bench: validate rounded Score and isolated zero-fit results

[Levius-Fubuki]

Purpose

Decider rounds a derived Score to two decimals and its probabilities to four. The benchmark currently rejects a valid score=0.88 with probabilities 0.1234/0.8766 because it compares against 0.8766 with a fixed 1e-4 tolerance. Bound both independent rounding errors, use an accurate weighted sum, and require the score to remain inside the exact level range.

Also recognize the released isolated-Score zero-fit outcome only when its zero probabilities and complete fit/calibration metadata are consistent. Other invalid distributions still fail, and wire probabilities are preserved without renormalizing them. The net diff is one production Python file.

Test Plan

Baseline: 7f39ac40902c374803992407bb26eeba29c8a588 (main). Head: de34df45dd42f290c030b094fb630d8e65a82052.

• python -m unittest discover -s tests/benchmarks -p 'test_*.py' -v
• python benchmarks/bench.py validate benchmarks/smoke.jsonl
• External regression harness: pinned legitimate Score and zero-fit responses; 1,152 deterministic independently rounded distributions across 2–10 levels; reversed label order; exact range; inconsistent/nonfinite values; incomplete/contradictory zero-fit metadata; huge integer metadata through actual request_one transport handling.
• Full diff and independent review; git diff --check.

Test Result

8 existing tests and 4 smoke requests pass. All 10 external regressions pass, including the 1,152-distribution corpus. Before the change, the focused regression suite reproduced four validation errors and an out-of-range acceptance. Independent review found huge integer fit metadata could raise an uncaught overflow; equality guards now reject it before float conversion, and six end-to-end malformed-response cases are recorded as invalid_response.

Demo / evidence

CPU wire validation only; no GPU inference, numerical model parity, latency or throughput claim. Python 3.12.13, httpx 0.28.1, macOS ARM64. Reference: decider-ai 1.8.1 source, Decider-2B v11 revision. Scalar rounding bound is 0.005 + 0.00005 * n * (n-1) / 2 + 1e-9; probability-sum allowance from #75 is unchanged.

Regression scripts, fixtures, logs and reviewer reproductions remain local artifacts following the author's core-code-only diff preference. Existing repository tests are reproducible with the commands above; the additional local harness is not part of repository CI. Demo/video: N/A for a response validator.

Self-review

• Reviewed the full diff and addressed findings.
• Focused benchmark validation change; no serving/model architecture changes.
• Ran relevant checks and reported their results and limitations.
• Description and claims match the implementation and available evidence.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.