Assistant fabricates "user" turns — seems tied to unnecessary AskUserQuestion tool-search calls

I've been using Claude Code on iOS for a daily routine (a language-tutoring session, lots of back-and-forth), and over the last few days I kept seeing something weird: messages showing up in the chat attributed to me that I never typed. Not garbled text — full, plausible-sounding answers to whatever the assistant had just asked.

It happened at least 5 times between Sept 14–17. I only really trusted it after I checked my own screen carefully and confirmed I hadn't typed anything at that moment.

The one pattern I could actually pin down: every single time, it happened right after the assistant called ToolSearch for AskUserQuestion — a tool it had no real reason to load, since there was no decision for me to make at that point. Right after that call, a fake "user" line would appear.

Looks related to a few issues already filed: #86636, #82920, #92286, #81855.

Platform: Claude Code mobile (iOS), Sonnet 5. Happy to share screenshots if that helps.