Follows #34

Breaking out deconv.py into:

• config.py: Holds a pydantic model for input params into deconv. It also freezes the data in the config
• utils.py: Holds small lightweight methods
• solver.py: implements the CVXPYSolver and OSQPSolver using base class DeconvSolver
• deconv.py: everything else

Testing and comparisions to old deconv.py

I ran some comparison sweeps across the old and new deconv code

solve

• solve looks to perfectly replicate the old code. The refactor didn’t change the core solver output.

difference in svals_len

• We are seeing a difference in some solve_thres behavior.
• The vast majority of solve_thres failures show the exact same debug signature:
• old_svals_len=999, new_svals_len=1
• That’s a bug in the old code I think: min_rel_scl filtering shrinks scals/bs but not svals/yfvals in the old code, so objective evaluation zips mismatched lists and effectively evaluates the “wrong” candidate(s).
• To verify, I implemented this bug in the new code and that results in identical results. This is left in currently if you enable legacy_bug in deconv.py.