docs(orchestrate): pin the target workspace before spawning; kill agents before closing panes - #6

Docs-only update to the orchestrate skill, from two failure modes hit in real multi-wave runs. No script changes — the wmux CLI already supports everything referenced.

1. Panes/agents spawn into the ACTIVE workspace, not the coordinator's

wmux layout grid anchors to the caller's WMUX_SURFACE_ID. In background or resumed coordinator sessions that env var is often stale (points at a dead surface), and the app then falls back to whatever workspace the user currently has active. spawn-agents.sh passes no --workspace either — so an entire orchestration grid plus its agents can materialize in an unrelated workspace the user is working in. Their --cwd still governs where they work, so the run succeeds, but it hijacks the user's screen and breaks pane-hygiene counting.

CLI facts the new text relies on (all already shipped): agent spawn, layout grid, list-panes, and agent list all accept --workspace <id>; list-workspaces returns { id, title, isActive, cwd } — and multiple workspaces can share a cwd, so the target must be resolved by id.

Changes:

• Phase 6b — resolve the target workspace at orchestration start and record it in state.json's workspaceId field (which has existed in the schema all along but was never populated).
• Phase 6d — pin before spawning: primary path is wmux select-workspace <id> right before spawn-agents.sh (works with the script unchanged); focus-preserving alternative is layout grid --workspace + per-agent agent spawn --pane. Also fixes the stale claim that the script creates panes via wmux split — it uses wmux layout grid (6d and 6e).
• Phase 7 — scope agent list with --workspace (the default is ALL workspaces, so the listing can include the user's own agents from other sessions) and never kill an agent the orchestration didn't spawn.

2. Teardown order: agent kill BEFORE close-pane

Closing an agent's pane first leaves the agent registry stale — agent list keeps reporting the agent as running, agent kill then fails with "Agent not found" — and the launcher process tree (shell → node → claude) survives the pane close, invisibly burning CPU and tokens.

Changes:

• Phase 7 wave transition — reap in order when closing finished agents' panes.
• Phase 9 — the ordering rule, the stale-registry failure mode, and the recovery path (kill the process tree by pid; judge liveness by the process, not the registry status).