single_sbatch routing: rework agent-driven (DSE/live-RL) handling and mixed-job guard ordering

Context

handle_dry_run_and_run (src/cloudai/cli/handlers.py) routes scenarios as:

agent_driven = [_is_dse_or_live_rl(tr) for tr in test_scenario.test_runs]
if args.single_sbatch or not any(agent_driven):  # in this mode cases are unrolled using grid search
    handle_non_dse_job(runner, args)
    return 0
if all(agent_driven):
    return handle_dse_job(runner, args)
logging.error("Mixing agent-driven (DSE / live-RL) and plain jobs is not allowed.")
return 1

The args.single_sbatch short-circuit is evaluated before the mixed-job guard. This ordering predates the live-RL work (it keyed on tr.is_dse_job originally) and is benign for DSE, because SingleSbatchRunner deliberately grid-unrolls DSE param spaces (unroll_dse / all_trs / handle_dse). But it has two rough edges:

1. Mixed agent-driven + plain under --single-sbatch bypasses the "mixing not allowed" error and silently routes everything through handle_non_dse_job.
2. SingleSbatchRunner has no live-RL path — it only branches on tr.is_dse_job. A live_rl_mode job under --single-sbatch would silently run once as a static job, because live-RL needs the in-process GymServer loop, not a batch grid.

A narrow stopgap guard was added for case (2) (error on live_rl_mode + single_sbatch) with a TODO pointing at this issue. This issue tracks the proper fix.

Proposed fix (dedicated PR)

• Reject mixed agent-driven + plain before the single_sbatch short-circuit.
• Make single_sbatch semantics explicit per agent type: DSE -> grid-unroll (current behavior), live-RL -> unsupported/error (or a live-RL-aware single-sbatch path if ever needed).
• Add regression tests covering single_sbatch=True for DSE, live-RL, plain, and mixed scenarios.

Files

• src/cloudai/cli/handlers.py (handle_dry_run_and_run, _is_dse_or_live_rl)
• src/cloudai/systems/slurm/single_sbatch_runner.py (all_trs, handle_dse)
• tests/test_handlers.py (_run_routing + routing tests)

Start in src/cloudai/cli/handlers.py at handle_dry_run_and_run and _is_dse_or_live_rl, then inspect all_trs and handle_dse in src/cloudai/systems/slurm/single_sbatch_runner.py. Use the _run_routing tests in tests/test_handlers.py as the entry point. Done means mixed jobs are rejected before single_sbatch routing, with regression coverage for DSE, live-RL, plain, and mixed scenarios.