Shared runtime manifest v1 (owner: Suite Lead, 2026-10-11)

Single schema for %LOCALAPPDATA%\Voxprint\shared\manifest.json (Linux ~/.local/share/voxprint/shared/manifest.json). Every app reads and writes THIS shape; no app-private variants. Blocker fix, applies in the current RC (RULES-FREEZE exception).

{
  "schema": 1,
  "components": {
    "<component-id>": {
      "kind": "runtime | wheels | model | tool",
      "version": "2.11.0+cu130",
      "path": "runtime/py314-torch2.11-cu130",
      "sha256": "<hash of the archive or lock file it was installed from>",
      "installed_at": "2026-10-11T08:40:00+03:00",
      "refs": ["enlarge", "canvas"]
    }
  }
}

Rules - Top-level keys are exactly schema (integer) and components (object keyed by id). Unknown keys are preserved on write, never dropped. - Component ids are lowercase [a-z0-9][a-z0-9._-]*, e.g. python-3.14, torch-2.11.0-cu130, onnxruntime-gpu-1.31, tensorrt-10, ffmpeg-8, model-<sha256-prefix>. - Reuse when id + version + sha256 match; otherwise install side by side under a new path. Never overwrite a component another app refs. - refs: add your app id on install, remove it on uninstall; delete the component only when refs is empty. - Writes are atomic: write manifest.json.tmp, fsync, rename; take the lock file manifest.lock (exclusive, 30 s timeout) around read-modify-write. - Reader migration: if a file has version + resources (old Canvas shape), convert each resource into a component and write back schema 1. Readers MUST accept the old Enlarge shape too (it already is schema+components). - The shared Python runtime includes torch (cu130) and onnxruntime-gpu only; installing CPU onnxruntime into it is forbidden (it shadows the GPU build). TensorRT 10 libs ship as component tensorrt-10 for ORT 1.31.