.gitignore

# ── Rust ──
/target/
# Fuzzing state. COMMITTED: fuzz/seeds/ (hand-written inputs + every promoted
# crasher — the permanent tier-0 replay set) and fuzz/artifacts/ (raw crash
# inputs as the soak found them). LOCAL: fuzz/corpus/ is libFuzzer's working
# state — thousands of coverage-minimized blobs meaningful only to its own
# coverage map, regrown from the seeds on any machine.
/fuzz/target/
/fuzz/corpus/

# ── OS ──
.DS_Store
*.DS_Store

# ── Editors ──
.vscode/
.idea/
*.swp
*.swo

# ── Python venv / cache (used by helper scripts only) ──
.venv/
__pycache__/
*.pyc

# ── Large local artifacts ──
models/
assets/*.bin

# ── Test fixtures ──
crates/lumen-mlx/tests/fixtures/*.safetensors
# Exception: the vision-tower golden is 2.9 MB, not GB-scale, and is the only
# thing standing between `gemma4_vision_parity` and a Python dependency at test
# time. Committing it keeps that test pure Rust and runnable from a clean clone.
!crates/lumen-mlx/tests/fixtures/gemma4_vision_golden.safetensors

# ── Personal AI workflow tooling (kept local; not part of the public project) ──
.ai/
.claude/
.lean-ctx/
.outline/
AGENTS.md
AI-RULES.md
LEAN-CTX.md
ai-system-install-guide.md
claude.md
scaffold.mjs

# ── Workspace symlinks ──
/lumen-rs

# Qwen3.5 / Qwen3.6 MoE backend is now included as an opt-in feature
# (`cargo build --features qwen3_5_moe`). All source files are tracked.

# ── Local debug / investigation artifacts ──
worker-hang-investigation.html

# ── Personal benchmarking + dev scripts (kept local; not part of the
#    public project). ──
# `scripts/*` and not `scripts/`: git does not descend into an excluded
# DIRECTORY, so a `!scripts/golden/` re-include below would never be consulted.
scripts/*
# ...except the ORACLES. A golden nobody can regenerate rots the moment the
# model or the reference implementation moves — `flux-scheduler-invariants` was
# exactly that, a test comparing against a /tmp dump that had ceased to exist.
# So the scripts that PRODUCE or COMPARE-AGAINST a reference are committed
# alongside the goldens they make. The other ~47 files here are one-shot A/B and
# profiling artifacts and stay local. See crates/lumen-mlx/tests/golden/README.md.
!scripts/golden/
!scripts/golden/*.py

# ── Moltis 협조 데이터 (외부 프로젝트 reference data + raw 입력 로그).
#    Personal eval / pruning ground truth 용도. Public repo 업로드 금지. ──
raw_inputs_domestic/
raw_inputs_foreign/
standard_names/
moltis_data/