Bug: realtime backend prompt contains leaked authoring instructions

What issue are you seeing?

After https://github.com/openai/codex/pull/17493 by @aibrahim-oai codex-rs/core/templates/realtime/backend_prompt.md currently appears to contain prompt-authoring/editing instructions that were committed into the final runtime prompt instead of being resolved into clean model-facing guidance.

Examples from the current file:

• Delete every rule that sounds corporate.
• `Add a rule: 'Never open with Great question...'
• Add this line verbatim at the end of the vibe section...
• Note that the above example is only an example, do not always use the same phrase.

These read like drafting notes or rewrite instructions rather than stable final prompt content.

Relevant file and commit:

• codex-rs/core/templates/realtime/backend_prompt.md
• 163ae7d3e6578c69afcf4ba80353bc76c86a683b

What steps can reproduce the bug?

1. Use the default realtime backend prompt path (i.e. do not override experimental_realtime_ws_backend_prompt).
2. Inspect codex-rs/core/templates/realtime/backend_prompt.md in the current tree, or let the realtime flow use it via prepare_realtime_backend_prompt(...).
3. Observe that the prompt includes authoring/editing instructions instead of only final model-facing instructions.

What is the expected behavior?

The realtime backend prompt should contain only the final intended instructions for the model.

Authoring artifacts like:

• rewrite instructions
• "add a rule" directives
• "add this line verbatim" notes
• meta-comments about examples/repetition

should be resolved before the prompt is committed and used at runtime.

Additional information

This is prompt hygiene rather than a feature request, and should have been caught during PR review.

The prompt is used by the experimental realtime backend flow here:

• codex-rs/core/src/realtime_prompt.rs
• codex-rs/core/src/realtime_conversation.rs

So this is model-facing runtime content, not just an internal drafting document.