feat(lifecycle): detached handoff supervisor with OS-owned callback

[nullStack65]

Detached T3 lifecycle handoff (supervisor + callback)

An agent running inside T3 cannot upgrade T3: the lifecycle action kills the agent
before it can report back. This adds a detached handoff that moves the action to
an OS-owned supervisor whose lifetime is independent of T3.

Draft. Non-overlapping with the Windows SCM lifecycle work on #10 (no shared
files: prepared #10 changes and this branch touch disjoint paths). Related
lifecycle hub: #10; environment coordinator: nullStack65/closura-agent-config#237.

What is here

• Envelope (lib/envelope.mjs): owner-private, credentials refused at write
  time, describes originating thread/project/machine, command, wait target,
  relaunch, readiness, identity and callback.
• macOS: transient per-handoff LaunchAgent (RunAtLoad, KeepAlive=false,
  Background). Its direct parent is launchd (pid 1), so T3/Electron job-object
  teardown cannot reach it. It removes its own plist and bootctl bootouts itself.
• Windows: transient current-user Scheduled Task (InteractiveToken,
  LeastPrivilege, one-shot). Implemented; not executed on a Windows host.
• State machine (lib/run.mjs): PREPARED → DETACHED → WAITING_FOR_EXIT → APPLYING → RELAUNCHING → WAITING_FOR_T3 → CALLBACK_PENDING → COMPLETE/FAILED,
  single-run claim, terminal-result idempotency that refuses to re-run the
  destructive command, bounded timeouts, and a hard rule that a still-alive
  waited-for pid fails the handoff rather than being force-killed.
• Independence proof (lib/process.mjs): parent chain walk asserting
  directParent === 1 and no initiator ancestor — required before the initiator
  may quit T3.
• Callback contract: terminal result envelope + gh durable receipt
  (or generic HTTP POST). No state.sqlite writes, no fabricated orchestration
  events, no copied auth tokens.
• Docs: docs/internals/lifecycle-handoff.md (Mac, Windows, callback
  contract, security model, recovery, agent invocation). Operator README under
  scripts/lifecycle-handoff/.

Evidence on this head

• Unit: node --test test/envelope.test.mjs test/run.test.mjs → 11 passed / 0 failed.
• Real macOS fixture: node --test test/fixture.integration.test.mjs → 1 passed.
  Real LaunchAgent; fake parent prepares a handoff and exits; the detached
  supervisor survives (independence proof: pid directParent=1, ancestry [1],
  independent=true), runs the harmless command, relaunches a fake service,
  delivers the callback exactly once (state=COMPLETE), unregisters its own job,
  and refuses to re-run on idempotent replay.
• Full suite: 12 passed / 0 failed.

Callback status

T3 exposes no supported CLI/REST callback today; the only mutation transport
is authenticated /ws orchestration.dispatchCommand. This PR therefore ships the
result-envelope + GitHub-receipt fallback behind a stable callback.kind, so a
future supported endpoint slots in without touching lifecycle logic.

NOT DONE (explicit)

• Windows execution on a Windows host (source + docs only).
• Real-T3 restart smoke is not run from an active initiating session (it would
  terminate that session); the mechanism is proven by the fixture. Run it from an
  idle T3 with a bounded, non-mutating command.
• No wiring into bin.ts; invocation is documented in the README.
• No packaging/CI/merge. Not independently reviewed.

[github-actions]

Thread transfer impact

⚠️ The latest CI run did not produce a thread transfer result for 59bfaab.

This comment will update automatically after the next completed run.

[nullStack65]

Closing as an obsolete proposal under the current approved delivery direction, not as superseded or implemented. Exact head 59bfaabcb52b5fc4e77228efdd2aaad97ae45ab5 contains unique detached OS-supervisor/state-machine/callback code that is NOT in main. It remains unwired to the supported CLI, lacks a supported T3 callback path, and records unexecuted Windows and real-T3 acceptance. The current lane explicitly excludes a new manager/orchestrator framework and callbacks; no independently useful small repair can be separated from this architecture. Preserve this PR and branch as unmerged design history; no equivalent self-upgrade capability or acceptance is claimed.

Audited against current main cb9636bfede88938f4b2358ba439884c52760b4e (tree 4b6973cd81a58ec85ec93938c072f31243377889) with Luna investigation and distinct independent read-only review, integrated by Sol. Current-main focused BootService/service CLI tests: 89 passed. Normal postmerge CI 37635278267 passed Check, Test, all three server shards, Rust and Release Smoke (conditional mobile native analysis skipped). Closure preserves the PR, discussion, source branch and commits; no stale merge or branch deletion. Source evidence does not establish installed/native acceptance; shared release remains owned by #20.