MAINT BREAK: Move include_baseline from Scenario constructor to initi…

[adrian-gavrila]

Description

Drops include_default_baseline from Scenario.__init__ and from every subclass constructor. Adds include_baseline: bool | None = None as a keyword-only parameter on Scenario.initialize_async, under
@apply_defaults.

This treats baseline like every other common runtime parameter (objective_target, scenario_strategies, dataset_config, max_concurrency, max_retries, memory_labels), so it can be overridden per
run from a config file, the CLI, or the frontend without rebuilding the scenario object.

Class-level controls

Two ClassVar[bool] flags on the Scenario base, mirroring the existing TARGET_REQUIREMENTS pattern:

class Scenario:
    SUPPORTS_DEFAULT_BASELINE: ClassVar[bool] = True   # capability: can it have one?
    DEFAULT_INCLUDE_BASELINE: ClassVar[bool] = True    # default when caller does not specify

• AdversarialBenchmark and Psychosocial set SUPPORTS_DEFAULT_BASELINE = False. Explicit include_baseline=True on these raises ValueError.
• Jailbreak sets DEFAULT_INCLUDE_BASELINE = False. It supports a baseline but does not include one by default (templates already dominate the run). Callers can opt in per run.

Resolution checks capability first, then default. A forbidden scenario can never silently inherit a True default.

Migration

# Before
scenario = RedTeamAgent(include_baseline=False)
await scenario.initialize_async(objective_target=target, ...)

# After
scenario = RedTeamAgent()
await scenario.initialize_async(objective_target=target, include_baseline=False, ...)

Testing

• pytest tests/unit/scenario: 555 passed.
• pytest tests/unit: 7460 passed.
• ruff check, ruff format, pre-commit (incl. ty type check): clean.
• Notebook docs regenerated end-to-end against a live target.
• New tests: initialize_async(include_baseline=True) on AdversarialBenchmark raises ValueError; include_baseline=False on the same is accepted.