HSLM backward on the GPU: dX and dW int8 tensor-core kernels written by a .t27 spec, measured against cuBLAS on an RTX 4090 #10847

HSLM training (specs/ml/train_hslm.t27, 2.13M ternary params) runs its backward pass on the CPU in f64: ternary_matmul_t_add (dX = dY x W^T) and outer_add (dW = X^T x dY). This issue gives that backward a GPU path through the t27c PTX backend of #10425: two int8 tensor-core kernels whose PTX is written by a .t27 spec, each with a CPU reference in the same spec, measured on an RTX 4090 against cuBLAS (the library torch.mm dispatches to). Wiring into the trainer is a follow-up; train_hslm.t27 is not touched here.

Boundary

• specs/igla/coder/hslm_bwd_ptx.t27
• bootstrap/gen/rust/igla/coder/hslm_bwd_ptx.rs
• .trinity/seals/coder_igla-coder-hslm-bwd-ptx.json
• bootstrap/src/main.rs
• bootstrap/src/host/driver.rs
• bootstrap/src/runtime_minimal.rs
• bootstrap/src/runtime_minimal_test.rs
• bootstrap/src/neural/attention.rs
• bootstrap/src/neural/hslm.rs
• docs/reports/orphan_modules.json

User Scenarios

• Given the spec, when t27c test-report runs it, then every test passes: the text pins of both kernels, the integer references on hand cases in the trainer's index layout, and the f64 error bounds of the quantized path.
• Given an RTX 4090, when t27c ptx-run --kernel 1 (dX) or --kernel 2 (dW) runs an HSLM shape, then it prints mismatches=0 against the t27 integer reference and the time per call.
• Given the same shapes, when cutlass_profiler runs cuBLAS SGEMM, bf16 and s8 GEMMs, then the PR reports both side by side.

Requirements

• FR-001: The dX kernel MUST compute C[m][j] = sum_i dYq[m][i] * w(j*K + i), the index layout of ternary_matmul_t_add, with W as 2-bit codes packed four per byte.
• FR-002: The dW kernel MUST compute C[i][j] = sum_r X[r][i] * dY[r][j], the index layout of outer_add summed over rows, with i32 accumulation and K <= 131071.
• FR-003: The spec MUST state the scales (dX: s_dy[row]; dW: s_x[i] * s_dy[j], per-column) and test an f64 error bound against an f64 reference.
• FR-004: The added foreign glue MUST be offset by deleted hand-written foreign lines (specs/policy/own_language.t27); the deleted files are orphans no mod reaches.
• FR-005: The report MUST be read against specs/t27t/t27t.t27 without flipping RUNNER_USES_INT_MM or F_MEASURED.

Success Criteria

• t27c test-report specs/igla/coder/hslm_bwd_ptx.t27 exits 0 on the lab.
• t27c ptx-run prints mismatches=0 for both kernels on 243x243, 243x729, 729x243 (padded to 256/736) at 27 and 1782 rows.
• The PR body has a table of ms per call and GFLOP/s for the kernels and for cuBLAS f32, bf16 and s8 on the same GPU.