[Yushu-Li]

I fixed unnecessary model requests when an LLM judge has zero weight.

The evaluator already ignores zero-weight feedback, but generate_all_with_context still calls every configured model before that filtering happens. A disabled judge therefore consumes a request, and an exception from it can discard the active judges' feedback. In an offline reproduction with weights [0.0, 1.0], both clients are called once.

Skip non-positive-weight models at dispatch. Keep a None response slot for each skipped model so the evaluator's response indices still match the normalized weights and prompt logs. Update the response/logging type annotations to describe those slots. Positive-weight model ordering, forwarded options, weighted aggregation, client initialization, and active-judge error propagation are unchanged. This does not add all-zero-weight configuration validation.

Five new async regression tests use the real ensemble and, for feedback checks, the real evaluator with a temporary evaluation file. They cover disabled judges in the first/middle/last positions, disabled-client errors before and after an active judge, unequal-weight aggregation and missing metrics, aligned prompt logging, ordinary response ordering/options, and active-client error propagation. On the original dispatch implementation, three test methods fail across six failing subtests/assertions; the two compatibility controls pass.

Validation on macOS / Python 3.12:

• All five new regression tests passed.
• All eight existing ensemble/LLM-feedback tests passed.
• OPENAI_API_KEY=test-for-offline-tests PYTHONPATH="$PWD" python -m unittest discover -s tests -p 'test_*.py' -v: 959 tests passed.
• Black and isort checks on the new test module, Python 3.10 syntax checks on all changed files, and git diff --check passed. Existing formatting in the production modules is preserved.

I used AI assistance for implementation, regression tests, and local validation. No paid model calls or model-backed integration runs were performed.