Model appends unrequested suggestions and anticipates the user's next step despite CLAUDE.md rules

## Summary

Claude Code appends content the user did not ask for: unrequested suggestions, alternative approaches, and sentences that anticipate the user's next action or intent. This happens even when the project CLAUDE.md explicitly forbids it.

## Reproduction

1. Put a rule in CLAUDE.md such as "answer only what was asked; do not add suggestions or predict the user's next step".
2. Ask a purely factual question (for example, "what are this machine's specs?").
3. The model answers the fact, then adds recommendations or next steps that were not requested.
4. Tell the model in chat not to do this. It complies for one turn, then the behavior returns.

## Expected

- Answer exactly the question asked, with no appended recommendations.
- Do not write sentences that predict the user's next action, intent, or the rest of their sentence.
- Honor a user-authored "no suggestions" rule consistently across the session, not just for the next turn.

## Actual

Helpful-sounding additions are appended regardless of the rule, and the rule decays after a few turns.

## Why it matters

Every unrequested proposal on screen is a forced context switch for a user who is doing their own design and judgment work. A "just the answer" behavior that is actually enforced, or stronger weight on user-authored no-suggestion rules, would fix it.

The report names no source file, entry point, or test. Start by reproducing the behavior with the stated CLAUDE.md rule and a factual question, then trace where session instructions are assembled and persisted. Done means the no-suggestions rule remains effective across turns and responses contain only what was requested.

Written by the indexing model from the issue text.