Token/cost usage meters and fan-out budgets

[liam-russell]

Problem

A five-worker fan-out (#137) spends real money invisibly. Nothing shows what a session, a subtask, or a whole orchestrated task (#134) has cost, and nothing stops a looping worker from burning budget all night.

Proposal

• Capture usage from ACP session updates where the agent reports it (best-effort; PTY-mode agents may report nothing — show "unknown" honestly rather than guessing).
• Meters: per-session cost/tokens in the chat header and sessions panel; rolled up per orchestration group in the stacked view (Stacked-work view: group workers under their orchestration branch, review merges, roll-up PR #136).
• Budgets: optional cap on a fan-out — at the threshold, pause remaining workers, raise an attention-inbox item (Attention inbox: one surface for everything that needs the human #146), and let the user top up or stop.
• Pricing data from the provider registry's model catalogs (AI provider registry + universal searchable model picker (OpenRouter, OpenAI-compatible, Ollama) #144 — OpenRouter's catalog includes pricing).