refactor(eval): unify evaluation strategy solvers using generic InferenceFormat solver

[gspencergoog]

Summary

Integrates the unified InferenceStrategy and Parser abstractions into the A2UI evaluation framework, replacing redundant solver scripts with a single generic format solver.

Changes

• Unification of Evaluation Solvers:
  • Deleted duplicate elemental and express solver files (eval/a2ui_eval/strategies/express.py, eval/a2ui_eval/strategies/elemental.py, eval/a2ui_eval/strategies/direct.py).
  • Implemented generic.py containing a unified format solver (format_system_prompt and compile_format_payload) that polymorphically initializes and executes the target format strategy.
• Payload Extraction and Parsing:
  • Implemented format-aware surface extraction using regex patterns in the generic evaluation solver.
  • Enabled the programmatic experiments API configuration in the evaluation harness to validate A2UI v1.0 specifications.
• Scorers & Infrastructure:
  • Updated scorers to use the new a2ui.strategies.schema module.
  • Refactored dataset.py and tasks.py to support the parameterized solvers.
• Testing:
  • Expanded test_strategies.py to verify prompt generation and response parsing across all strategies.

Impact & Risks

• None. This only affects the evaluation pipeline and tests.

Testing

Verify the changes by running both SDK and evaluation unit tests:

uv run pytest agent_sdks/python/a2ui_agent/tests/
uv run pytest eval/tests/

[jacobsimionato]

I don't quite understand why we need this dance. Either the surfaceId is dynamically set by the LLM, in which case we read it from the inference format as part of the parsing process, or else its injected from outside, in which case we supply it to the parser. Are you sure we need all this dynamic logic?