fix(evals): stop comments satisfying an expected API

[salmanmkc]

Description

A file that calls createGestureDetector, onThumbsUp and spawnConfetti, none of which exist, scores 0.778 out of 1.0 on gestures-thumbs-up. It gets there because the task lists "thumbs" in expected_apis, matching is a substring check over the whole file, and the file has a comment reading "detect a thumbs up gesture". So it collected a third of the API dimension for prose.

That's the dimension the whole thing exists to measure, and it was being paid out for a comment.

Now only comments are stripped before matching. Same file scores 0.0 on APIs and 0.667 composite, which is just the floor for parsing and not tripping the forbidden patterns.

I did try stripping string literals too, and that was wrong. Gesture names are passed as strings here, so setGestureEnabled('thumbs-up', true) is real usage and correct code would have started failing. Imports were never going to work that way either, since a specifier like @dimforge/rapier3d only ever appears inside a string.

Second commit tightens the four tasks that listed a bare word. "thumbs" becomes "thumbs-up", which is the actual gesture name, so invented code writing "thumbs up" with a space gets nothing. "query" becomes ".query(" so it has to be a call rather than the English word, and that still matches if someone aliases const ai = xb.ai.

On gestures-thumbs-up the invented file now scores 0 of 3 APIs, and one using enableGestures, setGestureEnabled('thumbs-up') and the gestureRecognition event target scores 3 of 3.

Tests are stdlib unittest, so nothing to install:

python -m unittest discover -s evals/prototypes -p 'test_*.py'

Neutering the stripper fails six of the nine, so they're actually holding something.

Worth saying plainly: these four tasks won't be comparable with the earlier published run, since that one handed out some of this credit for prose. Probably makes the with-skills vs without-skills gap look bigger than before, since invented code was sitting on a floor closer to 0.78 than 0.67.

Type of Change

• Bug fix
• New feature / enhancement
• New demo or sample
• Documentation update

Media / Screen Recordings & Screenshots (If Applicable)

Nothing visual, it's the scoring script for the eval prototypes.

Checklist

• Tested in simulator & device: Doesn't touch the SDK or anything that runs in XR. Verified by scoring an invented file and a correct one against gestures-thumbs-up before and after, plus 9 unit tests on the comment stripper.
• Large Assets ($\ge$ 1MB): None.
• SDK Dynamic Dependencies: No new dependencies, the tests use stdlib unittest rather than adding pytest.
• Security: No keys or secrets.