run: drain piped child stdout/stderr to prevent Ralph loop deadlock

[selimdev00]

Problem

In dashboard and swarm mode, spawnRalphLoop spawns the loop with stdio: ["ignore", "pipe", "pipe"], but nothing ever consumes the child's stdout/stderr — the dashboards read progress from .ralph/ state files, not from the pipe.

Node child-process pipes start paused. Once the loop (and the agent CLI it drives) writes more than the OS pipe buffer (~64KB), the child blocks on write() indefinitely. The loop then hangs forever: onExit never fires, so bmalph run --dashboard and bmalph run --swarm (where every worker is spawned this way) stall and never resolve. It only takes one verbose agent turn to cross 64KB.

Fix

When stdio is piped (inheritStdio === false), put both streams into flowing mode so their buffers drain and the child can always make progress:

if (!options.inheritStdio) {
  child.stdout?.resume();
  child.stderr?.resume();
}

This preserves existing behavior (dashboard mode already shows state-file progress, not raw child output) while removing the deadlock. The inheritStdio === true path is unchanged. Output is discarded, same as today — a follow-up could instead tee each stream to a per-worker log file under .ralph/logs/ for swarm debugging.

Tests

• Asserts stdout.resume() / stderr.resume() are each called once when inheritStdio is false.
• Asserts the inherited-stdio path (no stream handles) skips draining without throwing.
• Full unit suite passes (1837).