fix(desktop): finish regenerate archival before host shutdown

[AR307]

Quitting just after a regenerated answer finishes can dispose host-core while
session.saveActiveRevision is still pending. The turn has already left
activeTurns, and an empty message outbox does not prove the archive finished.
On upstream 0.17.0, a held archive RPC reproduces host disposal before archival.

Track terminal event-persistence promises. During quit, stop sidecar event
production and drain those writes, checkpoints and the outbox while host-core
is alive. The terminal-write wait uses the existing two-second quit budget;
unresponsive archival logs a warning and does not prevent exiting. The existing
single-RPC archive operation and stored transcript format remain unchanged.

Validation:

• New quit regressions fail on unmodified upstream because host-core is already
  disposed while the archive is held. After the fix, success, storage error and
  unresponsive archive scenarios pass using real event persistence, disk outbox
  and shutdown wiring. Only Electron and the host RPC boundary are simulated.
• Related persistence, checkpoint and regenerate tests: 30 passed on the
  standalone PR branch.
• Desktop TypeScript check, Electron build and runtime bundle: passed on both
  the combined candidate and this standalone branch, whose sidecar uses the
  unchanged upstream runtime code.
• node scripts/e2e-regenerate-quit.mjs: isolated Windows Electron, real
  sidecar and Rust host, controlled local provider. Regenerate through the
  context menu, quit immediately on the terminal event, restart, read both
  revisions, and verify there is no disposed-host archival error. All six
  assertions passed on this standalone branch as well as the combined candidate.

No user profile or paid model is used. macOS/Linux execution and GitHub CI have
not been run for this local draft.

Related work: #704 optimizes live branch storage, while #947 and #606 address transcript/outbox recovery. This change addresses terminal archival racing with host disposal during quit.

Validation candidate: fb48a3fb6d0a6941207fec2be09add64b3fcd397; upstream main: 72b5e826cb7a9928467091ccf745aa9b225eeb04.

[muzimu217]

Verified locally on head fb48a3fb6 (rebased on today's main, which now includes the merged #1435):

• New regression suite shutdown-regenerate-persistence.test.mjs: 3/3 pass — success, storage-error and unresponsive-archive scenarios, driven through real event persistence, disk outbox and shutdown wiring.
• Related existing suites show no regression: inflight-checkpoint 8/8, persistence-outbox 10/10, regenerate-branch-archive 4/4.

The implementation reads correct to me:

• trackWrite keeps a Set of in-flight terminal writes and removes on both settle paths, so nothing leaks; flush() loops Promise.allSettled until the set drains.
• The quit ordering makes that drain guaranteed to converge: sidecar event production is disposed before the drain, so no new terminal writes can arrive mid-flush. flushEventPersistence racing the existing QUIT_TURN_SETTLE_BUDGET_MS (warn-and-exit on timeout) preserves the bounded-quit contract instead of making quit hang on a stuck archive — the right trade for a shutdown path.
• Outbox flush stays after the drain, so the archive that motivated this fix can't be lost between persistence and outbox.

One nit, non-blocking: Summary.md at the repo root is new in this PR. If it's meant as PR documentation it probably shouldn't ship in the repo — worth confirming whether upstream wants a root summary file at all.

Everything I can check locally is green; the remaining call is the maintainer's on the ADR 0060 wording.