v0.2.4

[0.2.4] - 2026-07-10

Device-resident trunk for tt-bio gen (BoltzGen) — no structure-model code changed for
Boltz-2/ESMFold2/Protenix-v2 (the new TokenDistanceRecycle/TrunkModule params default to
off/None, purely additive).

Release gate (scripts/release_gate.py, examples/prot.yaml, 200 steps / 5 samples, seed 0):

model	CA-RMSD	TM	floor	result
Boltz-2	1.43 Å	0.944	≤3.0 Å / ≥0.75	PASS
ESMFold2	2.76 Å	0.798	≤4.0 Å / ≥0.65	PASS
ESMFold2-fast	1.74 Å	0.907	≤4.5 Å / ≥0.60	PASS
Protenix-v2	3.87 Å	0.706	≤6.0 Å / ≥0.50	PASS

No regression vs 0.2.3 (within TT diffusion's seed-to-seed variance band).

BoltzGen designability — n=8 fixed-length-100 designs, examples/binder.yaml: scRMSD
median 0.84 Å (resident) vs 0.91 Å (host), 7/8 designs ≤2 Å strict pass (comparable to host's
8/8) — no regression. Wall-clock (design + refold + confidence + analysis + filtering) 697 s
→ 479 s, ~31% faster. See docs/boltzgen-resident-trunk.md.

Added

• BoltzGen device-resident trunk — TokenDistanceRecycle (mirrors TemplateRecycle) keeps
  the per-iteration token-distance injection fully on-device, collapsing 4 host↔device
  crossings/iteration to 2 (only the template sub-module still round-trips). Boltz.__init__
  takes use_resident_trunk: bool = True; set false to fall back to the original host path.

Changed

• Promoted Protenix-v2's diffusion denoiser-unit and AttentionPairBias(has_s=True) ad-hoc
  checks to proper pytest cases (test-coverage only, no functional change).