Because people working with timestamps like to write code that says things like:

    // schedule another run for tomorrow
    schedule_event(now() + 86400)

...which doesn't actually work when leap-seconds are involved (things start to drift by a second). If you specify that leap seconds get replayed, it works.

Possibly its a historical accident: UNIX was released before the first leap second, and before anyone could have known they were needed (back when seconds were defined in terms of Earth's movement). Around its release time, the definition of the second was switched (making leap seconds potentially needed). Maybe by the time anyone realized, switching from UTC to TAI would have been too painful?

Not to mention that you can't know the UTC-TAI offset more than a few months into the future. We can not predict which years will have leap seconds inserted.

Unix timestamps do not handle leap seconds well at all. Obvious things like t₂-t₁ fail to provide the number of seconds between t₂ and t₁.