v0.5.0

[0.5.0] - 2026-07-27

Multi-sample folds are about 3x faster on Protenix-v2 and OpenDDE, which now draw every sample
from one batched device trajectory instead of looping one sample at a time. The parity reference
fixtures a clean checkout needs are published for the first time, so
scripts/fetch_parity_fixtures.sh works and the eight structure legs 0.4.0 had to skip are back
in the gate. RFdiffusion3 gains the parity, performance, and UX gate coverage it shipped without,
and a few percent of throughput.

Release gate (Blackhole P150a on pc, tt-bio 0.5.0): host suite 114 passed / 49 skipped;
packaging guard 15/15 data files and 31/31 declared dependencies present in the wheel and sdist;
UX gate PASS on every shipped surface, now including tt-bio design.

Parity gate (scripts/full_parity_gate.py, 22 legs): 20 PASS, 2 GAP, both GAPs
reproducing a committed GAP-evidenced verdict rather than drifting. The eight envelope structure
legs score for the first time since they were externalized, because the fixture asset they read is
finally published: Boltz-2 trp-cage 0.074 Å against a 0.145 Å envelope (ratio 0.51), Boltz-2 HSA
0.92 vs 1.33 (0.69), Protenix-v2 7ROA 0.049 vs 0.042 (1.16), Protenix-v2 HSA 0.050 vs 0.052 (0.96).
ESMFold2 (4 targets), ESMC-300m/600m, SaProt-35m/650m, OpenDDE-abag, BoltzGen (scRMSD pass-rate
100%, median 0.68 Å) and the RFD3 featurizer (43/43 keys bit-exact) all reproduce their recorded
verdicts, as do five of six Boltz-2 affinity legs; the sixth improves on its recorded gap.

Three of those legs pass on the absolute floor rather than a measured envelope, because their
cached bf16 and fp32 references are identical, which collapses the envelope to zero: Protenix-v2
ubiquitin (0.037 Å against the 0.05 Å floor) and both OpenDDE structure legs. The envelope test
assumes a model with a torch CPU path to recompute in two dtypes, which a ttnn-only port does not
have. Those OpenDDE legs are covered by the R/D/X device-vs-reference diagnostic instead, the right
scorer for such a port, and opendde-prot-prod passes it (X 4.824 against a 1.499 floor).
Restoring a meaningful envelope for ttnn-only legs is follow-up work, not a coverage hole.

The two GAPs are Boltz-2 7ROA no-MSA structure (kabsch_rmsd ratio 2.04) and FKBP12+SB3 no-MSA
affinity. Both are recorded GAP-evidenced in docs/implementation-parity.md. The 7ROA one was
root-caused this release: the pinned envelope seed lands that target in an unusually chaotic
reverse-diffusion trajectory in the reference, whose own bf16-vs-fp32 spread swings 3.45 Å, 2.16 Å
and 0.81 Å across seeds 0, 1 and 2, and the leg passes cleanly at seeds 1 and 2. Two on-device fp32
levers were tried and neither moves it, so it is a property of the reference trajectory, not a port
defect.

Perf gate (scripts/perf_regression.py, trpcage 20 aa single-sequence, warm 2+5, ±15%):
boltz2 1.164 structures/s (-2.2%), esmfold2 1.586 (-7.0%), esmfold2-fast 2.167 (-5.4%),
protenix-v2 2.192 (-8.0%), opendde 1.891 (-1.6%), esmc-300m 25.65 seq/s (+53.3%), esmc-600m 20.99
(+0.3%), esmc-6b 4.429 (+39.7%), saprot-650m 239.2 (+7.4%), boltzgen 0.01708 designs/s (-0.8%),
rfd3 0.1178 designs/s (-3.6%), boltz2-affinity 0.00909 affinities/s (-4.2%, median of three
isolated runs). opendde-abag measures 1.789 structures/s and is seeding its first baseline. No
model regressed beyond the threshold and nothing OOMed.

Added

• Per-sample chain-pair ipTM and per-chain pTM for OpenDDE — pair_chains_iptm and
  chains_ptm are now written for every sample in all_runs, like every other confidence
  scalar, instead of for the top-ranked sample only. For antibody-antigen work the
  antibody-vs-antigen chain-pair ipTM ranks candidates better than the global ipTM, and a
  winner-only value cannot rank the rest.
• RFdiffusion3 gate coverage — tt-bio design now has legs in the parity gate, the
  performance regression gate, and the UX gate. It shipped in 0.4.0 with none.

Changed

• Protenix-v2 / OpenDDE diffusion multiplicity batching — Protenix.fold / OpenDDE.fold
  now draw n_sample samples from one batched device denoise trajectory instead of looping one
  sample at a time (mirrors boltz2.AtomDiffusion.sample's multiplicity +
  max_parallel_samples pattern; --max_parallel_samples chunks batches too large to fit).
  Parity-verified against the established diffusion noise floor (batched-vs-unbatched drift
  within the seed-to-seed floor, not bit-exact, since diffusion is stochastic) on two hosts:
  Protenix X/floor 0.971-1.049, OpenDDE X/floor 0.995-1.003. Measured speedup at multiplicity 4:
  Protenix 3.19-3.57x, OpenDDE 3.15x.
• RFdiffusion3 is a few percent faster — the design matmuls that read a single tile of K now
  carry an explicit core-grid hint. It is bit-identical to the default and worth 5-15% depending
  on size. On one Blackhole p150a at 40 residues, batch 8 goes from 0.1216 to 0.1352 designs/sec
  and batch 1 from 0.0767 to 0.0807. The refreshed per-size table is in docs/rfd3-design.md.

Fixed

• OpenDDE re-ran a full offline MSA search on every multi-chain fold, and overwrote the MSAs
  the other models read. The paired-MSA helper had no cache check, unlike the unpaired path, so
  each fold searched the whole database again. Worse, it wrote {seq_hash}.a3m into the shared
  msa_dir, silently replacing the files Boltz-2 and Protenix-v2 read for the same chains.
  Paired results now live in msa_dir/paired/ and are only computed when absent, so a paired run
  can neither clobber the unpaired cache nor be served an unpaired file as if it were paired.
• The parity reference fixtures were documented but never published.
  docs/implementation-parity.md and scripts/fetch_parity_fixtures.sh both pointed at a
  parity-fixtures-latest release asset that did not exist, so on a clean checkout the eight
  envelope structure legs had no references and reported BLOCKED-REF-REGEN-NEEDED instead of a
  verdict. The asset is published and the fetch is verified end to end from a clean directory.
  Two fetch bugs went with it: a doubled parity-fixtures- prefix in the asset name, and a
  grep -F that treated the $ end-anchor as a literal character.
• Regenerating envelope references destroyed the fixtures' other provenance. The envelope
  path and the legacy R/D/X path share one meta.json with incompatible schemas, and the
  envelope regenerator overwrote the whole file, wiping the settings_tag and upstream-reference
  provenance the R/D/X scorer needs. That is what made five legs flip between blocked and passing
  every time either gate was regenerated. The regenerator now nests its own bookkeeping under an
  envelope key and leaves harvested fields alone.