feat: extend Foundry cloud evaluator coverage to 22 built-in evaluators (#51)

[Dongbumlee]

Summary

Extends AgentOps Foundry cloud evaluation to support 22 built-in evaluators (up from 8), covering all evaluator categories: quality, agent, safety, RAG, tool, and NLP. Adds CI/CD integration documentation with integration models, gating best practices, and evaluator reference.

Closes #51

Changes

Code (foundry_backend.py, runner.py)

• Expanded evaluator frozensets: response_completeness, groundedness_pro, retrieval, tool_selection added to existing sets
• New frozensets: _EVALUATORS_NEEDING_TOOL_DEFS_ONLY (tool_input_accuracy, tool_output_utilization, tool_call_success), _EVALUATORS_NEEDING_OUTPUT_ITEMS (task_adherence — uses {{sample.output_items}} instead of {{sample.output_text}})
• Fixed NLP evaluator names: bleu_score, rouge_score, gleu_score, meteor_score to match _to_builtin_evaluator_name conversion
• Added default init params: RougeScoreEvaluator requires rouge_type — defaults to rouge1
• Dynamic item_schema: Automatically includes tool_definitions and context fields when evaluators require them
• Refactored _default_foundry_input_mapping to frozenset-based routing (covers all evaluators, not just 4 hardcoded names)
• Improved error handling: Logs evaluator errors when score: null (e.g., safety evaluators in unsupported regions), improved runner error message with --verbose hint

Documentation (ci-github-actions.md)

• CI/CD Integration Models: PR quality gate, scheduled regression, post-deployment validation, multi-environment promotion, Azure DevOps pipeline
• Gating Best Practices: Threshold design, scenario-specific evaluator selection, comparison-based regression detection
• Supported Evaluators Reference: Complete table of 22 evaluators by category with inputs and requirements
• Troubleshooting: Safety evaluator region requirements, missing scores diagnosis

Tests

• ~20 new unit tests for all evaluator data_mapping patterns
• All 96 tests pass

Live Verification

All 22 evaluators verified end-to-end against live Foundry cloud evaluation (East US 2):

Category	Evaluators	Result
Quality	Coherence, Fluency, Relevance	✅
Agent	IntentResolution, TaskCompletion, TaskAdherence	✅
Similarity	Similarity, ResponseCompleteness	✅
RAG	Groundedness	✅
Safety	Violence, Sexual, SelfHarm, HateUnfairness	✅
Tool	ToolCallAccuracy, ToolSelection, ToolInputAccuracy, ToolOutputUtilization	✅
NLP	F1Score, BleuScore, GleuScore, RougeScore, MeteorScore	✅

Note

docs/analysis-issue-51-*.md are internal research/analysis documents created during the issue investigation. They should be removed before release — they are included in this PR for team review only.