fix(cleanup): preserve baseline seed experiments on standard cleanup

[stephen-chu]

Problem

scripts/cleanup.py → delete_ci_experiments() deletes every experiment linked to the dataset:

experiments = list(ls_client.list_projects(reference_dataset_id=datasets[0].id))
for exp in experiments:
    ls_client.delete_project(project_name=exp.name)   # deletes ALL, incl. baselines

That includes the baseline-haiku-… / baseline-sonnet-… experiments setup.py seeds as the demo's Haiku-vs-Sonnet "before" reference. So a standard cleanup empties the dataset's experiment list — even though the README promises "After cleanup, the demo is ready to run again — no need to re-run setup.py." The presenter has to reseed baselines manually after every cleanup.

Fix

Skip experiments whose name starts with baseline- (the prefix setup.py uses). CI/Engine experiments (engine-…) are still swept as before. Docstring, the [2/3] step message, and the README Cleanup section updated to say baselines are preserved.

Testing

Verified against a live dataset with two baseline seeds present:

[2/3] Removing CI/Engine experiments (keeping baseline seeds)...
  Deleted 0 experiment(s), kept 2 baseline seed(s).

Both baseline-haiku-… and baseline-sonnet-… survived; a non-baseline- experiment would still be deleted (logic unchanged for those).

Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com