fix: keep native voice recording through long turns

[rogerchappel]

What changed

• remove the native iOS Agent voice session's unconditional 20-second turn finalization
• wait for three seconds of sustained silence before finalizing, reducing cutoffs during natural pauses
• emit a one-time native.recording.long-form diagnostic after 20 seconds while capture continues
• add regression coverage to keep the hard duration cutoff from returning

Root cause

The native VAD completed a recording when either silence reached 1.8 seconds or total recording duration reached 20 seconds. A continuously speaking user was therefore forcibly sent to transcription at roughly 20 seconds, after which the partial transcript was immediately submitted to chat.

User impact

Regular, non-realtime Agent voice can now capture speech beyond 20 seconds and tolerates longer natural pauses before starting the assistant response.

Verification

• pnpm test -- src/lib/native-voice-session-generator.test.ts
• pnpm typecheck
• pnpm exec eslint apps/mobile/scripts/ensure-ios-voice-session.mjs src/lib/native-voice-session-generator.test.ts
• xcodebuild -project App.xcodeproj -scheme App -sdk iphonesimulator -configuration Debug CODE_SIGNING_ALLOWED=NO build
• pre-push pnpm build
• git diff --check

Remaining risk

Very long turns are still uploaded as one WAV/STT request. Rolling transcription segments should be implemented separately rather than expanding this focused fix.