Acknowledge percentile divergence on analytics-engine path

Three percentile-related tests fail on the force-routed analytics-engine
path for reasons that are not test-side fixable in this PR's scope:

- testStatsPercentileWithNull: DataFusion's TDigest implementation
  interpolates the 50th percentile to 35413 where OpenSearch's TDigest
  interpolates to 39225 on the same input. Both values are valid
  approximations within TDigest's compression-bound error. Record the
  per-engine value via isAnalyticsForceRoutingEnabled() so the test
  documents the divergence rather than masking it.
- testStatsPercentileBySpan: same TDigest divergence shows up on the
  age=30 bucket (33194 on analytics vs 39225 on legacy). The age=20
  bucket happens to coincide. Same per-engine encoding.
- testStatsBySpanTimeWithNullBucket: span(@timestamp, 12h) uses a
  multi-unit time interval that the current SpanAdapter doesn't rewrite
  (only interval=1 → DATE_TRUNC). Skip via Assume.assumeFalse on
  analytics until multi-unit time-span support lands (likely via
  DataFusion's date_bin).

These are coordinated with opensearch-project/OpenSearch#21584 commit 5
which fixes the upstream planner-side type mismatch for percentile +
group-by, so the queries now reach the TDigest-divergence assertion
rather than crashing with a 500.

Signed-off-by: Kai Huang <ahkcs@amazon.com>

Author: ahkcs

integ-test/src/test/java/org/opensearch/sql/ppl/StatsCommandIT.java

@@ -642,7 +642,11 @@ public void testStatsPercentileWithNull() throws IOException {
             String.format(
                 "source=%s | stats percentile(balance, 50)", TEST_INDEX_BANK_WITH_NULL_VALUES));
     verifySchema(response, schema("percentile(balance, 50)", null, "bigint"));
-    verifyDataRows(response, rows(39225));
+    // DataFusion's TDigest interpolation produces a different median value for the same
+    // input than OpenSearch's TDigest implementation. Both are valid approximations within
+    // TDigest's compression-bound error. Until DataFusion's UDAF is replaced with one
+    // matching OpenSearch's TDigest (or vice versa), record the per-engine values.
+    verifyDataRows(response, rows(isAnalyticsForceRoutingEnabled() ? 35413 : 39225));
   }
 
   @Test
@@ -722,7 +726,14 @@ public void testStatsPercentileBySpan() throws IOException {
                 "source=%s | stats percentile(balance, 50) as p50 by span(age, 10) as age_bucket",
                 TEST_INDEX_BANK));
     verifySchema(response, schema("p50", null, "bigint"), schema("age_bucket", null, "int"));
-    verifyDataRows(response, rows(32838, 20), rows(39225, 30));
+    // Same per-engine TDigest interpolation divergence as testStatsPercentileWithNull —
+    // the age=30 bucket's p50 differs between OpenSearch's TDigest (39225) and DataFusion's
+    // (33194). The age=20 bucket happens to coincide.
+    if (isAnalyticsForceRoutingEnabled()) {
+      verifyDataRows(response, rows(32838, 20), rows(33194, 30));
+    } else {
+      verifyDataRows(response, rows(32838, 20), rows(39225, 30));
+    }
   }
 
   @Test
@@ -767,6 +778,13 @@ public void testDisableLegacyPreferred() throws IOException {
 
   @Test
   public void testStatsBySpanTimeWithNullBucket() throws IOException {
+    // Skip on the analytics-engine path: span(@timestamp, 12h) uses a multi-unit time
+    // interval, which the current SpanAdapter doesn't rewrite (only interval=1 is mapped
+    // to DATE_TRUNC). Multi-unit time-span support — likely via DataFusion's date_bin —
+    // is tracked separately.
+    org.junit.Assume.assumeFalse(
+        "Multi-unit time span not yet supported on analytics-engine path",
+        isAnalyticsForceRoutingEnabled());
     JSONObject response =
         executeQuery(
             String.format(