fix(evaluation): avoid crash on first turn with empty user content

[anxkhn]

Please ensure you have read the contribution guide before creating a pull request.

Link to Issue or Description of Change

1. Link to an existing issue (if applicable):

No existing issue. Describing the bug below per the issue-template structure, as
CONTRIBUTING permits for a PR without an associated issue.

2. Or, if no issue exists, describe the change:

Problem:

PerTurnUserSimulatorQualityV1._evaluate_first_turn crashes with
AttributeError: 'NoneType' object has no attribute 'strip' when the first
invocation's user_content is a non-None Content whose parts are empty
(parts=[]) or None (types.Content()).

The method guards only the first_invocation.user_content is None case, then
does:

score = int(
    get_text_from_content(first_invocation.user_content).strip()
    == conversation_scenario.starting_prompt.strip()
)

get_text_from_content (llm_as_judge_utils.py) returns None when the content
has no text parts, because if content and content.parts: falls through for
empty/None parts:

if content and content.parts:
    return "\n".join([p.text for p in content.parts if p.text])
# implicit return None otherwise

So .strip() is called on None and raises. This is reachable through the
public Evaluator.evaluate_invocations, which calls
_evaluate_first_turn(actual_invocations[0], ...): any evaluated conversation
whose first user turn was built with empty parts (or carries only non-text parts
after text extraction returns None) crashes the evaluation instead of producing
a result.

Reproduction (text extraction results that reach _evaluate_first_turn):

                    parts=[]  ->  None      # old code: .strip() -> AttributeError
             types.Content()  ->  None      # old code: .strip() -> AttributeError
                   text part  ->  'hi'       # scored normally
     function_call-only part  ->  ''         # scored (legitimately FAILED)

Solution:

Extract the text once and return a NOT_EVALUATED PerInvocationResult when it
is None, before the comparison. This mirrors the two guards already in this
file: the user_content is None guard immediately above, and
_convert_llm_response_to_score's if response_text is None or not response_text: return AutoRaterScore().

user_text = get_text_from_content(first_invocation.user_content)
if user_text is None:
    return PerInvocationResult(
        actual_invocation=first_invocation,
        eval_status=EvalStatus.NOT_EVALUATED,
    )

score = int(
    user_text.strip() == conversation_scenario.starting_prompt.strip()
)

The fix is behavior-preserving for every case that did not crash: the guard is
is None only, so a non-text part (text extraction returns "", not None) is
still scored and legitimately marked FAILED. Only the two previously-crashing
inputs (parts=[], types.Content()) now return NOT_EVALUATED.

Testing Plan

Unit Tests:

• I have added or updated unit tests for my change.
• All unit tests pass locally.

Added a parametrized regression test
test_evaluate_first_turn_not_evaluated_when_user_content_has_no_text covering
types.Content(role="user", parts=[]) and types.Content(), asserting the
result is NOT_EVALUATED (score None) instead of raising.

$ pytest tests/unittests/evaluation/simulation/test_per_turn_user_simulation_quality_v1.py -q
33 passed

$ pytest tests/unittests/evaluation -q
486 passed

Effectiveness check: temporarily reverting the new guard makes the two new
parametrized cases fail with the exact
AttributeError: 'NoneType' object has no attribute 'strip'; restoring the guard
makes them pass. The test is not vacuous.

Manual End-to-End (E2E) Tests:

Not applicable; this is a unit-level crash guard in the evaluation module,
covered by the added regression test. The behavior is demonstrated by the
reproduction table above (empty/None parts extract to None, which previously
hit .strip()).

Checklist

• I have read the CONTRIBUTING.md document.
• I have performed a self-review of my own code.
• I have commented my code, particularly in hard-to-understand areas. (No new
  comments needed; the guard mirrors the existing ones directly above/below it.)
• I have added tests that prove my fix is effective or that my feature works.
• New and existing unit tests pass locally with my changes.
• I have manually tested my changes end-to-end. (Unit-level fix; covered by
  the regression test rather than an E2E flow.)
• Any dependent changes have been merged and published in downstream modules. (None.)

Additional context

The two new NOT_EVALUATED cases are consistent with how the aggregator already
handles non-evaluated turns (num_evaluated == 0 yields an EvaluationResult
without an overall_score), so a conversation whose only turn is empty no longer
crashes and simply contributes nothing to the score.

---

Diff summary

• src/google/adk/evaluation/simulation/per_turn_user_simulator_quality_v1.py
  (+10/-2): compute user_text = get_text_from_content(...) once; return
  PerInvocationResult(eval_status=NOT_EVALUATED) when user_text is None; reuse
  user_text in the comparison.
• tests/unittests/evaluation/simulation/test_per_turn_user_simulation_quality_v1.py
  (+26): parametrized regression test over types.Content(role="user", parts=[])
  and types.Content().

[adk-bot]

Thank you @anxkhn for your contribution! 🎉

Your changes have been successfully imported and merged via Copybara in commit c9bacd4.

Closing this PR as the changes are now in the main branch.