test(cli): subprocess integration tests for opencode acp

[kitlangton]

Summary

Adds the second long-lived-command builder to the cli-process harness:
opencode.acp(opts) spawns the real CLI in JSON-RPC-over-stdio mode and
returns a duplex handle (send/receive/close/exited) scoped to the
test's lifetime. Stdin EOF triggers a clean shutdown; the scope finalizer
falls back to SIGTERM after 2s for a hung child.

ACP frames each JSON-RPC message as one ndjson line on stdout. The builder
forks a scope-bound Stream.fromReadableStream + splitLines pipeline that
feeds parsed responses into a Queue.unbounded, so tests can yield* acp.receive
without worrying about backpressure or framing.

Two smoke tests:

• `initialize` round-trip — sends the protocol handshake from `src/acp/README.md` and asserts the response advertises the same protocolVersion and a non-empty agentCapabilities block.
• Clean shutdown on stdin EOF — proves the scope finalizer's `stdin.end()` triggers a graceful exit, not a SIGTERM-fallback exit.

Why this matters

Nothing currently exercises the ACP stdin/stdout protocol end-to-end. A regression that broke the framing, the handshake, the agent capabilities block, or the stdin-EOF clean-shutdown path would slip through every other test tier. The handshake test alone catches a huge class of "ACP doesn't even start" regressions that today only manifest when a user wires up an editor integration and it fails silently.

Stacked on #28263

This branches off the serve-builder branch since auto-merge on #28263 is still waiting on Windows checks. Once #28263 merges, the diff here shrinks to just the acp changes (the serve changes will already be in dev).

Test plan

• `bun run test test/cli/acp/acp-process.test.ts` — 2/2 pass locally in ~4s
• `bun run test test/cli/` — 323/323 pass, no regressions
• `bun run typecheck` clean

Stack

1. test(cli): subprocess integration tests for opencode acp #28265 👈 current
2. test(cli): help-text snapshots for every CLI command #28267