Swarm await/summary reports terminated workers as done with status 'unknown' instead of failed

Split out of #512 symptom 3. The boot-time death causes from that issue are fixed (see #512 for the commit list); this reporting gap is the residual.

Problem

When a swarm worker process vanishes rather than exiting through a normal completion path, await_members reports it as done with status "unknown", and "All members done" fires while every status is still unknown. A coordinator therefore treats a dead worker as a finished one and proceeds to synthesize from missing output.

Expected

• A member whose process disappeared without reporting should settle as failed, not unknown.
• await_members should not count unknown toward "all done"; either keep waiting or resolve it to a terminal failure.
• The summary should distinguish "reported ready", "failed", and "vanished/no status" so a coordinator can retry or salvage.

Why it matters

This is the silent-failure half of #512: even with boot failures now loud, a mid-run death still degrades into a plausible-looking success, which is the worst outcome for an orchestrator.

Start at the await_members entry point and trace how vanished workers receive their status and how the "All members done" condition is computed. Reproduce a worker disappearance, then verify that unknown workers do not count as done and the summary distinguishes ready, failed, and vanished members.