Extract paper structure

Use the primary paper and supplement. Locate official code when available, record its license, pin a revision, and inspect it to resolve paper omissions. Keep paper facts, official implementation details, and local design in separate fields; do not copy source text or code into the implementation map.

Produce a compact implementation map containing:

• task, inputs, covariates, tensor axes, output, and probabilistic semantics;
• preprocessing, normalization, decomposition, embeddings, main blocks, head, and inverse transforms in execution order;
• defining equations with paper section or equation references;
• loss, auxiliary objectives, initialization, defaults, and train/eval differences;
• shape invariants, sequence constraints, marks or adjacency contracts, and edge cases;
• official-code clarifications with revision and source path when available;
• unspecified details and decisions that would materially change fidelity.

For every implementable operation, include a component decision: reuse-existing with the component name and matched contract, extract-new with expected consumers, or model-local with the semantic mismatch. Prefer reuse-existing whenever equivalence is established.

Separate paper facts from implementation inference. Match each operation to an existing component only after checking mathematics, axes, normalization, masking, residual order, initialization, state, and outputs with:

uv run tsf component match <operation-and-contract-terms> --json
uv run tsf component show <candidate>

Deliver the map independently of code. It is complete when another worker can implement or audit the model without guessing a defining operation. Stop and surface ambiguity when choosing silently would change the named method.