Swap Agentd persistent-agent spawner from pi to Omni/Agent

Goal

Replace pi --mode rpc with agent (no args, stdin-mode) as the persistent agent runner in Omni/Agentd/Daemon.hs. This unifies persistent and one-shot runs under a single binary and gives persistent agents access to agentd's replay/debug tooling.

Current state

Omni/Agentd/Daemon.hs spawns persistent agents as pi --mode rpc, holds the stdin/stdout pipes in an in-memory RunningAgent map, and parses pi's bespoke JSONL schema from stdout.

New state

• Spawn agent (via whatever path lookup already exists for oneshot runs) with no script argument, so it enters stdin-prompt mode.
• Hold the stdin pipe the same way.
• Writing a prompt = hPutStr stdin promptText >> hPutChar stdin '\NUL' >> hFlush stdin.
• Cancel = signalProcess sigINT agentPid (no stdin message).
• Shutdown = close stdin (EOF) or send sigTERM.
• Stdout parser changes from pi JSONL to Trace.Event JSONL — which is the same format the oneshot path already parses, so this should share code with the existing oneshot event reader.

Acceptance criteria

• Creating a persistent agent via agentd's existing API (POST /agents) spawns agent (not pi) and the agent accepts prompts via null-delimited stdin writes.
• POST /agents/<id>/send successfully delivers a prompt and the agent responds.
• POST /agents/<id>/stop cleanly terminates the agent.
• Cancellation mid-prompt works via SIGINT (requires t-759.2).
• agentd replay <run-id> and agentd debug <run-id> both work for persistent agents — same as they already do for oneshot, because they read Trace.Event from the same event log.
• Existing persistent agent metrics (cost, tokens, iterations) still populate correctly, sourced from Trace.Event instead of pi JSONL.

Depends on

t-759.1 (stdin mode) and t-759.2 (signal handling) must land first.

Files likely to touch

• Omni/Agentd/Daemon.hs — main spawn code and stdin/stdout handling for RunningAgent
• Possibly shared event-parsing helpers between oneshot and persistent paths