chore: move test fixtures into tests/, flatten scripts/, drop paper source

.gitignore

@@ -1,32 +1,16 @@
-__pycache__/
-*.onnx
-
-# Legacy run directories at the repository root
-/trainings/
-/scores/
-/slurm/
-
 # Python tooling
+__pycache__/
 *.py[cod]
-.pytest_cache/
-.ruff_cache/
-.mypy_cache/
 *.egg-info/
 .venv/
-venv/
-htmlcov/
+.pytest_cache/
+.ruff_cache/
 .coverage
-profiles/
-*.trace.json
 dist/
 build/
-benchmark-results/
-validation_output/
 
 # Local data and generated outputs
-data/raw/*
-!data/raw/.gitkeep
-data/processed/
+data/
 outputs/
-checkpoints/
 logs/
+*.onnx

AGENTS.md

@@ -11,7 +11,7 @@ run from `cli.py` to the model in a few files.
 src/gnn4colliders/   flat modules; see docs/design.md for what each holds
 configs/             one self-contained YAML file per experiment
 scripts/             Slurm wrappers and small helpers (no core logic)
-tests/               pytest; fixtures in data/fixtures/testing
+tests/               pytest, with small ROOT fixtures in tests/fixtures
 Pretrained_GNN/      the published pretrained checkpoints (Git LFS)
 ```
 

data/raw/.gitkeep

@@ -1 +0,0 @@
-

docs/paper/paper.tar.gz

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:9587a5b67b73dfc0f88b2a641b98410623a592d47cedd8fc71b35d28e9e1dfbc
-size 142839

docs/usage.md

@@ -101,7 +101,7 @@ allocation and, on Perlmutter compute nodes, `UV_CACHE_DIR=$SCRATCH/.cache/uv`
 ```bash
 # One training per seed, each followed by test-split evaluation:
 for seed in 1 2 3 4 5; do
-  sbatch -A <account> --export=ALL,SEED=$seed scripts/slurm/train.sh train.fast=true
+  sbatch -A <account> --export=ALL,SEED=$seed scripts/train.sh train.fast=true
 done
 uv run python scripts/summarize_ensemble.py outputs/gamgam/seed*   # mean +- std
 ```
@@ -110,5 +110,5 @@ Under `torchrun` with more than one process, training is data parallel:
 `data.batch_size` is per process, training events are sharded over ranks
 (padded to equal length) and metrics are computed over all ranks' events.
 CUDA graphs need a single process (so `train.fast=true` does too; DDP is
-tested in float32 eager). `scripts/slurm/train_ddp.sh` launches all GPUs of
+tested in float32 eager). `scripts/train_ddp.sh` launches all GPUs of
 one or more nodes.

scripts/summarize_ensemble.py

@@ -1,7 +1,7 @@
 """Summarize an ensemble of training runs as mean +- standard deviation.
 
 Each run directory must contain ``history.jsonl`` and ``test_metrics.json``
-(written by ``scripts/slurm/train.sh``). The standard deviation
+(written by ``scripts/train.sh``). The standard deviation
 is over runs, as in Table 3 of Ho et al. (2026).
 """
 

scripts/{slurm/train.sh → train.sh}

@@ -1,6 +1,6 @@
 #!/bin/bash
 # Train CONFIG on one GPU, then evaluate its best epoch on the test split.
-#   sbatch -A <account> --export=ALL,SEED=1 scripts/slurm/train.sh [key=value ...]
+#   sbatch -A <account> --export=ALL,SEED=1 scripts/train.sh [key=value ...]
 # The paper's ensembles vary only the seed (SEED=1..5). Prepare the event
 # cache first: uv run gnn4colliders prepare configs/gamgam.yaml
 #SBATCH -N 1

scripts/{slurm/train_ddp.sh → train_ddp.sh}

@@ -1,6 +1,6 @@
 #!/bin/bash
 # Data-parallel training on all GPUs of one or more nodes:
-#   sbatch -A <account> -N 2 scripts/slurm/train_ddp.sh [key=value ...]
+#   sbatch -A <account> -N 2 scripts/train_ddp.sh [key=value ...]
 # data.batch_size is per GPU. CUDA graphs need a single process.
 #SBATCH -N 1
 #SBATCH -C gpu

tests/conftest.py

@@ -9,7 +9,7 @@ from gnn4colliders.config import Config, load_config
 from gnn4colliders.data import PackedEvents, pack_events, prepare
 
 ROOT = Path(__file__).resolve().parents[1]
-FIXTURES = ROOT / "data/fixtures/testing"
+FIXTURES = ROOT / "tests/fixtures"
 #: 64-event fixtures of the five GamGam samples, labels 0..4 (ttH..ZH).
 GAMGAM_FIXTURES = sorted((FIXTURES / "atlas_opendata_gamgam").glob("mc_*_64.root"))
 TTH_FIXTURE = FIXTURES / "ttH_NLO_64.root"