For the case when windings are tightly coupled, specifically for single layers 1:1, the geometry is close to a parallel-wire transmission line, and that gives leakage inductance proportional to winding length and Zo. It's a bit tougher when sections contain multiple layers (the layers act against each other in this way, reducing the coupling of more distant layers and multiplying the effect of wire length). This is easy to calculate.
Things become tricky when the geometry has no close contact, e.g. split bobbin, or windings on separate legs of the core. In this case, it may be easier to work with a coupled-inductors model, and later convert the coupling factor to leakage as needed.
Note that the secondary-shorted inductance has a maximum value of the primary magnetizing inductance, when k = 0. That is, there's no difference between Lp(s/oc) and Lp(s/sc). We can't really use an L network equivalent (i.e., series LL, parallel Lm (m = magnetizing), ideal transformer) to describe this situation. We need to use a pi equivalent (with Lp (shunt), Lm (series, m = mutual), Ls (shunt) and an ideal transformer), where Lm can take negative values if we set the transformer ratio to 1. (There always exists a pi or tee equivalent network for any transformer, but the values need not be physical, specifically for the case of large turns ratios and k I think.)
So, understand that when k << 1, it may be less meaningful to consider leakage inductance.
For a split-bobbin configuration, we can model each winding as a source of MMF (magnetomotive force), and estimate the reluctance between and around them. For an EE core for example, we can model the winding window (which is now partially unoccupied, i.e. the gap between windings) as a wide air gap, and thus it has a relatively high reluctance which shunts some flux from between windings (i.e., leakage), while most of the flux proceeds around the magnetic (core) loop. Note that the core is gapped, so has notable reluctance by itself, and this makes the window leakage relevant. How much so? Depends on the ratio of reluctances. We can adjust the air gap to trade off between them (while adjusting Np, Ns to maintain Lp, Ls as needed).
So, say we take the winding window past the middle of the winding, to the end of the core. This has a width of (window length)/2 (assuming the coil is centered), height of (core height or thickness), and length of (window height). This rectangular volume has an inductance of A_L = 2 mu_0 w h / l. (2 because the E core is symmetrical. Or put another way: remember to add up the "height" from both windows either side of the center leg.)
FWIW, A "tall" aspect ratio 'E' core, by itself (no E or I completing its magnetic path), gives a mu_eff around 10 I think, so this gives your worst-case A_L (at k = 0). That's with the coil in the middle of the core piece's winding window, on average; it should be somewhat higher if piled closer to the root, and lower close to the open face.
We can also add magnetic shunts, if we want lower k while maintaining high Lp and Ls. The shunt width and height should match the core leg's dimensions (to avoid saturation / core heating), and its length (which is the dimension spanning across the winding window) is adjusted to set the air gap for the LL as desired. In the extreme case, the shunt gap is zero and a complete (figure-8) path is made around each coil; now it almost doesn't matter the gap between cores, as this is identical to the case of two complete EI cores butted together. Which is obviously going to be pretty low k. So, between adding shunts (if needed), and adjusting the gap between E's, we can design Lm and LL (or k) as needed.
The effect of shunts is to make the winding window air gap shorter and more easily controlled (the more compact air gap has less fringing).
As for quantities -- typical gaps will be on the order of 10-30% of core thickness, I think? I don't think there's going to be any easy, fast rule to compute this, and your best luck will be with the winding-window-as-air-gap, and use of gapped magnetic shunts. These can be approximated as rectangular volumes of air gap, and the result should be in the right ballpark, probably uh, better than 50% error, maybe not much better than 20% unless you get lucky? The general case is tough, because fringing fields are everywhere, with overlapping air gaps with fields going diagonally across them, for which rectangular volumes will not give good results. FEA will give excellent results, but of course you need a full 3D magnetostatic simulator for that.
The other way to construct these from standard parts, I think is to shim apart two E's with ferrite blocks in the outer limbs, leaving an exaggerated air gap in the center. Use half-bobbins on each E (I guess you could cut down a pair of split bobbins, or use tall "EI" style E's, with two normal bobbins back to back, as a gapped EE assembly), or of course whatever hardware is appropriate (the gapped split bobbins made for this type of transformer). It's probably easier to make and measure an assembly, than to set up a model and simulate it...
Have fun!
Tim