qualifying sectors based on neighbours

[tomato42]

currently the qualification mechanism (for latent or reallocated sectors) uses quantiles, in practice the behaviour of sectors is much more complex – while when the block is in good condition, the reads will mostly follow normal distribution, in case the block has issues, there will be re-reads, and the times of those re-reads will be distributed depending on number of times the disk had to re-read the sector

in short, if the sector is good, there will be one normal distribution for the read times
if the sector is in worse condition, the times will come from multiple normal distributions, each separated from the other by the rotational delay (8.33(3)ms for 7200 rpm disk)

using the neighbouring sectors, we can tell if the sector comes from one of the re-read distributions, or from a regular read; if the read is near the multiple of the rotational delay, it means that it really required a re-read, if it doesn't, it likely means that the delay (or at least part of the delay) came from external sources, so the sector should be re-read

Testing if a sample is from the same distribution as other samples can be done using Kolmogorov–Smirnov test.
Handling of blocks with track changes and the slowing down as the reads get to the end of the drive will require a bit more complex approach though.