Add a model

Use the flat src/models/<module>/ layout. Models and methods are peers; do not create family directories.

1. Resolve the public name, lowercase module slug, parameters, input needs, output type, paper identity, and official-code facts when available.
2. Scaffold an unregistered workspace with paper facts and the component decision, for example uv run tsf model scaffold --name MyModel --paper-title "..." --paper-url <url> --venue <venue> --year <year> --components revin --params "enc_in:int,hidden:int=128". Select --task-mode explicitly when it is not ordinary time_series; provide code URL, revision, and license together when official code exists.
3. Before implementing any block, run uv run tsf component match <requirements> --json for each operation in the paper structure map and inspect promising candidates with component show. Record each operation as reuse-existing, extract-new, or model-local. When an existing component is mathematically and contractually equivalent, reuse it instead of creating a local copy. Retrieval is only a shortlist: verify shapes, axes, normalization, masking, residual order, initialization, state, and outputs first. Replace the placeholder in model.py; never import another named model package as an implementation dependency.
4. Complete spec.py with only the factory, parameter schema, config path, and runtime contract. Keep the parameter schema strict and expose exactly forward(self, x_enc, x_mark_enc=None, x_dec=None, x_mark_dec=None); do not accept catch-all arguments. Put descriptive and provenance facts in README front matter.
5. Complete the model card's method, structure, inputs/outputs, paper and codebase links, local implementation, differences, shared components, and constraints.
6. Add focused paper/equation and reference checks to verification/models.toml. Run those tests directly, remove every scaffold marker, then admit atomically with uv run tsf model add --name MyModel. Admission temporarily registers the model, executes unified verification and all audits, and rolls registration back on failure.

Success requires one indexed model card and runtime spec, a preset, passing unified verification evidence, truthful source and artifact facts, and a component decision for every defining operation. Declare every reused component in spec.py and list it in the model card. Use implement-model for the implementation before accepting the scaffold; use integrate-foundation-model instead for a released pretrained foundation runtime. Generated placeholder code is never a catalog entry.

Use curate-components when the task is broader consolidation across existing models; do not expand a single-model addition into an unsolicited refactor.