Summary

• The Sheet open-focus retry had grown into a polling loop (2s/50ms when this started, 10s/50ms on main now) because a closing sheet's focus restore had no way to know another sheet had opened — the new sheet had to fight the old sheet's restore back. This fixes the race at its source and makes the remaining settle work event-driven.
• Guard the close-restore with the open-sheet stack. Restore proceeds when the target lives inside the top-most open sheet (a stacked sheet handing focus back down) and is skipped only when the target is out in the background behind a sheet that took over. Separate Sheet instances cannot cancel each other's timers, so the module-level stack is the only place this is knowable.
• Clear this instance's own pending restore on open — the restoreTimersRef comment already claimed this, but the code no longer did it, so a close-then-reopen in one frame pulled focus back to the opener.
• Replace the polling interval with a panel MutationObserver, a focusin listener, a stack subscription, and a short backoff chain. Idle when nothing changes; still upgrades from the close-button fallback to a late-mounted [data-sheet-autofocus] child the moment it mounts.
• End the settle window on the first pointerdown/keydown so a deliberate interaction is never overridden, and bound reclaims that do not stick so a surface which re-focuses itself from its own focus handler cannot trade focus() calls with the controller until the stack blows. A reclaim that holds resets the count, so defending the panel repeatedly across the window is unaffected.
• Only reclaim focus from nothing (document.body) or from background the sheet itself deactivated. A popover portaled out of the sheet (mode-action-popup, source-preview-popover, composer host) keeps focus.
• Make the page background inert while the top-most sheet is open, so the browser refuses focus escape instead of the sheet correcting it. Skips live regions, the Next dev overlay/route announcer, data-sheet-inert-skip opt-outs, and any inert another owner set; only elements present at open are marked, so a surface portaled after open stays interactive.
• Extract the stack, scroll lock, inerting, and focus controller into src/components/ui/sheet-focus.ts so the rules are unit-testable directly rather than only through a rendered Sheet.

Follow-up to the focus retries added in #1141. The merge from main keeps the case #1213 widened the poll for (a data-sheet-autofocus child arriving behind a slow dynamic import) by matching that settle window — listening is idle, so the longer window costs nothing.

Verification

• npm run verify:pr-local — runtime, format, lint, typecheck, full unit suite, production build + client-bundle secret scan, and offline RAG fixture validation, all green on the final merged tree
• npm run verify:ui (test:e2e:pr, Chromium) — 273 passed on the merged tree, including every toBeFocused assertion in the smoke spec
• Flake soak: tests/ui-smoke.spec.ts --repeat-each=3 — 273 passed, no flake in the focus-heavy paths
• Browser soak against the running app: 120 open/close cycles of the phone mode-menu sheet — zero inert-marker leaks, zero body.overflow drift, zero page errors, focus restored to the opener every cycle (median 0 ms, max 10 ms)
• tests/sheet-focus-stress.dom.test.tsx (new, 13 cases): rapid open/close cycling, three-deep stacking, middle-sheet close, StrictMode double-invoked effects, sheets with no focusable target, document listener balance, inline (non-portal) inerting, detached roots, nested sheets, focus wars, and churn
• tests/sheet-focus.dom.test.tsx (new): reclaim rules, background inerting including opt-out and foreign-inert cases, the sheet-swap restore race, close-then-reopen, late-mounted autofocus upgrade, user-interaction cancellation
• tests/ui-accessibility.spec.ts (new case): proves the containment contract in a real browser — jsdom ignores inert, so this is the only place it can be proven
• tests/ui-overlay-css-contract.test.ts pins "no setInterval polling" and the stack-guarded restore

What stress testing found

Two defects, both fixed here before this was ready for review:

1. The first version of the stack guard skipped every restore while any sheet was open, so closing a stacked sheet dropped focus to the body instead of returning it to the sheet underneath. Caught by the three-deep stacking case.
2. A background surface that re-focuses itself from its own focus handler could trade focus() calls with the controller synchronously. Now bounded, and the controller yields.

Risk and rollout

• Risk: medium — background inerting changes what is interactive behind an open modal in real browsers (jsdom does not enforce inert). Focus behaviour itself is narrowed, not widened: the sheet steals focus in strictly fewer situations than before.
• Rollback: revert this branch's commits; sheet.tsx returns to the in-file stack and the 50ms polling retry.
• Provider or production effects: None

Notes

• data-sheet-portal is a new opt-in marker for portaled surfaces that are logically inside a sheet; data-sheet-inert-skip is the background opt-out.
• The dev server started for browser QA was stopped afterwards; worktree is clean and no repo-owned process was left running.