Evaluation: Streamline AI summary model #1130

Is your feature request related to a problem?
The best-effort run-level AI summary on judge evaluations was generated with OpenAI (gpt-5-mini, Responses API), while prompt improvement already ran on Anthropic. Two providers on the same feature area meant two credential paths and a duplicated Anthropic error-handling ladder.

Describe the solution you'd like

• Switch EVAL_SUMMARY_MODEL default from gpt-5-mini to claude-sonnet-4-6.
• generate_run_ai_summary builds its own Anthropic client from the platform-owned ANTHROPIC_API_KEY — drop the session / openai_client params; update the caller in crud/evaluations/fast.py.
• Use one messages.create call with structured JSON output (output_config json_schema, {"summary": ""}) instead of free-text Responses parsing via map_kaapi_to_openai_params + extract_response_text, removing the mapper's DB lookup from this path.
• Keep the best-effort contract: no API key, provider error, bad JSON, or empty note → ai_summary is None and the run still succeeds.
• Extract the shared 6-branch Anthropic error ladder into log_anthropic_error(exc, *, fn_name, context) in services/llm/providers/claude.py; use it from both summary.py and prompt_improvement.py.
• Update tests (test_run_ai_summary.py, test_fast_judge.py) and docs/wiki/modules/evaluations.md.

Comments: No API surface change. Deployments pinning EVAL_SUMMARY_MODEL to an OpenAI model must switch to a Claude model, and ANTHROPIC_API_KEY must be set or summaries silently stay empty.

Tracked and completed in #1128.