v0.2.1

[0.2.1] - 2026-07-09

Adds the ESMC embeddings capability merged since 0.2.0 (already hardware-gated at merge
time) and fixes packaging/docs metadata that was stale since 0.2.0. No model code changed
for existing capabilities — the 0.2.0 accuracy/perf/OOM gate still holds.

Added

• ESMC protein-language-model embeddings — tt-bio embed CLI + Python API
  (tt_bio.esmc.embed): per-residue and pooled embeddings from ESMC-300M/600M/6B, no
  folding head or MSA required. Parity vs reference ESMC: per-residue/pooled PCC
  0.9995-0.9999 across variants (normal and --fast).
• Automatic batching + length-bucketing for tt-bio embed on ESMC-300M/600M (~18.5x warm
  throughput vs unbatched); exact row-independence (masked batched output bit-identical to
  running each sequence alone), PCC 0.9996+.

Fixed

• pyproject.toml description was still "Boltz-2 implementation..." — now lists every
  shipped capability (Boltz-2, ESMFold2, Protenix-v2, BoltzGen, ESMC).
• pyproject.toml had no readme field, so the PyPI project page rendered with an empty
  long description — now points at README.md.
• README: pip install tt-bio (PyPI) is now the primary install path (the wheel has been
  on PyPI since 0.2.0); git/source moved to a secondary section. Intro paragraph now
  mentions ESMC embeddings. The dense Boltz-2/ESMFold2/Protenix-v2 feature-support
  paragraph is now a compact table.