Add Korean and English display and request-language support

[hyperdrivebeep]

Add independent Korean/English display selection and response-language selection for each submitted question. Accepted requests retain their language across queues, retries and recovery; subsequent questions in the same conversation may choose another language. Preserve original quotations and saved answers, and distinguish UNKNOWN state from its localized label.

Keep the composer compact: move persistent default-saving and language/capacity explanations into Settings, retain small per-request controls, and display the accepted language only when it differs from the next selection. The observed Chrome composer height decreased from 214px to 110px. Also localize the program-owned redacted span placeholder without translating authored source titles.

This release excludes the 20 command-input refactor files bundled in the local language checkpoint and all later refactor work. Existing README demo images and Toti are preserved.

Validation:

• Locked installs, Python lint, architecture and size budget passed. Final web suite: 754 passed, 1 optional skipped; lint/type/build passed. Public CI passed on b9e06c5.
• Full applicable Python run covered all 3,153 active cases (139 disabled-Hook cases excluded by existing policy): 3,149 passed, 2 failed, 1 skipped, 1 existing xfailed. No missing, extra or duplicate cases. The two failed cases passed individually with unchanged tests/time limits. One original failure was the 25-second process-start timeout; the other was browser acquisition returning zero artifacts, with its original cause not established. The initial full run is NOT relabelled all-green.
• Python sources remained unchanged from clean starting candidate 3b57366 while the requested frontend adjustment was made; all final web checks and 78 affected source/static/web-mount Python checks were refreshed on b9e06c5.
• Real Chrome workflow: Korean question with English response, followed by an English question with Korean response in the same conversation; saved answer text and quotes preserved across locale changes/reload. B remains unmeasured; the Korean follow-up verifies A=0.80 against target 0.95 (difference 0.15). Research judgment remains PARTIAL_HOLD, not a claim that a cause is established.
• Live calls: 16 sends, 15 completed; an initial gpt-6-astra planner call ended OAUTH_INCOMPLETE_RESPONSE without automatic retry. The two complete workflows used gpt-6-sol/low. This is not all-provider or general research-quality certification. An initial observer-only error caused zero sends and was corrected without changing product code.