.gitignore

# Rust
target/
# Anchored to the repo root: the generated llvm-cov report directory. An
# unanchored `coverage/` matches ANY directory of that name at any depth,
# which silently excluded `crates/larql-vindex/src/extract/coverage/` —
# a source module — from `git add`. Keep the leading slash.
/coverage/
**/*.rs.bk
.cargo/

# Python / maturin
.venv/
__pycache__/
*.pyc
*.pyo
*.so
*.dylib
*.egg-info/
dist/
build/
*.whl
# Don't ignore Rust source modules named `build/`
!crates/*/src/**/build/

# IDE
.idea/
.vscode/
*.swp
*.swo
*~
.DS_Store
.claude/

# Checkpoints / temp
*.checkpoint

# Weights and binary model files — never commit these
*.npz
*.npy
*.safetensors
*.gguf
*.ggml
*.pt
*.pth
*.ckpt
*.bin
*.pkl
*.h5
*.hdf5
*.vindex

# ...except safetensors HEADER STUBS under a test fixtures directory. A
# stub is a real upstream header with its payload truncated away — tens
# of KB standing in for gigabytes — and it has to be in the repo, because
# a range source's gates are only meaningful against metadata that came
# from an actual checkpoint. `granite_tied_weights` asserts each stub is
# under 100 KB, so this exception cannot quietly grow into real weights.
!crates/larql-vindex/src/format/vindex3/encode/source/tests/fixtures/**/*.safetensors

# Nix
result
result-*

# output
output/
data/
experiments/
vindexes/
.pids/
docs/replay/

# DEC-0 loadgen outputs (bench/dec0/prompts.txt stays tracked)
bench/dec0/residuals-*/
bench/dec0/*.log
bench/dec0/*.jsonl
bench/dec0/anchor_*.json
bench/dec0/dec0_*.json

# Experiment run logs written to the repo root by the EXP-* examples.
# Transient stdout captures, regenerable by re-running the example; the
# findings live in docs/ and the experiments registry, not here.
/run_*.log

# quant-obs calibration captures — regenerable via `larql dec-bench capture`
bench/quant-obs/residuals-*/
bench/quant-obs/*.log

# SENSITIVITY moment captures: ~114 MB each, dominated by the strided
# `ffn_samples` the down_proj reconstruction needs. Not banked — they are
# reproducible from the container plus the frozen token bank, and every
# consequence record carries the capture's `moment_artifact_digest`, so
# provenance survives without the bytes.
bench/prompts/quality-bank-1/*-moments-*.json
# R4.0-CAL-A's downloaded Gutenberg source cache — not banked. Each
# source is sha256-pinned in r4_cal_a_build.py, so `verify` re-downloads
# and checks rather than trusting a vendored copy; the ~3MB of raw text
# adds nothing a re-fetch-and-check doesn't already guarantee.
bench/prompts/quality-bank-1/.r4-cal-a-source-cache/