fix inconsistent results in ARM exact tests by removing purification from tests

Author: paulbkoch

shared/libebm/tests/boosting_unusual_inputs.cpp

@@ -2088,9 +2088,11 @@ TEST_CASE("stress test, boosting") {
       // terms.push_back({0, 1, 2, 3}); // TODO: enable when fast enough
    }
    const size_t cRounds = 200;
-   std::vector<IntEbm> boostFlagsAny{TermBoostFlags_PurifyGain,
+   std::vector<IntEbm> boostFlagsAny{// TermBoostFlags_PurifyGain,
          TermBoostFlags_DisableNewtonGain,
          TermBoostFlags_DisableCategorical,
+         // TermBoostFlags_PurifyUpdate,
+         // TermBoostFlags_GradientSums, // does not return a metric
          TermBoostFlags_DisableNewtonUpdate,
          TermBoostFlags_RandomSplits};
    std::vector<IntEbm> boostFlagsChoose{TermBoostFlags_Default,
@@ -2099,10 +2101,10 @@ TEST_CASE("stress test, boosting") {
          TermBoostFlags_MissingSeparate,
          TermBoostFlags_MissingDrop};
 
-   double validationMetric = 0.0;
+   double validationMetric = 1.0;
 
    for(IntEbm classesCount = Task_Regression; classesCount < 5; ++classesCount) {
-      if(classesCount != Task_Regression && classesCount < 2) {
+      if(classesCount != Task_Regression && classesCount < 1) {
          continue;
       }
       const auto train = MakeRandomDataset(rng, classesCount, cTrainSamples, features);
@@ -2159,9 +2161,13 @@ TEST_CASE("stress test, boosting") {
                                                  .validationMetric;
             }
          }
-         validationMetric += validationMetricIteration;
+         if(classesCount == 1) {
+            CHECK(std::numeric_limits<double>::infinity() == validationMetricIteration);
+         } else {
+            validationMetric *= validationMetricIteration;
+         }
       }
    }
 
-   CHECK(validationMetric == 42031.143270308334);
+   CHECK(validationMetric == 62013566170252.117);
 }