fix(pi): render completed messages and errors - #1495

Reviewed in depth before merge, including against the installed Pi package's actual RPC internals: message_end is authoritative and always follows deltas, but the existing first-text dedup guard in useAIChat (the same architecture the Claude provider uses) means normally streamed responses render byte-identically - no double-render. The no-delta path is real (agent-loop's !addedPartial branch), and the error branch targets the only place current Pi actually reports errors: the old message_update error mapping is dead code on current Pi, which is exactly why quota/auth errors rendered as blank successful answers. Verified the error path can't wedge the SSE drain (message_end → turn_end → agent_end still terminates normally) and the error field survives the subsequent result. Revert-verification: two of your three tests genuinely fail against main.

One follow-up we'll track on our side, not yours: in a no-delta session with tool use, the first-text guard keeps only the first assistant message's text - strictly better than today's empty response, and the same shape the Claude provider lives with, but a text-replaces-on-message-boundary refinement in useAIChat would improve both providers. Merging - excellent first contribution.