qector-batch-decodinglisted

Batch, streaming, and GPU-accelerated decoding paths for QECTOR. Covers the manual 13 CUDA / OpenCL bit-identity contract, the per-thread scratch sizing formula (manual 13.4), the warm-up / dispatch / hot-path pattern, the worker-thread determinism guarantee (manual 26.3), and the cold-path / hot-path reporting rule. Load when a question involves millions of shots, GPU acceleration, batch decode, or performance tuning.
GuillaumeLessard/qector-claude-plugin · ★ 1 · AI & Automation · score 62

Install: claude install-skill GuillaumeLessard/qector-claude-plugin

# QECTOR Batch Decoding Source of authority: v1.0.0 reference manual, chapters 13, 17, 20, 26. ## The batch and streaming surfaces (manual 16.2 - Provisional) | Class | Notes | | --------------------------- | ---------------------------------------------------------- | | `CPUBatchDecoder` | batch shape stable; performance workload-sensitive | | `BatchDecoder` | batch shape stable; performance workload-sensitive | | `CUDABatchDecoder` | per-thread port of the UF core; bit-identity tested | | `OpenCLBatchDecoder` | source build only; bit-identity tested | | `CUDABpOsdDecoder` | batched BP+OSD; single-shot CPU preferred | | `StreamingDecoder` | OR-accumulated history; simulation workflow | | `SlidingWindowDecoder` | exponentially decayed window; bounded tail | The library MCP server does **not** expose any batch tool (manual 8.0: the 8-tool library surface is frozen). The bench server adds `qector-research.hot_path_microbench` for per-shot latency distribution sampling, but it caps at `QECTOR_MCP_BENCH_MAX_BENCH_SHOTS` (default 5000) and the result is per-machine, per-workload, per-build only (manual 22.5). ## The bit-identity contract (manual 13.2, Theorem 16) > For any graphlike code and any syndrome `s`, the unweighted batch > kernels produce `c_