fix(eval): trim legacy trajectory span history

[Chibionos]

Summary

• Fixes AE-1482 in the uipath-python evaluator layer by preventing the legacy trajectory evaluator from serializing full raw span attributes into the judge prompt.
• Reuses the compact tool-history formatter already used by LLMJudgeTrajectoryEvaluator, keeping tool name, arguments, and output while dropping prompt-bearing LLM spans and arbitrary span metadata.
• Adds a regression test with prompt-sized LLM/tool metadata to ensure AgentRunHistory does not reintroduce the duplicated system prompt.

Context

The reported trace bloat came from the legacy evaluator copying entire OpenTelemetry span payloads into the evaluation prompt. Raw LLM spans can legitimately contain growing message history, but the legacy trajectory judge does not need that full payload; it needs the compact tool trajectory.

This does not implement the separate sub-agent/deeper-trace trimming toggle discussed for trajectory scope. It addresses the immediate duplication/bloat bug in the legacy evaluator prompt construction.

Validation

• uv run pytest tests/evaluators/test_legacy_trajectory_evaluator.py tests/evaluators/test_evaluator_methods.py -k trajectory -q
• uv run ruff check src/uipath/eval/evaluators/legacy_trajectory_evaluator.py tests/evaluators/test_legacy_trajectory_evaluator.py
• uv run ruff format --check src/uipath/eval/evaluators/legacy_trajectory_evaluator.py tests/evaluators/test_legacy_trajectory_evaluator.py
• git diff --check

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
100.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud