As suspected, this was caused by incorrect shader state management when switching between multiple shaders (thanks Firefox for verbose warnings).

Also, testing WebGL performance is very tricky. You can get different performance depending on the state of browsers / system (messed up state can be 10-66% slower):

1. freshly started Chrome can be much faster than Chrome that already ran few WebGL demos (which are now closed)
2. demo opened in new Chrome tab can be faster than demo reloaded in the same tab
3. even demos opened and closed in Firefox can slow down Chrome demos

This suggests that browsers do not clean GPU state properly :(