[FEATURE]: emit agent_error session/update from acp/agent.ts when LLM call fails after retries #26379

Feature hasn't been suggested before.

• I have verified this feature I'm about to request hasn't been suggested before.

Describe the enhancement you want to request

This is the producer-side companion to the typed AgentErrorUpdate session/update kind: actually emit the frame from acp/agent.ts handleEvent() when a non-tool LLM call fails after retries.

Background

The agent's event loop in acp/agent.ts handleEvent() already routes message.updated and tool-call events to the connected ACP client. Currently, when session/processor.ts halt() writes an error field on the assistant message after retries are exhausted, the agent loop has no case for surfacing it as an ACP frame — so the failure stays local to the agent process and the client sees a stalled turn.

Proposal

Add a case "session.error": (or equivalent) branch in handleEvent() that converts halt()'s recorded error into an AgentErrorUpdate frame and dispatches it on the existing session/update channel. With the typed kind in place, this is a small wiring change: read the error category from the recorded state, build the payload, send.

Out of scope

• Defining the wire contract for AgentErrorUpdate itself — that should land first as a type-only addition (the closed-by-bot-but-still-relevant discussion in the related typed-kind issue covers the rationale).
• The classification logic on the producer side (proxy / model adapter that decides "this is a budget error vs rate_limit vs unknown") — out of scope for this change; consumers can fall back to unknown until the producer side is wired.

Why this matters

Without this emit, an ACP client connected to opencode has no programmatic way to render "retry exhausted, call failed" — it has to rely on its own watchdog timeout, which is brittle and produces a UX of "the agent just stopped responding." Tool-call failures already have a typed channel; turn-level errors deserve the same.