GloriousFlywheel Advanced Runner Proof Matrix 2026-04-22
Snapshot date: 2026-04-22
Owner surfaces:
#320/TIN-335#333/TIN-371#338/TIN-377#340/TIN-378#335/TIN-376#312/TIN-330#44
This note turns the broad “advanced runners” productization language into one ordered proof matrix.
The goal is not to claim broad support. The goal is to say which runner-class lane is real now, which one is next, and what would count as proof before the platform broadens its claims.
Current Ordering
- KVM
- GPU / WebGPU / Dawn
- macOS
- riscv
- broader cross-forge or rarer follow-on execution lanes
Matrix
| Lane | Current state | Real proof already landed | Next truth surface | Why it is not mature yet |
|---|---|---|---|---|
| KVM | active adjacent platform ask | shared tinyland-nix-kvm lane is landed; repo-scoped nix-vm-test-kvm ran real terminal-first Rocky 10_1 smoke |
broader rockies execution, queue behavior, graphical/Budgie follow-on proof on honey |
execution is still too narrow and queue pressure still shapes the product boundary |
| GPU / WebGPU / Dawn | tightened bounded repo-scoped proof surface | Jesssullivan/cmux GPU Smoke Test (Self-Hosted) run 24756928163 completed success on main; job GPU smoke test (cmux-nix) built libghostty, built cmux-linux, ran Zig tests, and completed the GPU smoke workload behind the gpu-tests environment gate |
promote beyond repo-scoped cmux proof into one named shared honey GPU lane, then keep Dawn / WebGPU-specific canaries and operator docs as the broader follow-on |
the proof is real but repo-scoped, environment-gated, and still does not establish a shared ARC / honey GPU lifecycle or a Dawn / WebGPU contract |
| macOS | tightened bounded repo-scoped proof surface | tinyland-inc/lab Nix Build has repeated successful Build aarch64-darwin jobs on self-hosted runner xoxd-bates; lab#149 removed job-level continue-on-error, removed the permissive shell fallback on the Darwin build step, and re-proved Darwin on branch plus main |
decide whether to promote beyond repo-scoped proof into a platform-owned shared macOS runner contract | the proof is real and now stricter, but it is still repo-scoped, self-hosted-mac-specific, and not a shared platform-owned contract |
| riscv | not started | none | name the first workload, owner repo, and runtime boundary | the lane is still demand-shaped, not product-shaped |
| Cross-forge follow-ons | compatibility only | Forgejo honey proof exists as bounded compatibility evidence | repeatable GitHub-first runtime pattern before any parity claim | there is still no mature multi-forge runtime contract |
Explicit Backlog Mapping
#312/TIN-330: KVM execution substrate#338/TIN-377: boundedcmuxGPU proof surface#340/TIN-378: first sharedhoneyGPU lane after the boundedcmuxfloor#44: broader GPU / Dawn / WebGPU-specific slice after the first shared lane#333/TIN-371: broader advanced-runner proof matrix and sequencing#335/TIN-376: bounded macOS proof surfaceTIN-127: Linux-builder / FlakeHub publication boundaryTIN-165: bazel-registry truth convergence
What Graduation Should Mean
For any advanced runner class, “proved” should require:
- a named runner lane with explicit lifecycle ownership
- one real runtime proof surface on an authoritative branch
- explicit cache, bootstrap, and cleanup behavior
- one downstream canary or owned-repo workload
- public docs that describe the proved path rather than only prerequisites or design notes
Current Recommendation
- keep treating KVM as the only active richer-runner productization lane
- treat
cmuxGPU smoke as a named bounded proof surface, not as a shared platform-owned GPU contract - treat the next GPU move as one named shared
honeylane with one real canary before broadening into Dawn / WebGPU maturity claims - treat
labDarwin success as a named bounded macOS proof surface, not as a shared platform-owned contract - do not broaden riscv or cross-forge claims until each lane has a named proof surface