fix(agents): dedupe ReAct draft text per-round while preserving live streaming (#421)

[MarioCadenas]

Problem

With OpenAI-compatible Claude on Databricks Model Serving, the model emits a full draft answer ALONGSIDE its tool calls on every ReAct round, so the answer appeared 3-4x in the output. The earlier approach (branch fix/421-adapter-text-dup) buffered text and flushed it only on the terminal turn — but that made the final answer arrive all-at-once instead of streaming, which is unacceptable. This PR closes/supersedes that approach and references issue #421.

Root cause

The duplication was never in the adapter or the wire protocol. AgentEventTranslator already closes the current message item (response.output_item.done) when a tool_call/tool_result arrives and opens a NEW message item for the next round's text — so each round's text is already a distinct Responses-API output item with its own output_index/item_id. The duplication came from two consumers that concatenated deltas across those item boundaries.

The three-layer fix (respect the item boundaries)

• Adapter (databricks.ts) streams every round's text live (already matched main — no change needed).
• consumeAdapterStream keeps only the terminal round: a draft followed by a tool_call/tool_result is set aside as superseded, and the message open at end-of-stream (or the last draft if maxSteps is exhausted mid-tool-calling) is returned. This centrally dedupes thread history, the non-streaming JSON fullContent, runAgent, and sub-agents.
• useAgentChat reduces the wire events into an ordered, per-output-item list while streaming live. content now tracks only the LAST message item (the terminal answer), so flat-API consumers that only read content dedupe too.

New public hook API

export type AgentTurnItem =
  | { kind: "message"; id: string; text: string; status: "in_progress" | "completed" }
  | { kind: "tool_call"; id: string; callId: string; name: string; args: unknown; status: "in_progress" | "completed" }
  | { kind: "tool_result"; id: string; callId: string; output: unknown; error?: string };

UseAgentChatResult gains items: AgentTurnItem[] (ordered by output_index). The dev-playground agent route uses it to render a collapsible "Steps" disclosure (intermediate drafts + tool-call/result chips) above the prominent, live-streaming answer.

Tests

• consume-adapter-stream.test.ts: multi-round draft+tool_call returns the terminal text once; maxSteps-exhaustion returns the last draft; LangChain single message still replaces; mixed.
• use-agent-chat.test.ts: two message items separated by a function_call → items is [message, tool_call, tool_result, message], content === the LAST message text (not concatenated), deltas stream into the right item live.

All appkit (2147) and appkit-ui (319) tests pass; build, typecheck, biome, and knip clean.

This pull request and its description were written by Isaac.