v0.4.0 — DF-MP2 end-to-end validation + NKI energy kernel

Highlights

• Real-molecule DF-MP2 validation against PySCF (#11) — trnblas matches PySCF's own mp.dfmp2.DFMP2 reference to nanohartree precision on H2O/STO-3G, H2O/cc-pvdz, CH4/cc-pvdz, NH3/cc-pvdz. New pip install trnblas[pyscf] extra, runnable examples/df_mp2_pyscf.py demo.

• Fused MP2 energy-reduction NKI kernel (#15, Phase 1) — trnblas.nki.nki_mp2_energy with partition-dim sub-tiling. Validated on trn1 across nvir ∈ {8, 16, 64, 256, 448}. Scaffold landed; further perf work tracked under #15.

• DF-MP2 step-4 collapse (#14) — energy reduction replaced from nocc² sequential batched dispatches with one chunked GEMM via the algebraic identity T_full = X @ X.T. On trn1.2xlarge:

  Shape	Flops	Cold	Warm	TFLOPS
  small (128/16/384)	3.4 G	0.025s	0.008s	0.43
  medium (512/64/1536)	2757 G	12.9s	9.77s	0.28
  large (768/96/2304)	20352 G	65.9s	62.8s	0.32

• Trainium CI infrastructure — Terraform module for a persistent trn1 test instance, SSM-driven runners (scripts/run_neuron_tests.sh, scripts/run_df_mp2_bench.sh), docs at docs/aws_setup.md.

• NKI GEMM kernel wired to real nisa.nc_matmul with stationary tile reuse + HBM padding for arbitrary shapes. nki_batched_gemm dispatches per-slice through the cached kernel. 17/17 hardware tests pass.

• Repository transfer — now at trnsci/trnblas. Docs at https://trnsci.dev/trnblas/.

• neuronxcc floor bumped >=2.15 → >=2.24 (NKI 2.24+ nc_matmul calling convention) to unify with the rest of the trnsci suite.

See the full CHANGELOG for details.

---

⚠️ Erratum (v0.4.3): The "GEMM per-call kernel timing" and "DF-MP2 end-to-end" tables here reported trn1 numbers that were silently torch.matmul fallback on trn1's Xeon, not NKI on the Tensor Engine. Fixed in v0.4.3.