feat(i18n-ko): native surfaces and E2E fixtures

[scs0209]

Summary

Wire ko through core locale resolution, desktop native surfaces, and E2E fixtures. This follows the same pattern as #3853 (zh-TW): Korean copy for startup/quit/boot dialogs, diagnostics, notifications, permission overlay, and related main-process fallbacks; E2E fixture flags accept ko / ko-kr without the toLowerCase regression fixed for Chinese locales in #3853 review.

Renderer UiCatalog types require a ko key once ko is in UI_LOCALES. Catalogs outside the #3977–#3979 renderer slices use English stubs for now; session-collaboration-copy includes a real KO block.

Review note: This branch stacks on feat/zh-tw-locale (#3853). GitHub cannot target AcerYue0:feat/zh-tw-locale as the PR base on apache/maka, so the PR currently targets main and includes #3853 commits in the diff. Do not merge until #3853 lands. To review only the Korean work, use:
AcerYue0:feat/zh-tw-locale...scs0209:feat/i18n-ko-native-e2e-3980

Fixes #3980

Verification

• npm run build:test — pass
• node --test packages/core/dist/__tests__/ui-locale.test.js — pass
• node --test apps/desktop/dist/main/__tests__/e2e-fixture-locale.test.js — pass
• Playwright E2E (ko-locale.spec.ts) — not run locally (dev server not started)
• npm run lint / npm run format:check — not run for the full workspace

AI use

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool(s) and scope: Cursor — implementation, translation stubs, tests, and PR text. The update(i18n-ko): session-collaboration… commit includes a Generated-by: Cursor trailer.

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck and the affected suites pass locally

Does this PR entail a change in behavior?

• Yes — described under Summary above
• No

[ggbdpq]

Plan for the P1, folded with the two named downstream gaps so the slice lands as one coherent contract change:

1. Shared predicate in Core next to MessageContent: meaningful content = non-empty text OR quotes.length > 0 OR attachments.length > 0.
2. Durable admission: normalizeRootTurnMessageContent switches from the text-length rule to the shared predicate, keeping its own attachment-count, trust, and size checks.
3. Compaction estimate: estimateRuntimeEventChars counts the structured envelope (quote texts + attachment identity) so a structured-only user event is no longer estimated at zero and dropped by the history-compact gate.
4. Recap: the recap projection switches to the same visibility/content decision instead of content.text.trim().
5. Verification: storage-level admission unit tests (quote-only accepted; truly empty still rejected with the same frame error), estimate/recap unit tests, and a composition-level test driving a real SQLite-backed turn.start with a quote-only message through idle and queued paths to a durable Turn. Decoder-only examples are excluded per your note; a full Host-subprocess run remains the acceptance bar for closing desktop: quote-only and attachment-only messages require inline text #4804.

This slice deliberately does not touch the empty-bubble half of #4804 (out of scope as agreed). Will start after the current review round settles so the layers land and get re-reviewed together.