[Behavior Mgmt] Gemini CLI should build and lint after every file change

Tracks updating the system prompt and evals to ensure that Gemini CLI builds and lints and runs tests after most file changes.

Right now it seems to wait until the end of the scenario, often resulting in a tail of unnecessary changes or a 'run-on' whack-a-mole' of build and test fixes. This is particularly troublesome when doing [long, repetitive operations](#17107) because it makes it harder to resume in a new session.

The goal is to get the agent to work more like a human and aim to keep the code buildable at each stage.

**Scope of work:**
- Eval test that checks that the agent builds and runs tests after each file change.
- Prompt changes as needed to steer it to do so.