fix(evaluation): preserve trace boundaries in eval sets

[linhongyu510]

Summary

Fixes #1021

Fix trace-to-eval-set conversion when a tracing file contains more than one trace.

Previously, conversion state such as conversation, app_name, and user_id was shared across trace groups, while eval_cases.append() ran after the grouping loop. This could collapse multiple traces into one eval case and mix tool calls or metadata across trace boundaries.

This change:

• creates one isolated eval case per trace
• sorts spans by start_time before deriving input, output, and tool order
• resets conversation and metadata state for every trace
• includes the trace ID in generated eval IDs
• derives the eval-set timestamp from the earliest generated case

The regression test uses two interleaved, out-of-order traces and verifies that conversations, tool calls, metadata, and timestamps remain isolated.

Validation

• latest main with the regression test: fails with 1 eval case instead of 2
• uv run pytest tests/test_evaluator.py -q: 3 passed
• uv sync --all-extras && uv run pytest -n 16 -q: 2929 passed, 6 skipped, 6 subtests passed
• ruff check veadk/evaluation/base_evaluator.py tests/test_evaluator.py
• ruff format --check veadk/evaluation/base_evaluator.py tests/test_evaluator.py
• git diff --check

AI assistance

This change was developed with AI assistance. I reviewed the trace grouping logic, rebased it onto the latest main, and verified the regression against both the unfixed and fixed implementations.

[linhongyu510]

A previously completed 2929-pass run is documented in Issue #1021, but it was on the pre-rebase head and is not evidence for 7a83eb1., but it was on the pre-rebase head and is not evidence for 7a83eb1.

[linhongyu510]

Full-suite follow-up for current head 7a83eb1: I created a clean isolated Python 3.12.13 environment outside the repository, ran uv sync --all-extras (288 locked packages), then executed python -m pytest -n 16 -q. Final result: 3175 passed, 5 skipped, 66 warnings, 6 subtests passed in 396.41s. The OTLP exporter logged localhost:8000 connection failures during teardown, but pytest completed normally with a successful final summary. This supersedes my earlier incomplete full-suite attempt.

[linhongyu510]

Full-suite follow-up for current head 7a83eb1: I created a clean isolated Python 3.12.13 environment outside the repository, ran uv sync --all-extras (288 locked packages), then executed python -m pytest -n 16 -q. Final result: 3175 passed, 5 skipped, 66 warnings, 6 subtests passed in 396.41s. The OTLP exporter logged localhost:8000 connection failures during teardown, but pytest completed normally with a successful final summary. This supersedes my earlier incomplete full-suite attempt.