MAINT GCG: model progressive admission transitions (#2665)

[Vardhman Gupta (Kaap10)]

Description

Addresses item 3 of #2665.

This PR models and extracts the progressive admission and scheduling transition decisions from ProgressiveMultiPromptAttack.run() into a dedicated ProgressiveScheduleController component alongside ScheduleTransitionAction(Enum).

Summary of Changes

1. Extracted ProgressiveScheduleController & ScheduleTransitionAction in pyrit/executor/promptgen/gcg/attack/base/progressive_schedule.py:

   • Encapsulates progressive goal and worker admission sequence (goals admitted before workers).
   • Manages remaining step budget calculation and enforces exact-budget termination without dangling sentinel resets.
   • Explicitly tracks loss_is_measured as a boolean flag to protect against sentinel-overwrite regressions while preserving legitimate non-finite model losses (inf / nan).
   • Implements control-weight ratcheting (+0.01 up to 0.09) and threshold transitions (> 0.09 disables stop_inner_on_success).
   • Triggers final held-out evaluation (ScheduleTransitionAction.FINALIZE_AND_STOP) only when fully admitted and success is achieved.
   • Houses ProgressiveScheduleState dataclass in the module for clean modularity.

2. Refactored ProgressiveMultiPromptAttack.run() in pyrit/executor/promptgen/gcg/attack/base/attack_manager.py:

   • Delegates inner attack round preparation, remaining-step budgeting, and transition handling to ProgressiveScheduleController.
   • Preserves 100% backward compatibility for last_schedule_state, return signatures (tuple[str, int]), and parameter logging.

Tests and Documentation

Unit & Integration Tests:

• Added dedicated unit test suite in tests/unit/executor/promptgen/gcg/test_progressive_schedule.py (16 deterministic tests):
  • Input validation for non-positive goals and workers.
  • Initial state configurations (progressive_goals / progressive_models).
  • Goal-before-worker admission sequencing.
  • FINALIZE_AND_STOP action on success in fully admitted state.
  • Control-weight ratchet increments and upper threshold transitions (> 0.09).
  • Exact-budget exhaustion boundaries at goal, worker, and control-weight transitions.
  • Non-finite inner loss tracking and validation.
• Ran the full GCG unit and integration test suites:
  • pytest tests/unit/executor/promptgen/gcg/: 309 passed (100%)
  • pytest tests/integration/executor/promptgen/gcg/test_gcg_attack_wiring_integration.py tests/integration/executor/promptgen/gcg/test_gcg_integration.py: 16 passed (100%)

Code Quality & Static Analysis:

• ty check pyrit/: 0 errors (all checks passed)
• ruff check: 0 errors (all checks passed)
• ruff format --check: 1653 files already formatted
• pre-commit run: 18/18 hooks passed

[Vardhman Gupta (Kaap10)]

Hi Roman Lutz (@romanlutz), this PR is ready for review whenever convenient. Looking forward to your thoughts, and happy to make any adjustments needed!

[Vardhman Gupta (Kaap10)]

Thanks for catching that, Roman Lutz (@romanlutz)! Removed the duplicate class definition in attack_manager.py so the imported class is exported directly, and added regression assertions in test_run_state.py verifying both class identity and isinstance() checks against attack_manager.ProgressiveScheduleState.