numr 0.11.0

GPU scatter_reduce is deterministic and bit-equal to the CPU, and the CUDA client works from any thread.

Fixed

• CUDA and WebGPU scatter_reduce — float results no longer change from run to run.
  • Contributions are grouped by destination with a stable radix sort. Each destination then folds its run in source order, the order the CPU kernel uses.
  • F32 and F64 results on CUDA, and F32 results on WebGPU, are bit-equal to the CPU reference for every op: sum, prod, max, min and mean.
  • No float accumulates through an atomic. Before, an overlap-add (an inverse STFT) gave different low bits on each run.
  • WebGPU mean rounds exactly as the CPU does: f32(f64(acc) / count).
• scatter_reduce shapes — a source shorter than the destination on an axis other than dim gives the right result on CUDA and WebGPU, which assumed equal shapes. A source longer than the destination on such an axis returns an error on every backend. The CPU path wrote out of bounds.
• CUDA client on any thread — a CudaClient works from a thread that did not create it, with no bind by the caller.
  • Every direct driver call makes the client's context current on the calling thread first. This covers the allocator, the copies, events, the Sobol cache, CudaDevice::sync, CudaDevice::memory_info and the NCCL communicator.
  • Before, record_event, record_event_on_compute, sync and memory_info failed with CUDA_ERROR_INVALID_CONTEXT on a fresh thread. On a thread where another device's context was current, they acted on that device.
• CUDA scalar readbacks — histogram, logm, sqrtm, signm and the scan totals read on the compute stream and return copy errors.
  • Histogram, logm, sqrtm and signm read after the kernels that produce the value. Before, the read raced those kernels.
  • F32 inputs to logm, sqrtm and signm read their scalars as F32. Before, each read took eight bytes as one F64.
• CUDA copy and allocation errors — a device-to-host copy returns an error when the stream synchronize fails. An allocation error other than out-of-memory keeps its driver code instead of reading as OutOfMemory.

Changed

• GuardedStream::enqueue_permit returns Result<EnqueuePermit, DriverError>. It makes the stream's context current before it takes the capture lock.

• Integer scatter_reduce on CUDA and WebGPU — integers go through the same sorted pipeline. The old CUDA integer path scanned every source element per destination.

• WebGPU scatter_reduce limits — a call errors with BackendLimitation when the sort scratch exceeds the device's storage-binding limit, or the shape holds more than 8 axes after merging aligned ones.

Full diff: v0.10.0...v0.11.0