[Bug]: Streamed replies are rescanned from the top on every update#19992

Before submitting

• I searched existing issues and did not find a duplicate.
• I included enough detail to reproduce or investigate the problem.

Area

apps/server

Steps to reproduce

1. Ask any agent for a long reply (tens of KB) that streams inside a code fence or a long paragraph.
2. Watch server CPU while it streams.

Expected behavior

Per-update server work stays roughly constant as the reply grows.

Actual behavior

makeAssistantStreamingFilter (apps/server/src/orchestration-v2/assistantStreaming.ts) rescans the whole accumulated reply with splitBufferedAssistantText on every update past its 400 ms window, because no new markdown boundary appears. Work grows quadratically with reply length, on the event loop. A 50 KB fenced reply in 10-character updates costs ~282 ms of CPU; a ~50 KB newline-free paragraph ~129 ms.

Impact

Minor bug or occasional failure

Version or commit

main at 26fffe5ae

Environment

Any provider, any client. Server-side.

Logs or stack traces

None.

Workaround

None.

Fix: #19990

Implemented with Claude Opus 5.5 in T3 Code. Reviewed with GPT-6.1-Sol.