ACP client, then cursor and grok drivers over it

[chrisbbreuer]

cursor-agent and grok both speak ACP, so the client unlocks two of the three
remaining providers and should come before either driver.

Blocked: neither CLI is installed on the development machine. A transport
written against nothing is unverifiable code, which is why the drivers are
currently registered-but-unimplemented rather than sketched.

What exists already

• packages/drivers/src/conformance.ts — the spec every driver must pass. A new
  driver is done when ./buddy harness:doctor --conformance --driver <kind> is
  green.
• packages/drivers/src/cli-driver.ts — cursor and grok are registered and
  probe honestly today ("not installed", or "installed but the driver is not
  implemented yet"). They must never report ready until they can actually run
  a turn, or a session starts and dies on its first turn.
• packages/drivers/src/codex.ts — the closest reference. It is written against
  the protocol schema the CLI generates for itself rather than against guesswork.

Watch for

Both findings from the codex driver are likely to recur:

1. Terminal-event coalescing. Codex signals a failed turn twice — an error
   notification, then a failed turn/completed. Emitting on arrival ends the
   turn twice. The conformance suite catches this ("exactly one terminal event").
2. Wire enum spelling. Codex rejected onRequest where it wanted
   on-request, and only after a thread had already started. Pin the exact
   strings in a test.

Done when

harness:doctor --conformance is green for the driver against the real CLI, and
a turn with a tool call produces session.provider-bound, tool.call.began,
tool.call.ended, turn.completed in the log.