[Yushu-Li]

I fixed ensemble weights becoming unusable during normalization. For example, two finite weights of 1e308 sum to infinity, so the current division turns both normalized weights into zero and model sampling fails. Negative weights can also silently become positive probabilities when every weight is negative.

The constructor now rejects non-finite or negative weights, and non-empty ensembles with no positive weight, before initializing model clients. If the sum of otherwise valid finite weights overflows, it rescales by the largest weight before normalizing. Ordinary normalization, zero-weight slots, the caller's configurations and empty placeholder ensembles keep their existing behavior.

Validation:

• Seven new test methods cover equal/unequal overflowing sums, NaN/infinities, negative and all-zero weights, client initialization side effects, ordinary/integer/subnormal weights and empty placeholders.
• The regression methods reproduce the failures on unchanged main; existing-behavior controls pass.
• All 20 focused ensemble/judge tests passed.
• Full local unittest suite: 985 passed.
• Black checks and git diff --check passed.

I directed this contribution and used Codex assistance for reproduction, implementation, tests and code review. Verification used offline clients and unit-test fixtures; no paid provider calls were made.