I took a look through most of the EDF spec documentation and found the following for duration of a data record:
Q9. Do non-integer sampling frequencies (like 1/30 Hz) cause problems?
Not necessarily. Good viewers will count samples and compare these with "number of samples in a datarecord" and in this way count how many datarecords have been passed (and consequently how many "duration's of a datarecord"). Because this is all integer computation, there are no round-off errors. This is why EDF recommends the "duration of a datarecord" to be an integer number of seconds. In the 1/30 Hz example, "duration of a datarecord" and "number of samples in a datarecord" can be 30 and 1, respectively. Or 3600 and 120, respectively.
However, if a sampling frequency is 999.98Hz (for instance due to small inaccuracy of the ADC clock), 'integer EDF' would use datarecords of 50000s containing 49999 samples of each signal. Even if only one signal is in the file, there would be more than 61440 bytes in a datarecord. The official specs say that in that case the duration should be a float value less than 1s. This will inevitably cause a small round-off error in the timing. Item 10 of the programming guidelines explains that this error is negligible, even in extreme cases.
I may move your change to the code base if we find an influx in EDFs that do this.
In any case, thanks for all the information, definitely helps out the development of tools!