hotfix: disable RTC inference engine for ACT checkpoints

[Mokuroh54]

Summary

RTC (Real-Time Chunking) regenerates action chunks through lerobot's flow-matching processor — a path ACT checkpoints don't have — so an ACT run started under RTC breaks at rollout.

Both engine selectors (studio DeployPanel, landing InferenceModal) now force Sync whenever the selected checkpoint is ACT:

• The RTC SelectItem is disabled, with a new hint explaining why (en + zh-CN keys added).
• The submitted inference_engine is derived (isAct ? "sync" : selection), not a state reset, so a stale RTC selection from a previous checkpoint can never reach the backend — including across the policy-config fetch race.
• Mirrors the existing ACT-only gating of temporal ensembling; no backend or schema change (the request carries no policy_type, and the backend default is already sync).

Why into main

Hotfix for released code — RTC + ACT is selectable on main today. sync_staging will carry it back into staging automatically.

Testing

• npx tsc --noEmit -p tsconfig.app.json / -p tsconfig.node.json — pass
• npm test — i18n catalog parity + dynamic-key tests pass; the one failing test (LanguageContext.test.tsx › switches back to English) fails identically on pristine origin/main (pre-existing)
• npm run lint — 4 errors, all pre-existing in untouched files (baseline-identical)
• npm run build — pass (dist not committed; CI rebuilds it)

🤖 Generated with Claude Code

[Mokuroh54]

Superseded by #113, which refuses RTC server-side for every policy that cannot run it (not only ACT) and disables the option in both dialogs from the checkpoint's supports_rtc field. Stacked on #108 → #107.