Add an end-to-end runnable graphnet demo on the bundled Prometheus data

[sevmag]

What

The graphnet connection made demonstrable end to end on data everyone has: one example script holding every line of demo-specific code, runnable both as plain Python and through the Hydra launcher.

examples/graphnet_demo.py — the full SPINE loop on the 50-event Prometheus example file graphnet bundles (data/examples/sqlite/prometheus/prometheus-events.db):

• Reader: a graphnet SQLiteDataset (EdgelessGraph + NodesAsPulses + the stock Prometheus detector with replace_with_identity, which keeps every column raw including sensor_id) adapted via spine_graphnet.readers.GraphNetRawDataset. sensor_id is globally unique in the file (503 sensors) and serves directly as the data-carried sensor key. Prometheus rows are single photons with no charge column, so a tiny wrapper appends unit charge to complete the (x, y, z, t, charge) layout.
• Scaling: spine_graphnet.scaling.DetectorScaler (new, general) turns any graphnet Detector's feature_map into a SPINE FeatureScaler — unmapped columns pass through, query positions go through the detector's xyz entries. The demo pretrains in the exact standardization graphnet's own examples apply to this file (Prometheus()), so a downstream graphnet fine-tune sees the same feature space; the demo defines no scaler of its own.
• Geometry asset: built from the file itself (distinct sensors, full sorted neighbour lists), loaded with load_geometry(..., sensor_key="sensor_id").
• Selection: events filtered with can_always_split under the same knobs the task samples with, split 28 train / 7 val.
• Task + backbone: occupancy-only CurtainTask, tiny DeepIceBackbone (92K params), trained via fit() with CurtainValAUC.
• The payoff: the exported checkpoint loads back into a stock graphnet DeepIce with strict key matching, ready for supervised fine-tuning.

Hydra path for the same run — the script always stages a self-contained demo dir (db copy, geometry, selection parquets) and has --prepare-only; the experiment config (introducing the configs/experiment pattern) then drives the identical run. The data group targets graphnet_demo.demo_reader directly — running python examples/train_curtain.py puts examples/ on sys.path, so no demo code needs to live in the installed package:

python examples/graphnet_demo.py --prepare-only
python examples/train_curtain.py +experiment=prometheus_demo

Core changes riding along:

• fit() passes accelerator="auto" instead of pinning "gpu", so the demo runs on CPU-only machines.
• Random holdout mode removed — the temporal cutoff is the pretext; every benchmarked run trained with it. The never-used random sensor split cost a second branch in sample_event and can_always_split, two pass-through CurtainTask arguments, and a dead random_vis_frac knob in every caller's selection filter. Net −40 lines, no behavior change for any existing config.

Verification

Both paths ran end to end on a CPU compute node after every commit, most recently: plain script (selection 28/7, 15 epochs in about a minute, best val loss 0.678, strict load_state_dict into stock DeepIce passed) and Hydra experiment (staged dir, trained, exported checkpoint val loss 0.682, stored config confirms the graphnet_demo.demo_reader target and the nested DetectorScaler(Prometheus())). All pre-commit hooks pass.

🤖 Generated with Claude Code