Bug #36698 ROW_COUNT() is reset inconsistently

Description:
ROW_COUNT() is reset to -1 by a SELECT statement, but not by an INSERT statement which fails due to duplicate key constraints.

How to repeat:
CREATE TABLE t1 (id int primary key);

INSERT t1 VALUES (1); -- row count set to 1
SELECT 1; -- row count set to -1
INSERT t1 VALUES (1); -- row count untouched
SELECT ROW_COUNT(); -- Returns -1

TRUNCATE TABLE t1;

INSERT t1 VALUES (1); row count set to 1
INSERT t1 VALUES (1); row count untouched
SELECT ROW_COUNT(); -- Returns 1

Suggested fix:
An INSERT that fails should reset ROW_COUNT() to -1, just like a SELECT statement does, or else a SELECT statement should not reset ROW_COUNT(). 

For consistency, I'd say that the latter is preferable, since that's how functions like LAST_INSERT_ID() behave.

For the particular issue that raised this bug, a workaround is found by handling failed inserts differently, to avoid calling ROW_COUNT() in an ambiguous situation.

This may already have been addressed with bug 21818.