fix(evaluation): reject parallelism=0 in EvaluateConfig/InferenceConfig

[gaurav-gandhi-2411]

Related

#6951 is where this gap was publicly established: a repo-wide sweep for
floor-less count/limit parameters in src/google/adk/evaluation/*.py,
posted as a comment on that issue, first flagged
EvaluateConfig.parallelism/InferenceConfig.parallelism alongside several
other candidates (a collaborator, @surajksharma07, engaged on the thread and
confirmed the related evaluate_eval_set empty-collection defect). A
follow-up comment on the same issue narrowed the sweep after two of the
other candidates turned out weaker than first stated, and reaffirmed this
one — EvaluateConfig.parallelism/InferenceConfig.parallelism at 0 — as
the confirmed, evidence-backed gap this PR fixes.

🔴 Required Information

Describe the Bug:
EvaluateConfig.parallelism and InferenceConfig.parallelism
(base_eval_service.py:46,72) have no lower-bound constraint. Both values
are forwarded, unclamped, straight to asyncio.Semaphore(value=parallelism)
(local_eval_service.py:186,216). A negative value fails loud —
asyncio.Semaphore itself raises ValueError immediately on construction.
A zero value fails silent: the semaphore constructs successfully, and every
subsequent .acquire() blocks forever, since a Semaphore(0) can never be
released by anything. In LocalEvalService.perform_inference/evaluate
this means the CLI (or any caller) hangs indefinitely with no exception, no
error message, and no evaluation ever completing.

Steps to Reproduce:

>>> import asyncio
>>> asyncio.run(asyncio.wait_for(asyncio.Semaphore(0).acquire(), timeout=5.0))
# TimeoutError after 5s — hangs, does not raise on construction

>>> asyncio.Semaphore(-1)
# ValueError('Semaphore initial value must be >= 0') — raised immediately

Expected Behavior:
Same as the existing convention for count/limit-like fields elsewhere in
this package — pydantic.ValidationError at construction time for a
degenerate value, not a silent hang at call time.

Observed Behavior:
EvaluateConfig(eval_metrics=[...], parallelism=0) and
InferenceConfig(parallelism=0) both construct successfully today, then
hang the first evaluation/inference run indefinitely.

Changes

• src/google/adk/evaluation/base_eval_service.py: add ge=1 to
  EvaluateConfig.parallelism and InferenceConfig.parallelism, matching
  the existing convention for count-like fields in this package —
  JudgeModelOptions.parallelism_limit (eval_metrics.py),
  ToolThreadPoolConfig.max_workers (run_config.py), and
  ContextCacheConfig.cache_intervals (context_cache_config.py) are all
  already ge=1.
• tests/unittests/evaluation/test_base_eval_service.py (new file): add
  test_evaluate_config_rejects_zero_parallelism and
  test_inference_config_rejects_zero_parallelism, each asserting
  parallelism=0 raises pydantic.ValidationError.

Testing

pytest tests/unittests/evaluation/test_base_eval_service.py tests/unittests/evaluation/test_local_eval_service.py -q
→ 26 passed

pytest tests/unittests/evaluation -q
→ 872 passed

No regressions in either run.

Risk & rollback

Additive validation only — narrows accepted input, does not change behavior
for any value ever actually used (the default is 4; no caller or test in the
repo constructs either config with parallelism<=0). Revert is a two-line
removal of ge=1.