fix(cli): adk eval process exit code now reflects PASSED/FAILED

[gaurav-gandhi-2411]

🔴 Required Information

Describe the Bug:
cli_eval (the adk eval CLI command) computes and prints a real per-eval-set
"Tests passed"/"Tests failed" summary (eval_run_summary) but never calls
sys.exit — the process always exits 0 regardless of the printed verdict,
indistinguishable from every test genuinely passing. This makes adk eval
unusable as a CI gate on its own exit code: a pipeline step that runs adk eval ... && deploy will deploy even when every eval case failed.

Every other subcommand in this file with a pass/fail outcome already sets a real
exit code: run --query does sys.exit(exit_code) (cli_tools_click.py:961,
derived from run_once_cli's return), test does sys.exit(1) on a missing
test runner (cli_tools_click.py:1060). eval was the one command whose exit code
carried no signal at all.

Steps to Reproduce:
Run adk eval <agent_module> <eval_set_file> against an eval set containing at
least one failing case, then check the process exit code (echo $?).

Expected Behavior: A non-zero exit code when the printed summary shows any
failed test, so the command is usable as a CI gate.

Observed Behavior: Exit code 0 regardless of the printed "Tests failed"
count — no exit code, prior to this fix, ever distinguished a real failure from
a real pass.

Environment: google-adk main branch (this repo), CLI entry point
cli_tools_click.py::cli_eval.

Why this fix

After the existing eval_run_summary loop, sys.exit(1) if any eval set
recorded a failed test, else sys.exit(0) — derived from the same counts
already being printed, no new computation introduced.

Changes

• src/google/adk/cli/cli_tools_click.py: cli_eval now calls sys.exit(1 if total_failed else 0) at the end, derived from the same eval_run_summary
  counts already being printed.
• tests/unittests/cli/utils/test_cli_tools_click.py:
  • test_cli_eval_with_eval_set_file_path / test_cli_eval_with_eval_set_id:
    both eval cases have no invocations and no configured criteria, so neither
    produces a PASSED verdict ("Tests failed: 1" in their own printed summary,
    unchanged by this fix) — their exit_code == 0 assertion was only ever true
    because the exit code carried no signal; updated to == 1 with a comment
    explaining why.
  • New test_cli_eval_exit_code_reflects_final_eval_status (parametrized,
    2 cases): mocks _collect_inferences/_collect_eval_results to return a
    canned EvalCaseResult with a controlled final_eval_status, asserting
    exit_code == 0 for PASSED and != 0 (1) for FAILED.

Testing Plan

Unit Tests:

• Added/updated tests above.
• All unit tests pass locally.
• The any_failed_exits_nonzero case of the new parametrized test fails
  against pre-fix code (verified by stashing the fix) — confirming it
  actually catches the bug.

pytest tests/unittests/cli/utils/test_cli_tools_click.py -k eval: 15 passed
pytest tests/unittests/cli/utils/test_cli_tools_click.py: 98 passed
pytest tests/unittests/evaluation/: 838 passed (unaffected — this repo's baseline
  for this branch's diff; unrelated to the eval-service tests touched by a sibling
  PR I'm also preparing)
pre-commit (ruff, isort, pyink, addlicense, codespell): all pass on changed files

20 unrelated pre-existing failures (conformance/markdown report, cleanup-unused-files,
path-normalizer, cloud-run-deploy tests) reproduce identically on a clean
upstream/main checkout with no changes — confirmed before attributing anything to
this change.

Manual E2E: ran adk eval against a real eval set with a failing case pre- and
post-fix from a built shell: pre-fix echo $? → 0; post-fix echo $? → 1.

Risk & rollback

Confined to the end of one CLI command function; no change to any return value,
schema, or the printed summary itself. The only behavior change is the process exit
code, which previously carried no information — any caller currently ignoring it is
unaffected; any caller that (incorrectly, given the old behavior) assumed 0 always
meant "passed" will now see accurate signal. Purely additive; revert is a clean
single-commit revert.

[gaurav-gandhi-2411]

Added a case this PR's own diff was missing: an empty `eval_run_summary` (no eval case ever evaluated, e.g. an empty eval set) previously took the same `sys.exit(0)` path as "every case passed" — indistinguishable from a genuine pass. It now raises instead of exiting 0.

This is the same defect class flagged in #6951 for the `AgentEvaluator` path and fixed there in #6952 (raise on an empty result set rather than silently returning success); this PR adopts the identical approach for `cli_eval`.

New test: `test_cli_eval_empty_summary_does_not_exit_zero`, confirmed failing pre-fix.