We hope eventually to fix this issue (and some other problems described in the "FITS image format quirks" section of the PS1 DR2 caveats page), but since the PS1 archive includes 1.5 petabytes of PS1 images, the task of updating all the image files is not simple.
Why are the observation times in the header and catalog off by 35 seconds?
The times in the warp image headers and in the catalog (e.g., the obsTime column in the Detection table) are defined using international atomic time (TAI) rather than UTC time. Those times differ by the addition of leap seconds, which leads to header times that differ by 34 or 35 seconds from the UTC times. (See Rots et al. 2015 for more details.) If you are concerned with timing at this level, you may need to convert the times to UTC. For the warp images, the fix for this is to insert the keyword TIMESYS = 'TAI' in the header. FITS cutout images have a correct TIMESYS keyword (as of 2022 January 20), but full skycell FITS images do not have a TIMESYS keyword. For the database times, the fix is to add 34 or 35 seconds (depending on the date) to the TAI time to get the UTC time.
Thanks to Peter Van Wylen for discovering this issue and identifying the fix.