .gitattributes declares
# a filter for it. Without these lines, the Space's Docker build checks out a
# 133-byte pointer instead of the real file, and `COPY data/gold` copies that
# pointer into the image. The container then fails with
#
#   IO Error: The file "/app/data/gold/gridpulse_app.duckdb" exists,
#   but it is not a valid DuckDB database file!
#
# Putting these lines in the GitHub .gitattributes instead would be wrong twice
# over: those files are not in GitHub LFS, and declaring them would push them
# into it on the next commit, spending LFS quota to solve a problem GitHub does
# not have.
#
# ORDER MATTERS. Within this file the LAST matching pattern wins for a given
# attribute, so the broad line-ending rule has to come first. Putting it last
# would set text=auto back on every binary declared below it, undoing the -text
# that keeps git from rewriting bytes inside a DuckDB file.

# 1. Line-ending policy, carried over from the repository .gitattributes.
#    entrypoint.sh is the container entrypoint: CRLF here means "bad
#    interpreter" and a Space that never starts.
* text=auto eol=lf
*.sh                  text eol=lf
Dockerfile            text eol=lf

# 2. Large binaries, reconstructed from LFS on checkout. These come last so
#    their -text survives.
*.duckdb              filter=lfs diff=lfs merge=lfs -text
*.parquet             filter=lfs diff=lfs merge=lfs -text
*.pt                  filter=lfs diff=lfs merge=lfs -text
*.pkl                 filter=lfs diff=lfs merge=lfs -text

# 3. LightGBM writes models as text. point.txt is about 35 MB and is stored in
#    LFS; the quantile models sit just under the threshold and are not. The
#    smudge filter passes non-pointer content through untouched, so one pattern
#    covers both cases and keeps working if a retrain pushes them over.
artifacts/**/*.txt    filter=lfs diff=lfs merge=lfs -text