MAINT: Migrating context compliance to technique

[rlundeen2]

Summary

Migrates ContextComplianceAttack from a bespoke single-turn executor into a core simulated-conversation technique, mirroring the earlier role_play migration. The prompt-scaffolding work that used to live in the executor (fabricating a benign Q&A + offer, then sending a hardcoded "yes.") now belongs to the attack technique layer, per doc/code/framework.md.

What changed

• New technique wiring: context_compliance is now registered in pyrit/setup/initializers/techniques/core.py via AttackTechniqueFactory.with_simulated_conversation(...) with num_turns=1, a fixed final_user_message="yes.", and dedicated adversarial + simulated-target system prompts (paths inlined like every other technique).
• Factory extension: with_simulated_conversation now accepts simulated_target_system_prompt_path and final_user_message, appending the fixed final user turn as a static SeedPrompt.
• New dataset YAMLs:
  • datasets/executors/red_teaming/context_compliance/context_compliance.yaml (adversarial → benign question)
  • datasets/executors/simulated_target/context_compliance_target.yaml (simulated target → benign answer + offer)
• Removed the ContextComplianceAttack executor, its old dataset, and unit test; rewired the scam scenario's inline context_compliance branch to the factory; updated affected tests and docs; synced doc/bibliography.md.

Breaking change

Removes ContextComplianceAttack from the public API. Follow-up to the role_play technique migration.

Validation

• All affected unit tests pass; ruff format/check and ty clean; pre-commit hooks pass.
• Live end-to-end validated against gpt-4o (adversarial + target): benign question → offer → "yes." → objective delivered.