BendEx

Local Bend language lab and Bend Doctor prototype workspace.

Contents

• Bend/ — upstream HigherOrderCO/Bend shallow clone used as reference source/docs/examples.
• workspace/ — local notes, experiments, results, and original tools.
• workspace/src/bend-doctor/ — Rust bend-doctor CLI.
• workspace/demos/room-of-workers/ — "Room Full of Workers": five verified puzzle demos + presenter script explaining Bend's automatic parallelism to non-technical audiences.

Bend Doctor quick run

bend-doctor
bend-doctor --verbose
bend-doctor --baseline --cuda-probe --timeout 60

From source:

cd workspace/src/bend-doctor
RUSTFLAGS='-C linker=cc' cargo test
RUSTFLAGS='-C linker=cc' cargo build --release
RUSTFLAGS='-C linker=cc' cargo install --path .

This WSL2 machine has an NVIDIA GeForce RTX 5090 (32 GiB, compute 12.0) with CUDA 13.1 working in WSL as of 2026-07-15 (run-cu/gen-cu pass smoke). Two machine-specific requirements, both wired into ~/.bashrc: /usr/local/cuda/bin on PATH, and NVCC_APPEND_FLAGS='-Xcompiler -U_GNU_SOURCE' (glibc 2.42's C23 rsqrt/pown declarations clash with CUDA 13.1 headers). hvm is the patched fork build (SiliconState/HVM2 rev a6c01d71, HVM_THREADS support) reinstalled with nvcc on PATH so the CUDA runtime is compiled in. Measured baseline (bend-doctor --baseline --timeout 60): run-cu passes tree depth 13, fails at depth 14 — the HVM2 fixed GNet limit, same boundary class as the old RTX 4060 numbers. Use run-c/gen-c as the default baseline and keep run-cu workloads at or below depth 13; re-run --baseline after HVM/CUDA/driver/WSL changes.