Omit zero-millis fractional in resolved date-math timestamps

The previous {@code "yyyy-MM-dd HH:mm:ss.SSSSSSSSS"} format always
emitted a 9-digit fractional component, which Calcite's TIMESTAMP
literal parser rejected with IllegalArgumentException on
{@code yyyy-MM-dd HH:mm:ss.000000000}-shape inputs. Round at
millisecond precision and emit the {@code .SSS} suffix only when
non-zero so whole-second values (epoch-millis ending in {@code 000},
{@code "now"} resolved on an exact-second boundary) match the
canonical PPL form the analyzer accepts.

Fixes {@code testSearchWithNumericTimeRange} which previously failed
with {@code IllegalArgumentException: "2025-08-01 03:47:41."} after
the visitor resolved {@code latest=1754020061} (whole-second
millis-since-epoch) to a {@code .000}-suffixed literal.

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

Author: ahkcs

core/src/main/java/org/opensearch/sql/calcite/CalciteRelNodeVisitor.java

@@ -515,8 +515,16 @@ private static Optional<UnresolvedExpression> tryResolveOpenSearchDateMath(
     } else {
       return Optional.empty();
     }
+    // PPL TIMESTAMP literals downstream are parsed by Calcite, which accepts
+    // {@code yyyy-MM-dd HH:mm:ss} or {@code yyyy-MM-dd HH:mm:ss.SSS}. Round at
+    // millisecond precision; emit a fractional component only when non-zero so a
+    // whole-second value matches the canonical PPL form (the analyzer rejects
+    // {@code yyyy-MM-dd HH:mm:ss.000}-shape literals with a trailing-zero parse
+    // error).
+    long millis = result.getEpochSecond() * 1000 + result.getNano() / 1_000_000;
+    String pattern = (millis % 1000 == 0) ? "yyyy-MM-dd HH:mm:ss" : "yyyy-MM-dd HH:mm:ss.SSS";
     String formatted =
-        java.time.format.DateTimeFormatter.ofPattern("yyyy-MM-dd HH:mm:ss.SSSSSSSSS")
+        java.time.format.DateTimeFormatter.ofPattern(pattern)
             .withZone(java.time.ZoneOffset.UTC)
             .format(result);
     return Optional.of(AstDSL.timestampLiteral(formatted));