Relax realtime voice turn detection for slower speech

[rogerchappel]

Summary

• Pass OpenClaw realtime VAD tuning through the CrewCMD session endpoint
• Default realtime sessions to wait 2000ms of silence before ending a user turn
• Increase prefix padding to 500ms so slower starts and resumes are less likely to be clipped
• Add route coverage for the default tuning and explicit VAD overrides

Why

OpenClaw defaults realtime voice endpointing to roughly 500ms of silence. That is too aggressive for slower speakers and causes the agent to respond mid-thought.

Verification

• pnpm test -- src/app/api/runtimes/[id]/talk/realtime/session/route.test.ts src/lib/gateway-client-realtime.test.ts
• pnpm typecheck
• git diff --check
• push hook: pnpm typecheck
• push hook: pnpm build (passes with existing Turbopack NFT-list warning)

Risk

Low. Scoped to realtime session creation parameters and type coverage. The tradeoff is slightly slower agent turn-taking after the user finishes speaking.