[ataniz]

Fixes from the find-opening and track-launch journey traces:

• Skills read the tag tree once per store, stop at the first budget refusal, and type the full CLI command.
• They don't write their own fetch client with retries.
• Every skill writes its Runs line without asking.
• find-opening reads the chart when a category maps to one.
• track-launch words delisting risk only as pattern, lean or no signal, and reads more go-lives with an account.

shared/data.md drops the removed minute cap. check-contract reads the budgets' new wording, which had made the check fail on main.

The scripts and the eval move to TypeScript. Node runs them directly and CI type-checks them.