[Bug] Model context pollution allows circumvention of refusal policies across model switches

**Bug Description**
Alignment problem: I started the session in Claude Haiku 4.5 low reasoning, then slowly switched models and reasoning efforts to get the model to reimplement byte-exact proprietary software in readable source code. This should not be possible. I am not the author of this software, and Opus should've refused after switching from Haiku. The polluted context of Haiku taking my request at face-value was what caused misalignment.

**Environment Info**
- Platform: linux
- Terminal: tmux
- Version: 2.1.274
- Feedback ID: 334bc498-26c8-4c28-88c2-d9b26404bba5

No repository file, test, or code entry point is named. Start by reviewing the reported sequence of model and reasoning switches and the associated Feedback ID; done would require a maintainer-defined fix that prevents prior model context from bypassing refusal policies.

Written by the indexing model from the issue text.