fix(acp): return stopReason: cancelled (not end_turn) on user cancel

[truenorth-lj]

Issue for this PR

Closes #25899

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Returns stopReason: "cancelled" instead of "end_turn" on the ACP session/prompt RPC when the turn was interrupted by session/cancel, and omits the misleading all-zero usage field on cancel.

Agent.prompt() currently resolves the ACP RPC with stopReason: "end_turn" whether the turn finished naturally or was interrupted. ACP defines a dedicated stopReason: "cancelled" for the cancel case, so reporting end_turn makes a user cancel indistinguishable from a clean completion.

Cancelled turns can also return usage: { totalTokens: 0, inputTokens: 0, outputTokens: 0 }. The LLM stream is killed before the AI SDK emits its finish-step usage event, so the assistant message tokens are still at their zero initializer. Reporting zero is misleading; omitting usage lets clients treat it as unknown.

The fix detects msg?.error?.name === "MessageAbortedError" and uses that at both Agent.prompt() return sites that return an assistant message:

const cancelled = wasCancelled(msg)
return {
  stopReason: cancelled ? ("cancelled" as const) : ("end_turn" as const),
  usage: msg && !cancelled ? buildUsage(msg) : undefined,
  _meta: {},
}

The command-only return site is unchanged because it has no assistant message to inspect.

A larger follow-up could recover provider usage during interrupt cleanup, but this PR intentionally keeps the fix scoped to reporting the correct stop reason and avoiding known-bad zero usage.

How did you verify your code works?

• bun test test/acp/event-subscription.test.ts - 12 pass / 0 fail
  • prompt() returns stopReason: cancelled with no usage when message has MessageAbortedError
  • prompt() returns stopReason: end_turn with usage on normal completion
• bun run typecheck - clean

Screenshots / recordings

N/A - backend protocol fix.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[truenorth-lj]

@rekram1-node — would you have a moment when you're next in the ACP queue? Same area as the end_turn race fix in #25683 (still open, also pinging there) — both are about stopReason correctness on session/prompt completion.

This one returns cancelled rather than end_turn when the user cancels mid-turn, so ACP clients can distinguish a user-aborted turn from a clean completion. CI green, no conflicts. v2 of #25901 (template-bot auto-close).

Happy to add a regression test or reshape if anything's off.