--- Quote Start ---
It is random per place and route.
When it fails, it always fails at the same point after several resets or re-sof.
I change nothing between build.
no multi cycles. Few false path but for debug ports.
Voltage is ok (1.1V). Clocks are ok.
--- Quote End ---
If you change nothing between builds then I assume the fitting results should be more or less identical. Therefore (my guess) it is likely your problem is to do with reset functionality per se which appears at power up or afterwards. I believe the place and route variation is just coincidence of same issue.
I mean some registers are not under reset. You need to check that your control signals are all under reset. You can also uncheck powerup don't care in the fitter settings so that all registers power up low(except preset ones). Finally you may add signaltap to capture the events at failure. You may find a work around if you play with the duration and release of reset until you find a stable window.
also check for any floating inputs which are not driven at your development stage