Add adversarial/edge-case trap tasks

[ScuttleBot]

Goal

Add tasks where the obvious approach is wrong, testing genuine reasoning over pattern matching.

Task Types

Red Herring Tasks

• Provide irrelevant but distracting information
• Include "obvious" solution that fails on edge cases
• Context that suggests wrong approach

Edge Case Gauntlets

• Off-by-one scenarios in dates/times/counting
• Boundary conditions (empty lists, single items, max values)
• Unicode/encoding edge cases
• Timezone handling across DST boundaries

Inherited Mess Tasks (Recovery-Bench style)

• Workspace containing prior failed attempts that need cleanup
• Broken state that agent must diagnose before fixing
• Conflicting partial solutions left behind

Specific Task Ideas

1. The Misleading Log: Error message points to wrong root cause
2. Off-by-One Gauntlet: 5 date/time operations where edges matter
3. The Cleanup Job: Previous agent left half-done work with bugs
4. The Obvious Trap: Task where copy-paste solution from docs fails

Grading

• Binary: did they avoid the trap?
• Bonus: did they explain why the obvious approach fails?

Success Criteria

• Traps should catch >30% of models
• Tasks should differentiate reasoning vs. pattern matching

References

• ARC-AGI design philosophy
• Recovery-Bench: evaluating error recovery