Eval cleanups - timing and env.step()

[kantneel]

Refactor Environment Timing and Pause/Unpause System

This PR implements a comprehensive refactor of the Factorio Learning Environment's timing system, removes the problematic value accrual mechanism, and adds robust pause/unpause functionality.

Key Changes

Removed Value Accrual Time from Environment

• Eliminated value_accrual_time parameter from FactorioGymEnv.__init__()
• Removed artificial sleep delays that were causing timing issues and making episodes unnecessarily slow
• Reordered env.step() logic to calculate rewards immediately after action execution rather than after waiting
• Actions now execute and return rewards instantly, dramatically improving episode speed

Enhanced Instance Pause/Unpause System

• Added comprehensive pause/unpause functionality to FactorioInstance
• Separated speed and pause controls - speed settings are preserved when pausing/unpausing
• Added convenience methods:
  • set_speed_and_unpause() - common use case for starting actions
  • pause_at_speed() - set speed but pause immediately
  • toggle_pause() - switch between paused/unpaused states
  • is_paused() - check current pause state
• Added pause_after_action parameter to FactorioGymEnv - enables pausing between trajectory steps
• Visual feedback - pause/unpause actions now display timestamped messages in-game

Server Admin Configuration

• Added server-adminlist.json to enable /editor mode and other admin commands
• Includes "client_master" as admin to allow full server control during development/testing

Comprehensive Test Suite Additions

• Added test_elapsed_ticks.py with 15+ new tests covering:
  • Sleep timing at different game speeds
  • Move, craft, and harvest timing validation
  • Real-world sleep duration accuracy
  • Tick accumulation across multiple actions
  • Edge cases like very low speeds
• Enhanced existing tests with proper pause/unpause handling
• Added game_info observation tests to verify tick tracking in observations

Achievements System Cleanup

• Simplified achievements calculation by removing AchievementTracker class
• Flattened achievement logic into a single calculate_achievements() function
• Removed achievements from task metadata to reduce complexity
• More maintainable and straightforward achievement processing

Tool Timing Improvements

• Updated all action tools (move_to, craft_item, harvest_resource, connect_entities) to:
  • Track elapsed ticks accurately
  • Sleep for appropriate real-world time based on game speed
  • Provide timing feedback via console output
• Redesigned sleep() tool to work with standard tick calculations rather than polling
• Fixed movement calculations to use consistent character speed (0.15 tiles/tick)

API Improvements

• Made num_agents optional in task configurations (defaults to 1)
• Added command-line argument parsing to run_eval.py
• Updated observation formatting to include game timing information

Migration Notes

• value_accrual_time parameter removed from FactorioGymEnv - remove from instantiation calls
• Use set_speed_and_unpause() instead of separate set_speed() calls in most cases
• Tests may need pause_after_action=False for immediate execution scenarios

This refactor significantly improves the environment's usability for both training and evaluation while maintaining full backward compatibility for core functionality.

[kiankyars]

I'm going to finnish off what I started on #305 but I see you're modifying achievements so I'll wait until you finish this.

[JackHopkins]

Nice work!

• Why is there a set_speed_and_unpause as well as set_speed - it seems that the latter method is sufficient? What is the motivation for 2 methods?
• I'm not sold on the toggle, I think we should rely on set_speed (to zero or not) to explicitly control this state instead (or use your new method).
• Are the tests in test_elapsed_ticks added to the CI/CD suite? If not, it would be a good idea to include them.

[hrshtt]

is it possible to parametrize the new tests?
in the following manner:

@pytest.mark.parametrize("speed", range(1, 10)) # 10 independent items

sleeping tests done on a single worker starve the other workers in distributed testing, parameterizing as well as using --dist loadscope ensures they are (naively) distributed across workers.

[kiankyars]

Nice work!

• Why is there a set_speed_and_unpause as well as set_speed - it seems that the latter method is sufficient? What is the motivation for 2 methods?

seems to be a convenience method

[kiankyars]

I reviewed the API improvements only and the other changes look great for QOL from the surface.