Likely bug to be fixed

I think there is a bug in the Gregorian initialization because it does account for the hour in the reference epoch.

The gregorian initialization and formatter should account for the reference epoch of the time scale.

src/timescale/mod.rs

@@ -135,6 +135,13 @@ impl TimeScale {
             _ => 1900,
         }
     }
+
+    pub(crate) const fn ref_hour(&self) -> i64 {
+        match self {
+            TimeScale::ET | TimeScale::TDB => 0,
+            _ => 12,
+        }
+    }
 }
 
 #[cfg_attr(feature = "python", pymethods)]

tests/epoch.rs

@@ -900,11 +900,13 @@ fn test_from_str() {
 }
 
 #[test]
-fn test_from_str_tdb() {
-    use core::str::FromStr;
-
-    let greg = "2020-01-31T00:00:00 TDB";
-    assert_eq!(greg, format!("{:e}", Epoch::from_str(greg).unwrap()));
+fn test_ref_epoch_delta() {
+    let e_tai = Epoch::from_gregorian_at_noon(2022, 9, 6, TimeScale::TAI);
+    let e_tdb = Epoch::from_gregorian_at_noon(2022, 9, 6, TimeScale::TDB);
+    // Since we store the durations with respect to their time scale reference epochs,
+    // the difference between the same epoch but in different time scales should be
+    // exactly the difference between the reference epochs.
+    assert_eq!(e_tai.duration - e_tdb.duration, J2000_TO_J1900_DURATION);
 }
 
 #[test]