Mid-turn user messages arrive only with the next tool result, after a whole parallel batch has run

**What happened**
The user's correction was delivered "alongside the next tool result" after a twelve-call parallel batch had fully completed, so the model could not act on it until every call in the batch had already been paid for.

**Expected**
The message surfaced before further calls in the batch dispatch, or a visible interrupt that halts the batch.

**What the user said**
"The thing to say in the main conversation is one line: stop the shell reads, go back to the graph"

**Repro**
Send a message while a multi-call parallel batch is running.

**Environment:** Claude Code CLI 2.1.270 (Claude Code), macOS, auto mode, model claude-fable-5-1, session of 2026-09-14.