[JMacnar]

Fixes #418

Problem

Symmetric designs fail in collate_examples with a dtype mismatch on
translation[row, target] = t, before the first diffusion step. Traceback and repro in
#418.

get_cyclic_frames and get_dihedral_frames build a frame's translation as
np.array([0, 0, 0]), which is int64. framecoords_to_RTs divides the rotation by a
norm, so R comes out float, but T = Ori passes the integer dtype through.
AddSymmetryFeats.make_transforms_dict carries it into feats["sym_transform"].
symmetry_maps then writes it by index into translation = torch.zeros(b, l, 3), which
is float, and an indexed write requires the dtypes to match.

batching.py arrived with #401 (2026-10-14), so this path is only reached since then.
test_rfd3_batch_symmetry.py does not cover it, and could not: it builds sym_transform
by hand because it needs a non-zero translation and a different frame per example to tell
a per-example mapping from a shared one. That leaves the dtype the generators produce
untested, which is what the new test pins.

Fix

Give the translation a float dtype where it is created, at the four literals in
frames.py:

np.array([0, 0, 0])   ->   np.zeros(3, dtype=np.float32)

This changes no value. The translation is exactly zero in every generator, and consumers
already add it to float coordinates: atom_array.py:161 does atom_array.coord += T,
with a comment that T should be 0. The rotations were always float, since they come from
np.cos and np.sin; only the translation was integer.

I changed the generators rather than casting in framecoords_to_RTs or at
batching.py:190, because that is where a translation in angstroms is being stored as an
integer, and a cast further down leaves the next reader of it to trip over the same thing.

Verification

• New test test_frame_translation_keeps_the_rotation_dtype in
  models/rfd3/tests/test_symmetry_frames.py, parametrized over C2, C3 and D2. It goes
  through get_symmetry_frames_from_symmetry_id and decompose_symmetry_frame, the
  path a real design takes. It fails on production for all three groups and passes
  with the fix; the file goes from 23 to 26 passing tests.
• End to end: rfd3 design with {"length": 200, "symmetry": {"id": "C2"}} on the
  default settings. On production it raises in symmetry_maps; with the fix it
  completes and writes a two-chain C2 assembly (200 residues per chain, 0.03 A RMSD
  between one chain and the other rotated 180 degrees about z).
• Every symmetry and batching test file in models/rfd3/tests passes: 98 tests, no
  failures. (test_symmetry.py errors during collection on missing json test data, both
  with and without this change.)
• In use since 9 November on a C2, C3, C4 and C6 design campaign, about 4000 structures
  over thirteen jobs. Geometry checked on one C6 assembly of six 500-residue chains: each
  chain superimposes on the first at one of the six multiples of 60 degrees, every angle
  used once, worst CA RMSD 0.086 A, assembly centred on the origin to within 0.02 A.

Disclosure

This change was prepared with AI assistance. The test and the end-to-end run above were
both executed.