Fix ADP restraint config being silently lost on target rebuild

[kmdalton]

ADPSimilarityTarget.simu_sigma is public API with a deliberate setter, but there was no way to make a value survive. Refinement._init_targets builds TotalADPTarget(self.model, verbose=self.verbose) passing no restraint parameters, so every rebuild resets simu_sigma / simu_sigma_aniso to the constructor defaults (2.0 / 1.0). refine_rigid_body rebuilds once per resolution cutoff via _rebind_for_data -> _init_targets; the ensemble and create_from_state_dict paths rebuild too.

The result: set a sigma, run rigid body, and refinement silently proceeds at the default. No warning. There was no supported alternative -- no constructor argument, no CLI flag (--sigma-a-max is sigma_A, a different quantity) -- so post-construction assignment was the only way in, and it was exactly what got discarded.

This is the failure mode already documented on _xray_target_kwargs: "a second build site silently reverts whatever it forgets to pass, which once made five CLI flags no-ops." The x-ray targets were given a single source of truth for their construction kwargs; the ADP targets never were. This applies the same pattern.

• CombinedModelTargets takes component_config, {component: {kwarg: value}}, set before _create_targets() and exposed to subclasses via _component_kwargs(). A component name that matches nothing raises rather than no-op'ing, since a silent no-op is the bug being fixed. Config is deep copied so a later mutation of the caller's dict cannot reach the target.
• TotalADPTarget._create_targets splats the per-component kwargs.
• Refinement takes adp_restraints=..., stores it alongside the other pre-_init_targets configuration, and passes it on every rebuild with the same getattr fallback _xray_target_kwargs uses for the ensemble and state-dict paths.

Behaviour is unchanged when no config is passed.

LBFGSRefinement(..., adp_restraints={"simu": {"simu_sigma": 0.4}})

Verified against the reported scenario: post-construction assignment reads back as 2.0 after refine_rigid_body, constructor config holds at 0.4 through rigid body, a second get_scales, and refine_adp.

TotalGeometryTarget has the same latent issue -- its components are built with no configuration path either -- but nothing sets geometry component parameters today, so it is left alone. The base-class mechanism is generic, so wiring it up later is two lines. A CLI flag for --adp-restraints would make this reachable from torchref.refine; deliberately not bundled here.

Tests: tests/unit/test_adp_restraint_config.py, 7 cases covering defaults, propagation, survival across a rebuild, copy-not-alias, and both misspelling paths. Full unit + functional suite passes (1748 passed, 74 skipped), as do the 32 integration tests touching rigid body, ensemble, state-dict and CLI paths.

[kmdalton]

hi @HatPdotS , i ran into an issue with the ADP target weights being reset to a default value between macrocycles. Claude proposes this fix. let me know if it sounds right to you.

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

❌ Patch coverage is 76.92308% with 3 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
torchref/refinement/targets/combined.py	62.50%	3 Missing ⚠️

📢 Thoughts on this report? Let us know!

[kmdalton]

one more detail -- i was using a custom refinement loop which runs rigid body during every macrocycle. if i understand correctly, that was the root cause. i can dig in more if needed. just wanted a second set of eyes on it.

[HatPdotS]

I moved the weights into the refinement loss_state object, and I think these should survive the rigid-body step.
I have been moving away from the combined targets architecture because it quickly becomes very confusing. Can you check if custom loss state weights are lost after the rigid step?

[kmdalton]

looking into it...