Hosted GPU tier: endpoint, launcher, reaper, runner image, clients (GJOB-088)

.github/workflows/runner-gpu.yml

@@ -0,0 +1,49 @@
+# Build the hosted-tier GPU runner image and push to GHCR. Default = fast path from the prebuilt
+# binary release asset; set build_from_source=true to compile (see docker/runner-gpu/Dockerfile).
+name: runner-gpu
+on:
+  workflow_dispatch:
+    inputs:
+      kokkos_arch:
+        description: 'Kokkos arch (ADA89 | AMPERE80)'
+        default: ADA89
+        type: string
+      tarball_url:
+        description: 'Prebuilt lammps tarball URL (release asset)'
+        default: https://github.com/forcefieldsilicon/mdengine/releases/download/runner-bin-20260905/lammps-kokkos-ADA89.tar.gz
+        type: string
+      build_from_source:
+        description: 'Compile instead of using the tarball'
+        default: false
+        type: boolean
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+      packages: write
+    steps:
+      - uses: actions/checkout@v4
+      - uses: docker/setup-buildx-action@v3
+      - uses: docker/login-action@v3
+        with:
+          registry: ghcr.io
+          username: ${{ github.actor }}
+          password: ${{ secrets.GITHUB_TOKEN }}
+      - name: Pick Dockerfile
+        id: pick
+        run: |
+          if [ "${{ inputs.build_from_source }}" = "true" ]; then echo "file=docker/runner-gpu/Dockerfile" >> "$GITHUB_OUTPUT"; else echo "file=docker/runner-gpu/Dockerfile.prebuilt" >> "$GITHUB_OUTPUT"; fi
+      - uses: docker/build-push-action@v6
+        with:
+          context: docker/runner-gpu
+          file: ${{ steps.pick.outputs.file }}
+          push: true
+          build-args: |
+            KOKKOS_ARCH=${{ inputs.kokkos_arch }}
+            LMP_TARBALL_URL=${{ inputs.tarball_url }}
+          tags: |
+            ghcr.io/forcefieldsilicon/mdengine-runner-gpu:${{ inputs.kokkos_arch }}
+            ghcr.io/forcefieldsilicon/mdengine-runner-gpu:${{ inputs.kokkos_arch }}-${{ github.sha }}
+          labels: |
+            org.opencontainers.image.source=https://github.com/forcefieldsilicon/mdengine

.gitignore

@@ -1,13 +1,15 @@
-Packages
+__pycache__/
+! .env.example
 .build
-xcuserdata
-*.xcodeproj
-DerivedData/
 .DS_Store
-db.sqlite
-.swiftpm
 .env
 .env.*
-! .env.example
+.swiftpm
 .vscode
+*.pyc
+*.xcodeproj
+db.sqlite
+DerivedData/
 dist/
+Packages
+xcuserdata

CLAUDE.md

@@ -6,7 +6,7 @@ macOS MD workbench: `MDEngine` (SwiftUI+Metal viewer), `mdengine-cli`,
 (Gitinama Inc.).
 
 ## Commands
-- Build: `swift build` · Tests: `swift test` (16 cases, keep green)
+- Build: `swift build` · Tests: `swift test` (20 cases, keep green)
 - Release: `swift build -c release` — **ALWAYS run after changes**:
   `/opt/homebrew/bin/mdengine` and `mdengine-mcp` are symlinks into
   `.build/release/`; a debug-only build leaves every other session running
@@ -19,6 +19,10 @@ macOS MD workbench: `MDEngine` (SwiftUI+Metal viewer), `mdengine-cli`,
   `~/.mdengine/hosts.json`; test bed = `localhost-test` host over ssh to this Mac
   (own key in ~/.ssh/authorized_keys). `lmp` must be an absolute path — no login
   PATH over ssh. Release packaging: `make_tools.sh` (CLI+MCP tarball).
+- Hosted GPU tier (`Sources/LAMMPSCore/HostedClient.swift`, `hosted/CONTRACT.md`): dev
+  loop = `python3 hosted/mock/mock_endpoint.py --port 8788 --data <dir>` +
+  `MDENGINE_HOSTED_URL=http://127.0.0.1:8788/v1 MDENGINE_HOSTED_LAUNCH="{lmp} -in {input} -log log.lammps"`,
+  pod stand-in = `docker/runner-gpu/runner.sh` with `MDE_WORK`/`LMP` (see hosted/README.md).
 
 ## Gotchas that already bit
 - SPM does NOT prune deleted resources from an existing