Sequence models from MC2

[wkirgsn]

Summary

This PR adds support for the sequence-to-sequence H-field models from the MagNet Challenge 2025 alongside the existing power loss models, exposing both model families through a unified public API.

Given a future B waveform and temperature — plus an optional warmup window of past B/H samples — the new SequenceModel predicts the corresponding future H sequence directly on variable-length waveforms (no fixed-length resampling).

New files

File	Purpose
src_py/magnethub/sequence.py	Public SequenceModel wrapper (mirrors LossModel)
src_py/magnethub/paderborn_sequence.py	Paderborn GRU backend + equinox model building blocks
models/paderborn_sequence_modeling/*.eqx	15 trained GRU8 coefficient files
tests/test_sequence.py	20 unit tests (validation, shapes, accuracy, all-materials smoke)
tests/test_files/unit_test_data_sequence_3C90.csv	Measured 3C90 toy data for accuracy regression test

Changed files

File	Change
paderborn.py → paderborn_loss.py	Renamed for clearer distinction from the sequence backend
__init__.py	Registers the new modules
loss.py	Updated import path after rename
pyproject.toml	Inline dependencies (drops requirements.txt), adds jax/equinox, bumps to Python ≥3.13, includes .eqx in sdist
README.md	Rewritten intro presenting both model families as first-class citizens; added usage examples, supported-materials list, and folder-structure entries

Usage

import magnethub as mh

mdl = mh.sequence.SequenceModel(material="3C90", team="paderborn")

# simplest call — past defaults to zero
h = mdl(b_future, temperature)

# with warmup for better accuracy
h = mdl(b_future, temperature, b_past, h_past)

Supported materials (sequence models)

3C90, 3C92, 3C94, 3C95, 3E6, 3F4, 77, 78, FEC007, FEC014, N27, N30, N49, N87, T37

Design decisions

• Argument order: (b_future, temperature, b_past=None, h_past=None) — the required inputs come first; warmup is optional and defaults to a single zero sample (GRU skips warmup internally when past length is 1).
• Merged helpers: The equinox building blocks (Normalizer, GRU, RNNwInterface, etc.) live directly in paderborn_sequence.py rather than a separate helpers module, keeping the file count manageable.
• A–E material mapping: Competition-internal labels A–E are mapped to their real material names (A=3C92, B=3C95, C=FEC007, D=FEC014, E=T37) so users always use human-readable names.

Test results

32 passed, 2 skipped, 1 warning

The 2 skips and 1 warning are pre-existing (debug-gated paderborn loss tests and a CUDA driver warning from the sydney model).