FIX Keep original_float_value in line with the verdict for inverted and combined scores

[Utkarsh Bahuguna (u7k4rs6)]

Fixes #2894.

TrueFalseInverterScorer now drops original_float_value from the inverted score, since the wrapped threshold float describes the uninverted verdict. The true/false aggregators drop it when they combine more than one score, and keep it for a single score, where it still matches. normalize_score_to_float then falls back to the verdict for inverted and combined scores.

Tests use real PlagiarismScorer + FloatScaleThresholdScorer scores with no mocks: the inverted score normalizing to its verdict with no leaked float, a two-scorer AND composite normalizing to 0.0, and a single-scorer composite keeping the float. The first two fail on main. tests/unit/score: 2832 passed.

[Roman Lutz (romanlutz)]

I responded on the issue.

[Utkarsh Bahuguna (u7k4rs6)]

Went with that in 24fb820. The inverter now drops the child float instead of flipping it, so inverted and multi-score results both normalize from their own verdict. Direct threshold scores and single-scorer composites keep it.

[Roman Lutz (romanlutz)]

This makes sense to me but I want Richard Lundeen (@richlundeen) to take a look. AFAIK this won't break anything.

[Utkarsh Bahuguna (u7k4rs6)]

Makes sense to land #2916 first. I tried merging it into this branch locally and it goes in cleanly. The only thing that changes is the assertion in test_threshold_and_inverter_preserve_each_judgment_async that expects original_float_value on the inverted result, which this PR would flip to not present. With #2916 the threshold's own float stays on its intermediate score, so dropping it from the wrapper's copy doesn't lose anything. I'll merge main here and update that test once #2916 is in.

[Utkarsh Bahuguna (u7k4rs6)]

Merged main now that #2916 is in, and updated test_threshold_and_inverter_preserve_each_judgment_async in 7653834. The inverted result no longer has original_float_value, and the retained threshold score still has 0.8. tests/unit/score and tests/unit/executor pass.

[Utkarsh Bahuguna (u7k4rs6)]

The merge queue failures were my test fixture still calling add_message_to_memory, which main now deprecates and the suite turns into an error. It uses add_message_to_memory_async now (0cb7b7e), and I merged main again. tests/unit/score and tests/unit/executor pass locally. Auto-merge should be able to pick it up once CI runs.

[Utkarsh Bahuguna (u7k4rs6)]

Richard Lundeen (@richlundeen) CI is green on 0cb7b7e now, but my push turned auto-merge off. Could you re-enable it when you get a chance?

[Utkarsh Bahuguna (u7k4rs6)]

The one failure here is unrelated: tests/unit/scenario/core/test_scenario_partial_results.py::test_run_async_cancellation_persists_progress_cleans_workers_and_resumes timed out on macOS 3.13 at its 5s wait_for (line 506), and the rest of the matrix got cancelled after it. This PR only touches the score wrappers, the test passes locally (5 of 5), and main's latest build_and_test also has an async timeout flake (c2ae5ee). The branch is already up to date with main. Could someone re-run the failed jobs? I don't have rights to.