Resolve an AI-add figure from an intent

The AI-add omnibox: infer the chart kind (unless pinned), bind the fields via one Haiku call, clamp them deterministically, then resolve the figure through dash-render with FIELDS ONLY (never the intent — ADR D5; the LLM never authors the spec or numbers). Returns the resolved preview directly. Synchronous (not the async-job path chat/ask use): there is no canvas append and no bake to run — just one small Haiku call + one engine-route resolve — and the editor awaits this result to open the studio, so a job + poll would only add latency. The blocking LLM/profile/render work runs off the event loop.