Module mcp 

Phase 4: AOT GGUF → P64 LLM-weight container compiler. Phase 6 / task #12: safetensor (+ MLX) source parsing + dtype gate for the streaming transcoder. Task #12 / STELLAR §A: BitNet b1.58 ternary quantization codec (compression during transcode). Task #12 / STELLAR §A: tensor-name → engine GEMM-role mapping + the ternary (FFN-only) policy. Task #12 / STELLAR §A: native GPU dispatch of the ternary GEMM kernel + on-device parity test. STELLAR §A A1a: GPU top-K reduction — CPU oracle + host merge + the WGSL kernel. STELLAR §A A1a: native GPU dispatch of the top-K reduction + on-device parity test. STELLAR §A AH-track H0: host topology + capability sensor (enumerate all adapters; discrete vs unified). STELLAR §A AH-track H1(a): cross-circuit GEMV benchmark → measured capability matrix (CPU/iGPU/GPU). STELLAR §A AH-track H2: residency + device-priority planner (discovery → employment plan, D31). STELLAR §A AH-track H1(a) cache: CBOR hardware passport (cache discovery, fast-boot skip, D26). STELLAR §A A0 (D17/D22): shared native LLM benchmark harness — the one measurement surface for the existing F16/Q8 path and the future ternary/top-k paths. STELLAR §A W2 (D17): per-kernel GPU timestamp profiler for the LLM forward/decode path. STELLAR §A W3: in-project GPU↔CPU kernel-parity oracle (error metrics + synthetic quant weights). STELLAR §A W1: in-project quality oracle (perplexity / KL / coherence + the quant quality gate). STELLAR §A AWQ: activation-statistics capture (the AWQ forward hook) for calibrated quantization. Model-Context-Protocol server surface (moved from the crate root per MODULE_REORG_PLAN.md; crate::mcp_server / crate::mcp_cooperation paths are preserved by re-exports in lib.rs). mcp_server owns the *_impls submodules.

Modules

mcp_cooperation

MCP agent cooperation (Track M — task #17/#16/#18).

mcp_server