So, friends, I did the experiment.
The setup was two 5mm LEDs (I'm not sure what type exactly, but most probably they have 60 degs of light distribution and 40 mcd of maximum light - still didn't get the way they measure this intensity): red with 330 Ohm series resistor and green with 160. Both with 5V supply and AVR microcontroller.
With this setup I was able to see the blink as short as 1 us for green and 2 us for red LED. I should point out that I was in well lit room but I put my palms around the LEDs to make a 3 inches deep well around LEDs. I looked directly on the LEDs and I was expecting the blink. So this light is definitely not enough to notice the blink if you are not expecting one.
The current can be estimated as 3.8 Volts / 330 Ohms = 11,5 mA for red and 23 mA for green.
So the electrical power is 11,5 mAmps * 1,2 Volts = 14 mW for red and 28 mW for green.
Sequentially the blink electrical energy was as low as 28 nJ (nano Joules !!!) in both cases. Which is about ten times more than I expect to spend on a blink!
I test this on my wife and my 7-yo daughter. Same thing.
Regarding the energy distribution versus time:
Unfortunately I wasn't able to change resistors so I made just one thing: I put the LED to a constant light mode with 1% PWM. And I did not notice any difference if I change the frequency (1 us blink each 100 us is equally lit as 100 us blink each 10 ms). This is not exactly what I need but it looks like it's not a big deal how I will distribute the power in time.
Regarding the sensitivity of the different areas of an eye: I was able to see the blink only if I look exactly on the LEDs. If I shift the eye sight axis a little bit - I wasn't able to see anything. The same thing I noticed with constant lighting.