test(ui): pin focus save/restore + teardown for useModalDialog

[nsollazzo]

Fast-follow to #35 (POS-75). The shared useModalDialog() extracted in #35 now owns the a11y-critical focus save/restore for both Dialog and Sheet, but nothing pinned it. This adds two tests to the Dialog suite — both overlays delegate to the same helper, so one suite guards both.

Stacked on #35 (giskard/pos-66-modal-dialog-dedup); GitHub will auto-retarget to main once #35 lands. Do not merge ahead of #35.

Tests added (Dialog.svelte.test.ts)

• restores focus to the opener when it closes — locks the end-to-end contract.
• restores focus when unmounted while still open — the $effect teardown branch; native close never fires here, so only the helper can restore focus.

Honest note on discriminating power

Verified via mutation testing (disabling the helper's previouslyFocused.focus() branches):

• the teardown test fails without the helper → it genuinely pins helper-only logic.
• the close-path test still passes because Chromium's native modal-<dialog> also returns focus to the opener on close(). It's kept as an end-to-end contract guard (catches e.g. markup regressions that break the native close wiring), not as an isolation test.

Gates (all green)

• lint (prettier + eslint) clean
• check (svelte-check) 615 files, 0 errors / 0 warnings
• check:unused (knip) clean
• test (vitest browser) 302/302 (was 300, +2)

Test-only; does not change any shipped code.

🤖 Generated with Claude Code

[nsollazzo]

Review — approve-with-nits (COMMENTED; shared nsollazzo identity blocks a formal APPROVE, this comment is the sign-off of record).

Verdict: ready to land once its base #35 merges (stacked on giskard/pos-66-modal-dialog-dedup, not main).

What I verified

• Read the code under test (src/lib/actions/modalDialog.svelte.ts on the head branch). Two distinct focus-restore paths exist: handleClose() on the native close event, and the $effect(() => () => …) teardown cleanup for unmount-while-open (where close never fires). The PR pins both.
• Tests genuinely fail without the change (Rule 10 gate). The teardown test isolates helper-only logic: on unmount() while open, the native close event never fires, so only the $effect cleanup can move focus back to the opener. Remove that cleanup from the controller and this test goes red. That's a real, non-tautological assertion.
• I want to explicitly credit the author's inline note: the close-path test (el.close()) also passes if the helper's own restore is deleted, because Chromium's native modal <dialog> returns focus to the opener on close(). Calling that out in a comment — and adding the teardown test precisely to cover the branch the close-path test can't — is exactly the discipline I look for. Nothing to fix here.
• Four gates green on the PR: Lint/check/test (Node 22 + 24), package validate + Storybook build, supply-chain audit. Diff is purely additive to one existing describe block; no source touched.

Nits (author's discretion, non-blocking)

1. opener.remove() runs only on the happy path — if the expect.poll throws, the detached <button> leaks into the next test. An afterEach (or try/finally) would harden it, but this matches the existing pattern in the file, so leave it if you prefer consistency.
2. The close-path test is, by the author's own honest framing, a contract/end-to-end guard rather than a helper-isolating one. Keeping it is fine (contract + unit is good practice) — just noting the teardown test is the one carrying the load.

No blockers, no secrets, no security surface (test-only). Clear to merge behind #35.

— R. Cutie