feat(ai): structured tool-call display (AI Elements–style)

[os-zhuang]

What

Replaces the one-line Ran query_data caption with a collapsible, structured tool block — a React Native take on Vercel AI Elements' Tool component (which is web/DOM-only, so it can't be ported directly).

• lib/ai-chat.ts — the SSE parser now captures rich ToolInvocation objects ({ id, name, input, output, state }) by merging tool-input-available + tool-output-available events keyed by toolCallId (was: just names). onUpdate streams the live tools; toolCalls: string[] kept for back-compat.
• components/ui/ToolInvocations.tsx — per-tool collapsible row: tool name + running/done status, expanding to reveal the input args and the result (mono, scrollable).
• store messages carry tools: ToolInvocation[]; app/ai.tsx renders the block under assistant turns.

Why custom (not the library objectui uses)

Confirmed from objectui's source: its chat UI is Vercel AI Elements (vendored MIT, shadcn/DOM) on top of @ai-sdk/react useChat. Neither ports to React Native — AI Elements is web-only (<div>/Radix/streamdown/shiki), and useChat hits a React 19.2.0 peer gap on Expo 55. So we mirror the design in custom RN components.

Tests

• Parser captures {input, output, state} keyed by toolCallId — asserted against the real captured SSE tool format.
• ToolInvocations component: collapsed-by-default, expand reveals input/result, running hides result, one row per tool.
• Screen integration + store carry-through.
• npm run lint 0 errors; jest 1255/1255 (153 suites).

Verification note: tested at every layer against the real server's SSE shape; the visual block itself is standard NativeWind (not screenshotted this round — MMKV/native-gated screens aren't browser-testable, but this one could be device-verified).

🤖 Generated with Claude Code