unicode/norm: do not compose across a ccc=0 character that combines backward

unicode/norm/composition.go

@@ -506,6 +506,16 @@ func (rb *reorderBuffer) compose() {
 			}
 		}
 		b[k] = b[i]
+		if b[k].ccc == 0 {
+			// b[i] is itself a starter, so it becomes the base for what
+			// follows and blocks any later mark from an earlier starter
+			// (D115). Updating s only in the combinesBackward branch above
+			// misses this: a ccc-0 character that combines backward does not
+			// end the segment, so it reaches here with s still pointing at
+			// the previous starter, and a following mark then composes
+			// across it.
+			s = k
+		}
 		k++
 	}
 	rb.nrune = k

unicode/norm/normalize_test.go

@@ -653,6 +653,21 @@ var appendTestsNFC = []AppendTest{
 	{"\uff9e", grave(30), "\uff9e" + grave(29) + cgj + grave(1)},
 	{grave(30), "\uff9e", grave(30) + cgj + "\uff9e"},
 
+	{ // A ccc=0 character blocks composition (UAX #15, D115) even when
+		// it combines backward and so cannot end its segment. U+09BE
+		// composes with U+09C7, but here it only stands between the
+		// starter and U+0300, which must therefore not compose with "i".
+		"",
+		"i\u09be\u0300\u0316",
+		"i\u09be\u0316\u0300",
+	},
+	{ // the same shape with a ccc=0 character that does not combine
+		// backward, which has always been handled correctly
+		"",
+		"i\u0903\u0300\u0316",
+		"i\u0903\u0316\u0300",
+	},
+
 	// Tests designed for Iter.
 	{ // ordering of non-composing combining characters
 		"",