Fix(eval): evaluate in :task_cleanroom images

[klieret]

Submissions are now evaluated in the artifact-free cleanroom image by default instead of the full :task build environment, so a submission can't rely on build artifacts leaked into :task. --image-tag stays as an explicit override (pass --image-tag task for the full build env).

This avoids drifts between the inference image and evaluation image.

Internal-reference: fdbd6657

Closes #9

[Copilot]

Pull request overview

This PR changes ProgramBench evaluation to run submissions in the artifact-free :task_cleanroom Docker image by default, aligning evaluation with inference and preventing reliance on leaked build artifacts. It keeps --image-tag as an explicit override to use the full :task environment when desired.

Changes:

• Switch default image_tag from task to task_cleanroom in the core evaluator (eval.py) and batch evaluator (eval_batch.py).
• Update the CLI programbench eval default --image-tag to task_cleanroom.
• Expand CLI help text to explain why task_cleanroom is the default and how to override back to task.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
src/programbench/eval/eval.py	Updates the evaluator’s default Docker image tag to task_cleanroom.
src/programbench/eval/eval_batch.py	Updates batch evaluation defaults to use task_cleanroom unless overridden.
src/programbench/cli/main.py	Updates CLI default and help text for --image-tag to default to task_cleanroom.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.