Add IDEA simulation and MLPF pipeline integration

[jpata]

Summary

Adds end-to-end FCC-ee IDEA support to MLPF:

• simulation, reconstruction, production configuration, and Snakemake workflows
• detector-aware track and calorimeter postprocessing, including separate Cherenkov and scintillation features
• TFDS datasets, training and evaluation recipes, and regression tests
• reproducible k4geo and k4RecCalorimeter integration checks

The PR also adds CLD particle-gun validation samples and updates shared pipeline configuration used by CLD and IDEA.

Current limitations

This pipeline is intended for engineering validation, not physics-performance studies. Tracks are truth-seeded and baseline candidates use target information, so the resulting metrics are not yet suitable for physics claims.

IDEA calorimeter energies are currently uncalibrated. Validation therefore focuses on data representation, truth relations, and energy closure.

GGTF tracking is disabled by default because the MLPF workflow uses truth-seeded tracks and GGTF significantly increases runtime and memory use. It can be enabled with ENABLE_GGTF=1.

Key4HEP dependencies

The reproducible workflow targets Key4HEP 2026-04-08:

Component	Branch	Commit
k4geo	key4hep-2026-04-08-mlpf	b313130
k4RecCalorimeter	key4hep-2026-04-08-mlpf	25016a3
FCC-config	key4hep-2026-04-08-mlpf	cdc8b78
key4hep-sim	key4hep-2026-04-08-mlpf	c09ebd3

key4hep-sim@c09ebd3 pins the other three commits and was verified from a fresh recursive clone.

Upstream staging PRs

• k4geo: #1, #2
• k4RecCalorimeter: #1, #2, #3
• FCC-config: #1, #2
• key4hep-sim: #2

Validation

• IDEA pipeline tests: 25 passed (tests/test_postprocessing_idea.py, tests/test_idea_pipeline.py, and tests/test_interleaved_shard_sampler.py)
• repository pre-commit hooks pass
• k4geo and k4RecCalorimeter build and install against Key4HEP 2026-04-08
• FCC-config compiles; workflow and component scripts pass bash -n
• a fresh recursive clone of key4hep-sim@c09ebd3 resolves to the pinned commits
• PODIO validation passes on a real IDEA event with 4,371 clusters and 4,405 cluster-truth links

Initial training results

Attention versus HEPTv2 on the initial dataset: