v0.6.4

[0.6.4] - 2026-08-19

P300 Blackhole cards fold again. On a 110-core grid (every P300), a mid-size
protein+ligand target died in the triangle multiplication with "statically allocated
circular buffers clash with L1 buffers": the chunk-width budget behind that op was
measured on a 130-core card and admits widths that do not fit on a tighter one. The
trimul now catches the clash, which throws before anything runs, and retries one chunk
width narrower. Narrowing is bit-exact (the width only partitions an independent-channel
sum), and the failing width is remembered per shape, so a process pays one failed
compile and every later call starts narrow. Reported by Taylor Singletary in #11; his
grid sweep is what pinpointed the threshold.

Added

• scripts/release_gate.py --model l1-budget, in the default arm set: a release leg for
  the class of defect #11 belonged to. It runs the trimul chunk-width budget for every
  part class in L1_BUDGET_PARTS and folds #11's own target across the grid ladder the
  running card can express, so a budget fitted on one card cannot ship unchecked on a
  card with fewer cores. docs/part-l1-budgets.md carries the measured per-part figures.
• OpenFold3 --single_sequence folds upstream's no-MSA mode through a one-row
  alignment, and upstream's own OpenFold3 inference suite now runs against this port
  with committed verdicts (docs/openfold3-upstream-suite.md).
• The release gate checks RELEASE_GATE_MSA_DIR before it opens a device and names the
  a3m files to seed. A dir that covered one target used to fail an unrelated-looking
  accuracy arm an hour into the run.

Fixed

• Mid-size targets no longer die with an L1 circular-buffer clash on 110-core
  Blackhole grids (P300/P300C). The triangle multiplication's channel chunk narrows to
  the widest width that fits, falling back to DRAM residency at the floor, and outputs
  are bit-identical to grids that never clashed. (#11)
• A caught trimul clash now says so on stderr. tt-metal logs the clash at critical
  before raising, which reads like a fatal error even though the retry succeeds. (#11)
• TT_VISIBLE_DEVICES accepts PCI bus addresses (0000:01:00.0), the form ttnn's
  device open takes, resolving them to device indices; an unknown entry fails with a
  message naming the index form. (#11)
• Opening a whole P300 board pair no longer fails with "Physical chip id 0 not found
  in control plane chip mapping". The 1x1 mesh-graph descriptor a lone P300 chip needs
  is now applied only when exactly one chip is visible. (#11)
• ttnn-only models refuse --accelerator cpu/gpu on every path, and the CPU/GPU path
  no longer requires the ttnn wheel. (#10)
• tt-bio predict exits nonzero when a run loses targets instead of reporting success
  on an empty result set.