v0.6.7

[0.6.7] - 2026-08-23

Fixed

• OpenFold3: a YAML msa: pointing at your own alignment file crashed the fold with
  IndexError: list index out of range. The vendored parser keeps only files whose stem is
  one of its known MSA sources and dropped everything else, so there was nothing left to
  index. Your file is now exposed to the parser under the canonical name, bytes untouched.
  Present in every release that shipped the OpenFold3 msa: key, 0.6.6 included; the
  committed examples/prot_custom_msa.yaml was one of the inputs that crashed.

• OpenFold3: cyclic: true on a chain now raises instead of folding it as a linear chain and
  reporting success. The vendored tree carries neither Chain.cyclic nor the cyclic_mask
  feature it derives, so the flag reached nothing. Use --model rf3 or boltz2 for cyclic
  chains.

• OpenFold3: a CCD ligand's reference conformer was built without stereochemistry, so the
  generator picked a handedness per centre at random and that arbitrary choice became a model
  input. Chiral centres are now assigned from the CCD entry first, so the conformer keeps the
  handedness the code names. --model openfold3 is polymer-only today and refuses ligand
  chains, so no fold in this release reached it; the other models build their ligands on their
  own paths and were never affected.

• OpenFold3: an MSA deeper than its per-source cap is now truncated, the way the reference
  truncates it. The vendored parser dropped the truncated copy and returned the full
  alignment, so a deep alignment reached the featurizer whole and the model saw a different
  set of rows than the reference did. Nothing changes below the caps: all seven OpenFold3
  parity legs sit under them and reproduce their committed numbers.

• Protenix-v2 and OpenDDE fold more accurately. Two bugs in the pair trunk both models share
  are fixed: the mask marking which residue pairs are real reached only one of the two triangle
  multiplications, and OuterProductMean added its output bias without the scale the reference
  applies. Every Protenix-v2 and OpenDDE structure leg in the accuracy gate now lands inside the
  reference's own seed-to-seed spread; some fell outside it before. The other models on that
  trunk reproduce their published numbers unchanged. See docs/implementation-parity.md.

• RF3 folds are back to full speed. 0.6.6 turned on the accurate softmax for Protenix-v2 and
  OpenDDE, and it reached two extra sites inside RF3's pairformer that were never meant to get
  it: 512 aa went from 82.5 s to 111.8 s. The setting is scoped now and the structure is
  bit-identical to what 0.6.5 produced.

• RoseTTAFold3 folds crashed on a clean pip install. biotite was declared without an upper
  bound, so a fresh install resolved 1.7.1, which removed two internals the vendored AtomWorks
  featurizer uses; every --model rf3 fold died at import before reaching a card. The
  requirement is now biotite<1.7. If you already have biotite 1.7 in an environment, pip install -U tt-bio will downgrade it. Affects every release that shipped RoseTTAFold3.

• full_parity_gate.py --workers no longer ssh-es a host to itself. The fleet short names
  qb1 and qb2 are recognised as their own boxes, and any host that is genuinely remote is
  probed once before the first fold: reachable, not this same machine, and the card node
  present. A bad worker name fails preflight in seconds instead of turning every device leg
  into an instant error.

• The release gates refuse to run on a Python environment that does not satisfy tt-bio's own
  declared dependencies, naming what is missing or out of bounds. Before, a gate host missing
  one package reported the model that needed it as a failure instead.

Performance

• OpenFold3 folds 704 aa 1.34x faster (43.193 -> 32.230 s) and RoseTTAFold3 1.14x
  (45.332 -> 39.808 s). Both are bit-exact, so no prediction moves. The gain is at the sizes
  where the accurate softmax used to give up on splitting its work and run one unblocked pass;
  512, 576, 640, 768, 896 and 1024 aa already split and are unchanged. See
  docs/openfold3-port.md.