Control v2 Cleanup

[mattlevine22]

Summary

This PR hardens and cleans up the closed-loop control implementation introduced in control_v2. It clarifies simulator contracts and current limitations, fixes online-filter initialization and randomness handling, improves typing and documentation, and adds focused coverage without changing the MPPI algorithm.

(heavy use of codex, but reviewed and iterated most things and culled a lot of the big changes it made to code + docs)

Key changes relative to control_v2

• Requires controlled simulation to use filter_source="cuthbert" and raises a clear error for unsupported backends.
• Gives the initial online filter update a valid preceding time, preventing degenerate zero-duration transitions and possible NaNs in time-scaled dynamics.
• Separates environment and filter PRNG streams and honors filter_config.crn_seed.
• Makes initial_policy_state an explicit Simulator.simulate argument.
• Rejects ctrl_times and ctrl_values when controls are computed online.
• Validates that a policy returns one control vector with the model’s declared control dimension.
• Explicitly rejects plated controlled simulation; n_simulations > 1 also remains unsupported.
• Improves type annotations, variable names, and docstrings around the controlled simulator and one-step Cuthbert filtering.
• Moves the extensive control-loop explanation into dedicated API documentation and adds the control tutorials and API page to the documentation navigation.
• Corrects control tutorial imports, equations, terminology, and typographical errors.
• Keeps MPPI’s implementation and dsx.simulate rollout semantics unchanged, while correcting inaccurate documentation about its formulation, differentiable fields, and dt default.
• Removes obsolete commented-out tests and adds focused coverage for handler routing, policy output shape, and unsupported plated execution.

[MatthieuDarcy]

Regarding this, I added a draft PR #314 to illustrate some of the issues and limitations with using the CD-dynamax backend. Not a serious proposal, more of an illustration of some of current limitations

[mattlevine22]

thanks! Another way to deal with that is to update CD-Dynamax to expose lower-level predict/update steps.

And then integrate them into a unifying mid-level Dynestyx Filter object (#295 )

[MatthieuDarcy]

I think that's the correct approach, it probably would be useful for other projects as well. I can take a look at it soonish.

[mattlevine22]

I wouldn't prioritize the CD-Dynamax side too heavily, because once #299 lands (better discretizers), we will be able to use those w/ Cuthbert filters to get most of the CD-Dynamax features.

But indeed, setting up #295 will be generally very useful for many other projects---I think a great direction to get us up to cutting-edge would be to implement/learn "enhanced filters" this way.

[MatthieuDarcy]

Could it be worth it to add the basic tutorial into the gentle introduction?

[mattlevine22]

Let's make that a separate PR so we can have a chance to look more carefully at the "pedagogy" and make sure it is as simple and clear as possible?

[MatthieuDarcy]

Sounds good! I think it would be very nice to have a basic example of online control in the intro tutorial

[mattlevine22]

Opened issue #316 for this.

[MatthieuDarcy]

Seems good, I focused on understanding substantial changes to the code, but the documentation is much improved, thanks Matt!

[DanWaxman]

looks mostly there to me, a few small nits

[DanWaxman]

sorry what is bootstrapping here?

[mattlevine22]

I think it is just how to start with prev_state=None; this was @MatthieuDarcy 's wording, perhaps through Claude haha. I had codex re-word this stuff...

@MatthieuDarcy could you double check its explanation? my brain is a bit split