feat: recover detached agent runs

[roackb2]

What changed

• add a Check status recovery path when retained live replay is unavailable
• persist the active run's base source revision alongside its sequence cursor
• reuse one MotionDoc reconciliation policy for live terminals and durable status recovery
• auto-apply a completed detached result only when the current deck still matches the run base; otherwise retain the result for the existing Apply agent result / Keep mine choice
• persist nonterminal cursors only after editor event handling succeeds, and preserve the cursor while rechecking the same active run
• replace the destructive native confirm with a focus-managed Radix AlertDialog
• extend Playwright coverage for replay expiry, automatic recovery, manual-edit protection, and both reset-dialog choices

Why

A replay-expired run was reported honestly, but the editor remained wedged until a page reload and could not safely reconcile the eventual durable result. This makes that state recoverable without duplicating Heddle's run service or sacrificing SlideX's source-revision/undo boundary.

The base source revision is product state, so it stays in SlideX's project/run binding. Heddle continues to own event cursor correctness, replay, and terminal semantics.

Dependency and rollout

• stacked on and intended to merge after #7, which itself follows feat: bind agent conversations to project lifecycle #6
• the agent remains behind the default-off editor flag
• Heddle #238 is still required to avoid replay-from-zero when the stored cursor remains available

Validation

• clean npm ci
• npm run test:agent (7 passed)
• npx tsc --noEmit
• npm run lint (0 errors; 11 existing warnings)
• npm run test:agent:e2e (3 browser scenarios passed)
• feature-enabled production build
• independent browser check: meaningful page and agent panel, no error overlay, no captured console errors

npm audit --omit=dev remains unchanged at the existing Next/Geist PostCSS advisory (3 moderate); pinned Radix AlertDialog adds no new production finding.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5ce6b7098e

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Guard base revision by the restored run id

When a stored binding is stale and the session endpoint reports a different active run, this restores the old run's baseSourceRevision onto state.activeRun.runId. In that scenario (for example a duplicated tab sharing the session starts another run), if live replay later expires, checkRunStatus will compare the current deck to the stale hash and can auto-apply the new run's latestMotionDoc instead of forcing review. Please only carry binding.baseSourceRevision forward when binding.runId === state.activeRun.runId; otherwise clear it like the polling path does for run-id changes.

Useful? React with 👍 / 👎.