Withhold experiment comparisons when attribution is invalid

Gate experiment charts on comparison validity, not merely on whether each denominator is nonzero. Render a genuine 0/n as 0% and 0/0 as unavailable. Withhold the arm-to-arm comparison when allocation changed during the window, episode identity turned over, or attribution is otherwise invalid, even if separate per-arm rates are computable.

Keep each percentage beside its numerator and denominator, and define the operational outcome in the interface. A status such as "issue closed" must not silently become "user recovered" or a causal claim. This pattern makes descriptive telemetry more honest; it cannot repair missing attribution or prove causality.