feat(eval): implement deterministic code-based scorers (exact_string_match, numeric_match)

[ulises-jeremias]

Summary

I’d like to contribute a focused improvement to the evaluation module by implementing the deterministic code-based scorers that are currently placeholders:

• exact_string_match
• numeric_match

This would improve reproducibility for scenarios where deterministic checks are more appropriate than LLM-only judging.

Current State

At the moment:

• src/evaluation/scorers/code_based.py contains NotImplementedError for both functions.
• docs/evaluation.md marks these scorers as skeletons/placeholders.
• src/evaluation/tests/test_scorers.py currently verifies the not-implemented behavior.

Why this is useful

• Reduces reliance on llm_judge for objective scenario types.
• Improves repeatability and debugging of eval results.
• Enables clearer per-scenario scorer routing via scoring_method.

Proposed MVP Scope

1. Implement exact_string_match with clear normalization rules (minimal and explicit).
2. Implement numeric_match with robust numeric parsing and tolerance support.
3. Register both scorers in the scorer registry.
4. Update tests from placeholder checks to behavior checks.
5. Update docs/evaluation.md to reflect availability and usage.

Out of Scope (for this first PR)

• No multi-judge panel logic.
• No changes to llm_judge rubric behavior.
• No large evaluator architecture refactors.

Acceptance Criteria

• exact_string_match and numeric_match execute without NotImplementedError.
• Scenarios using scoring_method route correctly to deterministic scorers.
• uv run pytest src/evaluation/tests passes.
• Documentation reflects the implemented scorer status and expected input fields.

Notes

I can keep this first PR intentionally small and test-focused for quick review.
If maintainers agree with this scope, I can open a draft PR proposal right away.