feat(agent): cancellation, timeouts, typed errors, and lifecycle helpers

[ThomasK33]

Why

An external team integrated @coder/ai-sdk-agent into a durable deep-research
fan-out (Vercel Workflow) and reported that the friction was concentrated in
cancellation, timeouts, quota visibility, error typing, and lifecycle. This PR
hardens those paths, then adds a quality + correctness pass on top.

What changed

Cancellation & timeouts

• abortSignal now interrupts the server-side run, not just the local
  WebSocket — a wedged turn stops generating and releases its workspace instead
  of running on, orphaned. Tearing down a stream() (cancel) interrupts too.
• New requestTimeoutMs: a per-segment time budget. On expiry the run is
  interrupted and the call rejects with a retryable CoderChatError
  (kind: "timeout") instead of hanging. Documented as per-segment, with
  abortSignal: AbortSignal.timeout(ms) for total wall-clock.

Typed, retryable errors

• A stream that ends before the turn settles rejects with CoderChatError
  (kind: "stream_closed", retryable) rather than reporting a truncated result
  as a clean stop.
• Transport failures and server error events are classified into typed,
  retryable errors (no more bare DOMException/non-retryable transport errors
  bypassing the documented retry contract).

API & lifecycle

• CoderAgent.listModels() to discover valid model hints.
• await using support via Symbol.asyncDispose (interrupt + archive on scope
  exit), so cleanup rides scope exit instead of a forgotten finally.
• A warning when responseFormat/JSON-schema output is requested — chatd does
  not constrain output server-side; use @coder/ai-sdk-provider +
  generateObject instead.

Docs

• Agent-vs-provider decision guide (agent/provider/root READMEs), error handling,
  workspaces & quota, a durable-workflow recipe, and timeout/cleanup semantics.

Process

Built from the integration feedback, then refined with a /simplify pass and a
max-effort /code-review pass. The review caught a real regression the simplify
pass had introduced (a caller's own AbortSignal.timeout being misclassified as
the agent's timeout) plus several edge cases in the new error/interrupt paths —
all fixed here, each with a regression test.

Testing

• 49 unit tests in @coder/ai-sdk-agent (8 new, covering abort→interrupt,
  timeout, stream-close, transport error, caller-timeout classification, cancel,
  and the responseFormat warning); 178 across the repo. All green.
• pnpm check (format + lint + typecheck) and pnpm build pass.

Known follow-ups (out of scope)

• A caller-aborted createChat leaves a chat whose id we never received, so it
  can't be interrupted — needs client-generated ids or a reconcile sweep.
• The requires_action safety-counter (pre-existing) can finish a turn as
  tool-calls with no tool calls if action_required never arrives.

🤖 Generated with Claude Code