Add Neptune backbone + masked point modeling (arXiv:2510.01733)

[sevmag]

What

Adds the method from "Reducing Simulation Dependence in Neutrino Telescopes with Masked Point Transformers" (Yu, Kamp & Argüelles, arXiv:2510.01733) as SPINE plugins, so it can be compared to CURTAIN on a shared backbone — isolating the pretext from the architecture.

• NeptuneBackbone (src/spine/backbones/neptune.py) — the paper's per-hit encoder: per-hit content MLP (charge only) + learned additive 4D positional MLP + a plain transformer + masked-mean readout, no FPS. It mirrors the paper's reference implementation (the prometheus branch), not upstream main (which has since moved to parameter-free 4D RoPE + FPS patchification) — pinned to the published architecture so comparisons stay faithful to the paper's method. encode honors an optional batch["pos_mask"], swapping masked hits' coordinates for learned mask embeddings before the positional MLP.
• MPMTask / MPMHead (src/spine/pretrain/mpm/) — masked point modeling, 1-to-1 with the paper: mask a fraction of hits' space-time coordinates (spatial / temporal / spatiotemporal), reconstruct them per token with smooth-L1. Token content is position-free so the reconstruction is non-trivial. Rides the existing collate/engine contract — no engine or CURTAIN changes.
• EncodedEvent.cls is now optional (Tensor | None = None). Per-token pretexts (MPM) read tokens and ignore cls; the only consumer that needs it (CURTAIN's QueryCrossAttnEncoder) now guards for None with a clear error. Existing backbones still fill cls, so behavior is unchanged.
• Configs: configs/backbone/neptune.yaml, configs/task/mpm.yaml.

Why

To answer, cleanly, whether CURTAIN's causal light-front forecasting pretext beats generic masked reconstruction — by running both on the same backbone with the same fine-tune, so the pretext is the only variable. (The two backbones' inductive biases interact with the pretext: DeepIce's learned Minkowski attention bias fights time-masking, whereas Neptune's additive positional MLP + mask embeddings tolerate it — which is exactly why a shared, neutral backbone is the fair place to compare pretexts.)

Validated

Unit smokes (synthetic batches, run on a GPU node):

• Backbone: correct shapes, masking measurably changes the encoding across all three modes, grads flow to the mask embeddings.
• MPM: all three modes compute, backward reaches the reconstruction heads and the mask embeddings, and the make_sample / collate data path runs end-to-end.
• EncodedEvent constructs without cls.

Scope / follow-ups (not in this PR)

• Fine-tuning a Neptune-pretrained encoder (spine_ft.py --backbone neptune) — the fine-tune harness currently hard-codes a graphnet DeepIce.
• A full train_curtain.py backbone=neptune task=mpm pipeline run.

Opening as draft because the end-to-end train → fine-tune → compare path isn't exercised yet; the plugins themselves are self-contained and validated.

🤖 Generated with Claude Code

[sevmag]

Re-scoped: #15 bundled the Neptune backbone + MPM pretext with the cls-optional change. Closing in favor of #16 (cls-optional only). The Neptune/MPM work stays on the neptune-mpm-backbone branch for a separate PR if wanted.