Integrate a foundation model

Use the official package, loader, configuration, and released weights. Do not reimplement the pretrained network, copy its source, convert checkpoints without a demonstrated need, or bundle weights in ModernTSF.

1. Confirm the primary paper, official repository, package version, checkpoint identifier, pinned revision, license, input semantics, outputs, and supported horizons. Treat an architecture-only local model as a different claim.
2. Check whether the official package can coexist with the main environment. If dependencies conflict, keep it in a compatible provider environment and use the same src/models/_foundation/ boundary; do not relax core dependencies.
3. Reuse FoundationModel and the closest official runtime adapter. Load only from an explicit local path with offline mode enabled. Never trigger a network request during model construction, verification, or an experiment.
4. Add one ordinary flat src/models/<slug>/ entry. Its factory receives verified local artifacts, declares inference-only, and exposes the canonical four-input interface. Do not add a provider registry or a foundation category.
5. Document official behavior, checkpoint facts, cache preparation, preprocessing, channel treatment, limitations, and every adapter transformation in the model card. Keep upper-layer methods as their own flat entries and compose them only through explicit experiment configuration.
6. Verify offline failure, official reference outputs, tensor axes, quantiles, finite values, CPU behavior where supported, batch/sequence boundaries, and state loading. Training and gradient checks are not-applicable only because the declared runtime is inference-only, not because verification was skipped.

Stop if the license, official loader, checkpoint identity, or input/output semantics cannot be established. Adding the interface alone does not admit a provider as a catalog model.