`EvalScorer.score()` applies missing-tests weight even when tests were skipped

[ChulingNH]

Found while reading backend/app/code/eval/scoring.py:112-172.

The weights are:

WEIGHTS = {
    "syntax": 0.30,
    "risks":  0.20,
    "lint":   0.10,
    "tests":  0.40,
}

In _score_dynamic, when the dynamic eval was SKIPPED, the method explicitly does:

if result.status == ExecutionStatus.SKIPPED:
    scores["tests"] = None  # Not applicable
    return scores

Then score() computes the weighted sum (line 145-148):

for dim, weight in self.WEIGHTS.items():
    if dim in dimensions and dimensions[dim] is not None:
        weighted_sum += dimensions[dim] * weight
        total_weight += weight

That branch correctly drops tests when None, and divides by total_weight instead of the full 1.0, so the math is sound for SKIPPED. Good.

But: _score_dynamic is only called when dynamic_result is not None (line 134). If a caller invokes score(static_result, dynamic_result=None), the tests key never enters dimensions at all — score() divides by 0.30 + 0.20 + 0.10 = 0.60 and the syntax dimension alone can contribute up to 100 * 0.30 / 0.60 = 50 to the overall. With risks=100, lint=100 we get an overall=100.0, grade A, even though no tests ran.

That's the bug: the static-only path produces grades equivalent to a tests-run path because total_weight is normalised away.

Fix options

1. Penalty for missing dynamic eval. If dynamic_result is None, treat tests as 0.0 (not None) so the divisor stays 1.0 and "no tests" caps overall at 60%.
2. Explicit "static-only" grade. Return EvalScore(overall=..., grade="N/A-static") when dynamic is absent, instead of pretending the score is comparable.

Option 1 is closer to how academic graders treat "didn't submit tests" — zero, not "skip the criterion".

Severity: Medium. The bug is invisible until you compare two projects, one of which ran tests and one of which didn't, and discover their overall scores are nearly identical.

[hiSandog]

A skipped dimension should either make the overall score explicitly incomplete or cause the remaining applicable weights to be renormalized; treating its 0.40 as an implicit zero conflates “not run” with “tests failed.” The result should expose both raw dimension scores and the denominator/coverage used so consumers can distinguish a high partial score from a fully evaluated one. Tests should cover skipped, failed, passed, and all-dimensions-unavailable cases.