fix motif drift for symmetric fused-motif designs (#324)

Issue
When scaffolding a fused/contiguous motif under symmetry (inference_sampler.kind=symmetry), the motif drifts over the trajectory so the output no longer matches the input motif. Symmetric inference applies a per-step COM recenter, then rebuilds the symmetric copies from the ASU. A fused motif is part of the generated chain, so it isn't marked FIXED_ENTITY_ID and nothing corrects for the shift. The recenter, propagated through the symmetry transforms, moves the motif off its input pose and the error accumulates over the trajectory.

Fix
In SampleDiffusionWithSymmetry.apply_symmetry_to_X_L, we skip the COM recenter when a contiguous motif is present, so it stays put, and the symmetric copies rebuild from it correctly.

Scope
Scoped to is_motif_atom_with_fixed_coord & is_sym_asu & ~FIXED_ENTITY_ID and guarded by not self.allow_realignment, so it's a no-op for de novo symmetric, unindexed motifs and the non-symmetric sampler.

Author: TanushGoel

models/rfd3/src/rfd3/model/inference_sampler.py

@@ -6,6 +6,7 @@
 
 import torch
 from jaxtyping import Float
+from rfd3.inference.symmetry.atom_array import FIXED_ENTITY_ID
 from rfd3.inference.symmetry.symmetry_utils import apply_symmetry_to_xyz_atomwise
 from rfd3.model.cfg_utils import strip_X
 
@@ -381,9 +382,18 @@ def apply_symmetry_to_X_L(self, X_L, f):
         # update symmetric frames to correct for change in global frame
         symmetry_feats = {k: v for k, v in f.items() if "sym" in k}
 
+        # a contiguous motif anchors the frame, so skip the COM recenter (avoids drift)
+        fixed = f.get("is_motif_atom_with_fixed_coord")
+        asu, ent = symmetry_feats["is_sym_asu"], symmetry_feats["sym_entity_id"]
+        held_motif = (
+            not self.allow_realignment
+            and fixed is not None
+            and bool((fixed & asu & (ent != FIXED_ENTITY_ID)).any())
+        )
+
         # apply symmetry frame shift to X_L
         X_L = apply_symmetry_to_xyz_atomwise(
-            X_L, symmetry_feats, partial_diffusion=("partial_t" in f)
+            X_L, symmetry_feats, partial_diffusion=("partial_t" in f) or held_motif
         )
 
         return X_L