What you found
Treatment was targeted on X the whole time, but the selection-bias term went to zero the moment X stopped moving Y(0). ATT and ATC stayed apart, because the targeting rule still picked the units with the biggest effects. Those are two different things, and only one of them biases the naive comparison.
E[Y|D=1] − E[Y|D=0] = E[Y(1) − Y(0) | D=1] + ( E[Y(0)|D=1] − E[Y(0)|D=0] )
The first term is ATT. The second is selection bias, and it contains only Y(0): no Y(1), no τ, nothing about how the effect varies. Targeting on who benefits changes ATT relative to ATC — it does not, by itself, bias the naive gap for ATT. Targeting on who starts higher does.
Random assignment removes both: the treated and control groups become interchangeable in Y(0) and in τ, so the naive gap, ATT, ATC and ATE all coincide (up to the luck of the draw with ten units).