[user story] debugging predictions by inspecting simulations

Following internal discussion, I'm posting a quick user story here. cc @dotsdl @dwhswenson

When FE predictions are poor versus experiment, many variables can be the cause, even additively. One of the first (good practice would be: always the first) thing to check is whether ligands are retaining their binding modes along the performed simulations. Ligands being ejected out of binding pockets is not uncommon even at these short timescales - obviously this messes up FE predictions.

Currently, at least on alchemiscale, there is no way of retrieving the simulation from a computed edge. Downloading a whole trajectory (or set of replica trajectories) can be costly and may not be fully necessary (although it would make for some pretty viz). A more scaleable approach is pulling down a PDB file of the last frame in the trajectory. Binding modes can then be easily inspected per simulation. In terms of the repex protocol, pulling down this PDB file for each replica would be very valuable, as the binding mode transition can be monitored along the edge.

I'm not sure to what extent this is currently possible in OpenFE - perhaps @dotsdl can shed some light on what features would be required to get this functionality be enabled for hook-up to alchemiscale?

Start by reviewing OpenFE's existing simulation and trajectory retrieval interfaces, then examine how computed edges and replica outputs could be exposed for alchemiscale. Clarify the required integration points and whether the final frame is available per simulation. Done means defining a supported way to retrieve a last-frame PDB for each replica without downloading full trajectories.