feat(release): recursion guard for the health gate; run-task drift warning; M20 packages/* decision recorded (#39)

- release.sh health gate runs with MOSAIC_ENSURE_SKIP=1: the gated task run
  cannot re-enter release self-determination
- run-task.sh warns on release drift instead of silently using a stale image
- ROADMAP: M20 decision recorded (packages/* monorepo at usurpation,
  continuity-first); restructure sequenced as M20 phase 1

Closes #39

skills/ms-agent-watch/SKILL.md

@@ -0,0 +1,69 @@
+---
+name: ms-agent-watch
+description: Use for all instances where a watch, wait, or agentic status check is needed. This skill avoid the need for wait cycles and other methods used to wait for an action outcome.
+disable-model-invocation: false
+---
+
+# ms-agent-watch — self-armed condition watches
+
+One CLI call arms an isolated `systemd --user` transient timer per watch. Each
+tick is a fresh, cgroup-isolated process; no loop, no script, no orphan. You
+never `sleep`, never write a watch script, never background anything.
+
+## Tool
+
+```bash
+tools/agent-watch/agent-watch.sh start \
+  --name <lowercase-id> --session <your-tmux-session> \
+  [--socket <tmux-socket>]            # REQUIRED for mosaic-fleet seats using tmux
+  --when '<shell command>'            # exit 0 = met; quote it
+  --message "text delivered to you" \
+  [--class actionable|reaction|human|digest|terminal-log] \
+  [--interval 30] [--timeout 3600] [--repeat] [--quiet-timeout]
+
+agent-watch.sh list                   # your host's watches
+agent-watch.sh log <name>             # what it did and why
+agent-watch.sh stop <name>            # retire + clean state
+```
+
+## Rules
+
+1. **Name it for the thing watched** (`ci-pr1350-green`, `peer-orch-01-reply`),
+   not for yourself — names appear in `[watch:<name>]` message prefixes and
+   in `systemctl list-timers 'agent-watch-*'`.
+2. **Interval floor is 10s.** A watcher is a fallback cadence. If you find
+   yourself wanting 1s polls, the real fix is an event, not a faster watch.
+3. **Timeout is real** (default 1h): on expiry you get one terminal-log note
+   (unless `--quiet-timeout`) and the watch retires. A watch is never forever;
+   re-arm deliberately if the condition is still pending.
+4. **Conditions are cron-style**: clean-ish environment, `cwd=$HOME`. Use
+   absolute paths. Do NOT rely on ambient credentials — resolve tokens through
+   the git credential helper or absolute service-credential paths.
+5. **rc=2 delivery is DELIVERED** (text reached your pane as a draft — FLEET-COMMS
+   E7); the watcher never retries it. Real failures retry twice then retire loudly.
+6. **A broken condition (exit ≠ 0/1) retires the watch** with a terminal-log
+   note. Check `log <name>` before re-arming — re-arming a broken condition
+   without fixing it just burns another timeout.
+7. **`--repeat` re-arms after every delivery.** Default is one-shot on purpose:
+   each watch should correspond to one pending fact.
+8. **Retire your watches** when the mission closes (`stop`). `list` shows
+   stale-state entries whose timer is gone; stop removes those too.
+9. Fleet seats MUST pass `--socket mosaic-fleet` — the default is the default
+   socket and the delivery will not reach you.
+
+## When NOT to watch
+
+- Waiting on another SEAT: send them a message (agent-send / comms) instead —
+  a watch polling their output is a loop wearing a hat.
+- Waiting on yourself: just do the next thing.
+- Something that already has a wake path (fleet watcher injects comms
+  automatically): those arrive without any watch.
+
+## Failure model (what you'll see)
+
+| Symptom | Meaning |
+|---|---|
+| `started watcher ...` then nothing, timer inactive | condition broke (exit ≠ 0/1) or timeout hit — `log <name>` says which |
+| `[watch:x] ... timeout after Ns` | retired; re-arm only if still relevant |
+| delivered (rc=2) in log | delivered as draft into your pane — go read the pane |
+| delivery failed rc=1 after 3 attempts | your session was gone; restart it, then re-arm |