keryx-miner-supr v0.13.2 — AMD OPoI inference fix

keryx-miner-supr v0.13.2

Fixes AMD OPoI GPU inference. On AMD/OpenCL rigs the OPoI self-test could fail with
no trusted ggml device matches the selected OpenCL worker PCI allowlist or
invalid value for main_gpu, leaving the miner stuck at "no models ready — mining
suspended". Two fixes:

• Device-index translation. llama.cpp (b10015) builds its GPU list from discrete
  GPUs only — integrated GPUs and CPU backends are filtered out. The Vulkan inference
  route now maps the selected card's ggml device index to that discrete-GPU space, so
  rigs with an integrated GPU present no longer overflow main_gpu and fail to load the
  model.
• PCI-allowlist fallback. When the mining GPUs expose no verifiable PCI identity
  (the OpenCL platform reports none, or the Vulkan driver lacks VK_EXT_pci_bus_info),
  automatic inference placement no longer fails closed — it falls back to the largest
  discrete GPU (pre-0.13.1 behaviour). Strict matching is kept whenever PCI is
  readable. KERYX_LLAMA_VK_DEVICE=<n> still pins a specific card explicitly.

Validated end-to-end on a multi-GPU AMD rig (integrated GPU + discrete cards): model
loads, OPoI tier is serveable, shares accepted.

NVIDIA, macOS and Windows builds are functionally identical to v0.13.1 (they do not use
the AMD Vulkan inference route); they are rebuilt at v0.13.2 for a consistent release.
Everything from v0.13.1 carries forward: the Matrix TUI dashboard + runtime telemetry
(--no-tui/--force-tui), JSON stats API, and all v0.13.0 correctness/autotune work.

Assets

NVIDIA modern (RTX 20xx–50xx) and legacy (incl. GTX 10-series / Pascal) lines, each as
HiveOS, mmpOS and SMOS packages plus a plain linux-x86_64 tarball; Windows (NVIDIA PoM);
macOS arm64; and the AMD/OpenCL line. Docker: ocminersupr/keryx-miner-supr:0.13.2 and
:latest. Checksums per platform.