fix(cli): fresh ephemeral room agents + clean channel teardown

[rogeriochaves]

Why

Ephemeral "rooms" of agents spin up a handful of Claude sessions under short readable slugs (e.g. dax, enzo, faye) and have them coordinate in a shared channel. Because the launch derives a deterministic session id from the slug (uuidv5(slug)), reusing those names across rooms breaks in two ways:

• kanban launch defaults to resume, so a recycled slug reloads a stale or unrelated prior conversation under that id, or dies with "No conversation found".
• Even with no-resume, claude --session-id <uuidv5(slug)> collides with the slug's own prior transcript and errors "Session ID already in use", which breaks the card-to-session link so missions never deliver and the room looks dead.

Separately, channel delete removed the metadata but left the append-only <name>.jsonl, so a channel re-created with the same name replayed stale history.

What

• kanban launch --no-resume: a forceFresh LaunchOption that gates the resume check, for ephemeral agents that should always start clean.
• Under forceFresh, mint a unique random session id for the launch instead of uuidv5(slug). The stable readable tmux name is preserved, so the card-to-session link holds and the agent stays deliverable.
• channel delete now removes the <name>.jsonl log too, covering every delete path (CLI, reset tooling, room teardown). A re-created channel starts fresh.
• Test for the fresh-launch path: a resumable transcript is ignored and a unique id is minted.

Notes

Extracted from hackathon work on a voice-orchestrated room of agents, rebased onto latest main. The handle-from-tmux-session fix from that batch is already covered upstream by the channels + CLI work in #105, so it is not included here.

Tests

pnpm test in cli/: 274 pass, 0 fail (including the new forceFresh launch test and the updated deleteChannel test).