Assistant text in the same message as an AskUserQuestion/ExitPlanMode tool call is silently dropped (not rendered, not persisted to session JSONL)

Environment: Claude Code 2.1.218, macOS 26.5.2 (arm64). Reproduced under WezTerm + herdr, but both are ruled out as causes — the text is missing from the session JSONL on disk, which sits upstream of any terminal rendering.

Summary: When the model emits a text block followed by an AskUserQuestion (or ExitPlanMode) tool call in the same assistant message, the text block is silently discarded. It is not rendered in the UI, does not appear in the Ctrl+O transcript, and is absent from the persisted session file (~/.claude/projects/<project>/<session>.jsonl). Only the question/plan dialog survives.

Impact: Any analysis or context the model writes immediately before asking a question is lost — the user sees context-free question dialogs (or a plan-approval dialog that looks like the model ignored their last message). Because the text is also missing from the session JSONL, the content is permanently lost on resume/compaction: the model itself loses its own prior output.

Repro (observed three times in one session):

1. In plan mode, prompt a task that leads the model to present findings and then call AskUserQuestion in the same turn.
2. Model emits [text block with findings] + [AskUserQuestion tool_use].
3. Observe: only the question dialog renders.
4. grep the session JSONL for phrases unique to the text block → zero hits, while the AskUserQuestion tool_use entries are present.
5. Control: a text block in the same turn ahead of a Write/Edit tool call is persisted and rendered normally — the drop is specific to messages containing AskUserQuestion/ExitPlanMode.

Expected: Text blocks preceding an AskUserQuestion/ExitPlanMode call render above the dialog and are persisted like any other assistant output.