v0.7.1

[0.7.1] - 2026-08-25

Fixed

• Protenix-v2, OpenDDE and OpenDDE-abag were padding the atom axis and not the token axis. At any
  token count that is not a multiple of 64, the ragged tail reached the triangle attention as real
  key columns rather than masked ones, and both the stock and the fused attention read them: on a
  probe, relative error against the aligned answer is 0.914 ragged against 0.038 padded. A
  98-residue fold presented 1208 ragged calls out of 1208 on Protenix-v2 and 1216 on OpenDDE, which
  carries a third axis in its structural-token refiner at roughly twice the residue count and so is
  essentially never aligned. All three axes now pad to a multiple of 64, mask, and slice back;
  TT_BIO_PROTENIX_TOKEN_BUCKET=0 restores the old path for an A/B and
  TT_BIO_PROTENIX_TOKEN_PAD_MULTIPLE overrides the multiple.

  Where the pad is 0 the fold is byte-identical, so a token count that is already a multiple of 64
  costs nothing and changes no number: at 512 residues both arms return CIF 5e404779d791fa8f on
  Protenix-v2 and agree to -0.101 s over eight interleaved pairs on OpenDDE. Where the pad is not 0
  you pay for the columns you added, and at real fold lengths that is a win rather than a cost:
  298 residues round to 320 and run 4.8 % faster on Protenix-v2 and 6.0 % faster on OpenDDE. Short
  folds are the exception and they are worse, because the fold is dispatch-bound before the pad
  arrives: 20 residues round to 64 and lose about 17 % of throughput. That is the smoke input
  perf_regression.py times, so its baseline rows were re-recorded deliberately rather than
  waived. docs/size-generality.md has the size-by-size reading.

• A job that never got the card now says so instead of looking like a bad result. A co-tenant
  holding a device used to surface as RuntimeError: every local worker exited, and inside a gate
  as an accuracy verdict against a fold that never ran. Contention now has its own reserved exit
  code, 75, carried from the worker through the predict fan-out to the CLI.

• opendde-trpcage-nomsa reports a real verdict in full_parity_gate.py again instead of
  BLOCKED-REGEN. The harvest step wrote the regenerated fixture under the source id rather than the
  destination id, so the leg never found what it had just produced.

• The benchmarks page. Rows that are hidden no longer reach the charts and tables, a stray script
  fragment no longer prints as page text, and design rows state the batch size they were measured
  at. A row is only published once every processor column in it has been measured: partial rows
  wait in perf/page_rows_pending.json and scripts/site_publish_guard.py enforces it.

• An explicit $RF3_CKPT is no longer outranked by a stale checkpoint in the legacy cache
  directory. The RF3 accuracy cell fell back to ~/.cache/tt-bio/rf3/ whenever the path it was
  given did not exist, so on a host that still had an old copy there it silently folded that copy
  instead of failing on the path the caller named.

Added

• RFD3 block-sparse atom attention, opt-in with RFD3_BLOCK_SPARSE=1. The atom site's
  128-neighbour index is block-sparse rather than row-sparse, so a block of neighbouring query rows
  shares a key window narrow enough to run as a batched dense matmul. Worth 2.982 s/design on a
  target it fits (6051 atoms, median of 7 rounds, range 2.218-3.988). It stays off by default
  because it is target-specific, not because it is unproven: the query block has to divide the
  tile-padded atom axis, which at the default block size holds for 288 of the 11745 atom counts
  between 256 and 12000. Every other target takes the dense fallback, which is always available and
  caps the downside at the dense chain. Not bit-exact against the dense path, and the only
  difference is the softmax reduction order.

• RF3 has an accuracy floor at 997 aa, not just at 117. release_gate.py --model rf3-1024aa folds
  7EIP on the device and gates CA-RMSD against the deposited crystal at 4.0 A; full_parity_gate.py
  scores the same leg. Measured 1.9687 A, against a reference of 2.0092 A.

Gates

Cut from a tree that passed, on one Blackhole p300c host (tt-quietbox2, Python 3.12.3, ttnn 0.68.0):

• Packaging: all 61 expected data files and all 43 declared runtime dependencies ship in both the
  wheel and the sdist and land on disk after a clean install.
• Parity (full_parity_gate.py, 41 legs, 201 min): 34 PASS, 4 GAP that reproduce their committed
  values, 1 PASS-caveated, 1 fixture awaiting regeneration, 1 FAIL. No leg drifted, errored or was
  skipped. The FAIL is af2ig-trunk-device, whose floor was recorded on a 13x10 board with the
  template on host and is read here on an 11x10 board with it on card; AF2-IG has no CLI path in
  this release.
• Accuracy floors (release_gate.py): every arm passed, including Boltz-2, ESMFold2, Protenix-v2,
  OpenDDE, OpenFold3, RF3, OpenBind, BoltzGen, RFD3, Nesso-1, ESMC and the OpenDDE antibody-antigen
  DockQ floor. PXDesign fit-RMSD 4.909 A against a 15.0 A floor.
• Throughput (perf_regression.py, +-15%): 14 of 14 models within tolerance on their recorded
  p300c baselines. The three models whose numerics this release changed clear their re-recorded
  post-fix baselines: Protenix-v2 +13.9%, OpenDDE -0.3%, OpenDDE-abag +0.6%.
• CLI (ux_regression.py): every surface cleared progress, parse and results/manifest shape.

Not covered, and worth naming rather than dropping: the size ladder has no baseline for this card
type, so it did not run; RF3, Nesso-1 and single-sequence ESMC-300M have no throughput baseline on
this card type; RFD3 block-sparse is dark on the gate's own fixture, so its coverage is its own
tests rather than a gate leg; and tt-bio design --model pxdesign is gated through the library
rather than end to end through the CLI.