Fix "read-tree -m A B" priming the cache-tree

In 456156d a shortcut to priming the index tree reference was
introduced, but the justification for it was completely bogus.

"read-tree -m A B" is to take the index (and the working tree)
that is largely based on (but does not have to match exactly) A
and update it to B, while carrying the local change that does
not overlap the difference between A and B, so there is no reason
to expect that the resulting index should match the tree B.

Noticed and test provided by Heiko Voigt.

Signed-off-by: Junio C Hamano <gitster@pobox.com>

builtin-read-tree.c

@@ -199,14 +199,9 @@ int cmd_read_tree(int argc, const char **argv, const char *unused_prefix)
 	 * "-m ent" or "--reset ent" form), we can obtain a fully
 	 * valid cache-tree because the index must match exactly
 	 * what came from the tree.
-	 *
-	 * The same holds true if we are switching between two trees
-	 * using read-tree -m A B.  The index must match B after that.
 	 */
 	if (nr_trees == 1 && !opts.prefix)
 		prime_cache_tree(&active_cache_tree, trees[0]);
-	else if (nr_trees == 2 && opts.merge)
-		prime_cache_tree(&active_cache_tree, trees[1]);
 
 	if (write_cache(newfd, active_cache, active_nr) ||
 	    commit_locked_index(&lock_file))

t/t1001-read-tree-m-2way.sh

@@ -392,4 +392,20 @@ test_expect_success \
      git ls-files --stage | tee >treeMcheck.out &&
      test_cmp treeM.out treeMcheck.out'
 
+test_expect_success '-m references the correct modified tree' '
+	echo >file-a &&
+	echo >file-b &&
+	git add file-a file-b &&
+	git commit -a -m "test for correct modified tree"
+	git branch initial-mod &&
+	echo b >file-b &&
+	git commit -a -m "B" &&
+	echo a >file-a &&
+	git add file-a &&
+	git ls-tree $(git write-tree) file-a >expect &&
+	git read-tree -m HEAD initial-mod &&
+	git ls-tree $(git write-tree) file-a >actual &&
+	test_cmp expect actual
+'
+
 test_done