@wenathlan/saddle 1.8.18 → 2.0.2

package/docs/viability.md

@@ -0,0 +1,777 @@
+# Viability Proof - Why e2ugh Is Possible
+
+This is the key document of the repository: the argument that a 100 percent
+software engine can present CPUs, memory and GPUs that do not physically
+exist, execute real workloads on software rasterizers, and pass the hardware
+probes that AI agents and framework installers actually perform - the thesis
+that virtual hardware is functional spoofing, not fabricated throughput. Every
+claim below is anchored to a verified, dated source from the 2026-08-22
+research pass: the full methodology, the multilingual corpus design, the
+145-entry numbered bibliography and the three roadmap appendices (A, B, C)
+live in this document; the implemented
+capabilities live in the domain modules (`scheduler.ts`, `compute.ts` and
+kin, after the v3 redistribution of the retired `future.ts`); the optimization
+techniques that make the stack fast live in `performance.md`; the isolation
+stack that makes it safe lives in `security.md`; the engine landscape lives
+in `alternatives.md`; the virtualized hardware catalog itself (EPYC 9965, RTX
+5090, RTX PRO 6000, MI350X and the generated artifacts) lives in
+`hardware.md`, and the runtime tiers that execute it in `virtualization.md`.
+
+Contexts covered (25): functional spoofing, honest execution, procfs
+channels, sysinfo hooks, NVML shims, Khronos loaders, virtual nvidia-smi adapter 575.57.08,
+CUDA 12.9, memoverlay, dolos, GpuAdapter, detection-depth ladder, CPUID truth,
+MemorySwap -1, ShmSize 2g, MR !17813, MR 31551, Lavapipe CTS, Rusticl 3.1,
+Firecracker 125 ms, QEMU 11.1.0 MTTCG, GitHub Actions free, GHCR, static
+binary bypass, SECCOMP_RET_USER_NOTIF.
+
+## Contents
+
+1. [The thesis: functional spoofing, honest execution](#thesis)
+2. [How tools actually read hardware (attack surface)](#attack-surface)
+3. [The detection-depth ladder](#ladder)
+4. [The technical proof table (14 techniques)](#proof-table)
+5. [Why the earlier sixty-page "impossible" analysis was wrong](#refutation)
+6. [Performance envelope](#envelope)
+7. [The slowness question: answered 4.7x-12.35x](#slowness)
+8. [Honest limitations](#limitations)
+9. [Getting started on GitHub in an afternoon](#afternoon)
+10. [Cost analysis: zero dollars](#cost)
+11. [Verification reports](#verification)
+12. [Research methodology](#methodology)
+13. [The multilingual corpus](#multilingual)
+14. [Numbered bibliography (145)](#bibliography)
+15. [Appendix A. v5 future feature ledger (F-001..F-065, set A)](#appendix-a)
+16. [Appendix B. v6 future feature ledger (F-001..F-065, set B)](#appendix-b)
+17. [Appendix C. Backlog from the v4 archive (13 items)](#appendix-c)
+18. [Appendix D. Planned externals (not implementable in-engine)](#appendix-d)
+
+## The thesis: functional spoofing, honest execution
+
+The engine does not pretend to be physically faster than the host. The
+thesis is narrower and provable: standard Linux tooling reads hardware
+through a small number of well-defined channels - procfs files, the
+`sysinfo` libc call, the Vulkan/OpenGL/OpenCL loaders, and the NVML library -
+and each of those channels can be intercepted or replaced without root and
+without patching the tools themselves. The result is a sandbox where:
+
+1. `lscpu`, `free`, `htop` and language runtimes report the virtual board
+   (192 EPYC cores, 128 GB or up to 1 TiB of RAM) because they parse
+   `/proc/cpuinfo` and `/proc/meminfo`, which the engine renders
+   byte-realistically (`virtualcpu.ts`, `virtualmemory.ts`).
+2. `nvidia-smi`, `gpustat` and `pynvml` report an RTX 5090, RTX PRO 6000 or
+   A100 because NVML entry points are shimmed through `LD_PRELOAD` with
+   `dlsym(RTLD_NEXT)` chaining, and the CLI itself can be replaced by a
+   pixel-compatible virtual adapter emitting the confirmed header `NVIDIA-SMI 575.57.08
+   Driver Version: 575.57.08   CUDA Version: 12.9`.
+3. `torch.cuda.is_available()`, `clinfo`, `vulkaninfo` and `glxinfo` succeed
+   because the execution is real: llvmpipe (OpenGL 4.6 core, 161/161
+   extensions, AVX-512 via Gallivm), lavapipe (Vulkan 1.3-conformant, 1.4
+   exposed) and Rusticl (OpenCL 3.1) genuinely run compute - on CPU threads,
+   not silicon.
+
+Spoofing the reports and executing on software is the entire trick. It fools
+the overwhelming majority of real-world probes - framework guards, installer
+checks, agent reconnaissance - while never claiming throughput the host
+lacks. A tool that reads `/proc`, calls NVML, or queries a Khronos API sees
+the virtual board; a kernel-level audit (CPUID instruction, actual core
+count, TPM) sees the truth, and the engine never claims otherwise. That is
+the precise sense in which the engine is a 99 percent illusion: depths 1-3
+of probing (below) cover what essentially every agent, installer and
+benchmark harness does before running code; depths 4-5 do not, and the
+engine says so.
+
+## How tools actually read hardware (attack surface)
+
+| Tool | Read channel | Interception point |
+|---|---|---|
+| `lscpu` | sysfs plus `/proc/cpuinfo` | generate/replace `/proc/cpuinfo` per man page lscpu(1) |
+| `free` | parses `/proc/meminfo` | generate/replace `/proc/meminfo` per man page free(1) |
+| `clinfo` | OpenCL properties via ICD dispatch (`/etc/OpenCL/vendors/*.icd`) | point the ICD at Rusticl over llvmpipe |
+| `vulkaninfo` | Vulkan API through `libvulkan.so.1`, ICDs under `/usr/share/vulkan/icd.d` | set `VK_DRIVER_FILES` to `lvp_icd.x86_64.json` |
+| `glxinfo` | GLX implementation on an X display | Xvfb with `+extension GLX`, `LIBGL_ALWAYS_SOFTWARE` |
+| `nvidia-smi`, `gpustat`, `pynvml` | NVML (`libnvidia-ml.so`) | LD_PRELOAD NVML shim; or replace the `nvidia-smi` binary |
+| Python `psutil`, Go `mem` | libc `sysinfo()` | hook `sysinfo` via `dlsym(RTLD_NEXT)` |
+| `torch.cuda` | libcuda/libcudart entry points | GpuAdapter-style API interception; NVML shim for presence probes |
+
+## The detection-depth ladder
+
+Probes exist at different depths, and the engine answers each depth with the
+cheapest mechanism that satisfies it, exposing each rung as a distinct
+runtime tier (Docker, QEMU TCG, Firecracker) through the orchestrator, so an
+operator pays only for the depth their workloads need.
+
+| Depth | Example probe | What satisfies it | Engine tier |
+|---|---|---|---|
+| 1. Parsing /proc | `lscpu`, `free`, `htop`, most installers | byte-realistic generated cpuinfo/meminfo (memoverlay/dolos technique) | Docker |
+| 2. libc calls | `sysinfo()`, `uname()` consumers (psutil, Go mem stats) | `dlsym(RTLD_NEXT)` hook with `mem_unit` normalization | Docker |
+| 3. Hardware APIs | NVML, `nvidia-smi`, `clinfo`, `vulkaninfo`, `glxinfo` | NVML shim + virtual CLI + real Mesa software drivers | Docker |
+| 4. Instruction level | CPUID, MSR reads, core-count truth | emulated CPU with `-cpu EPYC-v5` and virtualized SMBIOS | QEMU TCG |
+| 5. Kernel objects | TPM, real device PCI enumeration | honest boundary: not virtualized without KVM-tier passthrough tricks | Firecracker (KVM hosts) |
+
+The ladder framing is the practical answer to "can it be detected?": yes, at
+depths 4-5, by tools that almost no agent or framework uses before running
+code; and no, at depths 1-3, which is where 99 percent of real probing
+happens. The engine's claim is exactly that scope.
+
+## The technical proof table
+
+Each row is one load-bearing technique, its mechanism, the proof that it
+works, and the primary source. Together the rows cover the full spoofing
+surface: CPU, memory, GPU presence, GPU APIs, execution engines, packaging
+and continuous integration.
+
+| # | Technique | Mechanism | Proof | Source |
+|---|---|---|---|---|
+| 1 | memoverlay (memory virtual identity) | `LD_PRELOAD` intercepts `open/open64/openat/fopen/fopen64/freopen/freopen64`, redirects `/proc/meminfo` to a regenerated `/tmp` file; an inotify (IN_MODIFY) thread keeps values fresh; `FAKEMEM_MEM=16G` sets the size | 362-line C file fetched and read; builds with `clang -shared -fPIC memoverlay.c -ldl` | https://github.com/stantheawesomeman/memoverlay |
+| 2 | dolos (CPU/sysinfo virtual identity) | Hooks `sysinfo()` via `dlsym(RTLD_NEXT)`; false `totalram` normalized by `mem_unit`, `freeram` at ~25 percent, plus `uname`/`gethostname` hooks; global install via `/etc/ld.so.preload` | 144-line hooks.c fetched and read | https://github.com/cdt4/dolos/blob/main/dolos/hooks.c |
+| 3 | fake-nvidia-smi | Pure-Python (stdlib only) generator with profiles h100 81,559 MiB/700 W, a100 40,536 MiB/400 W, v100/p100/k80/t4; randomized temp/power/bus-id; output matches the official manual | README (163 lines) fetched; header confirmed `NVIDIA-SMI 575.57.08 / CUDA 12.9` | https://github.com/pogusthewhisper/fake-nvidia-smi |
+| 4 | NVML unified shim | `libnvml-unified.so` via `LD_PRELOAD`; `nvmlDeviceGetCount` falls back to `cudaGetDeviceCount`, `nvmlDeviceGetMemoryInfo` reads `/proc/meminfo`, `nvmlDeviceGetName` uses `cudaGetDeviceProperties`; index-as-pointer handles | Built for DGX Spark GB10 (Ubuntu 24.04.3, driver 580.126.09, CUDA 12.8); README fetched (230 lines) | https://github.com/rick-hsu/nvml-unified-shim |
+| 5 | GpuAdapter (CUDA API interception) | Intercepts `libcuda`, `libcudart`, `libcublas`, `libnvidia-ml`, `libnccl` via `LD_PRELOAD` (Linux) / `DYLD_INSERT_LIBRARIES` (macOS); virtual `nvidia-smi` including `--query-gpu`, topo, MIG; `--devices "a100:2,h100:2"` | README fetched (1,200 lines), repo verified HTTP 200 | https://github.com/FanBB2333/GpuAdapter |
+| 6 | Docker memory virtualization | `--memory-swap -1` grants unlimited swap (docs: tools like `free` report host swap); `--shm-size 2g` fixes the PyTorch DataLoader Bus error from the 64 MB default; `vm.overcommit_memory=2` semantics | Official CLI reference and resource-constraints docs fetched | https://docs.docker.com/reference/cli/docker/container/run/ |
+| 7 | llvmpipe AVX-512 | Gallivm emits 512-bit vector code with runtime CPU detection (merged Mesa 22.3 via MR !17813); `LP_NATIVE_VECTOR_WIDTH=512` selects it | Merge request confirmed via Phoronix coverage linking the MR | https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17813 |
+| 8 | LP_MAX_THREADS | llvmpipe raster threads raised 16 to 32 by MR 31551 (compile-time ceiling in `lp_limits.h`); `LP_NUM_THREADS=0` selects auto | Mesa env-vars and llvmpipe docs fetched | https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/31551 |
+| 9 | Lavapipe (Vulkan on CPU) | Software Vulkan driver exposing Vulkan 1.4 (Mesa 25.1+), officially conformant at Vulkan 1.3 (CTS 1.3.1.1, SPI submission 2022-07-19); selected with `VK_DRIVER_FILES` (VK_ICD_FILENAMES deprecated) | Khronos conformant-products list checked | https://www.khronos.org/conformance/adopters/conformant-products/vulkan |
+| 10 | Rusticl (OpenCL on CPU) | Rust OpenCL frontend, sole Mesa OpenCL since Clover was deleted in Mesa 25.2; OpenCL 3.1 in Mesa 26.2; `RUSTICL_DEVICE_TYPE=gpu` spoofs the device class | Same-day OpenCL 3.1 support reported; release notes fetched | https://www.phoronix.com/news/OpenCL-3.1-Same-Day-Rusticl |
+| 11 | Xvfb headless GLX | Virtual X server provides a GLX display without hardware: `Xvfb :99 -ac -screen 0 1920x1080x24 -nolisten tcp +extension GLX +render -noreset` | Man pages and a public docker-opengl entrypoint fetched | https://man.archlinux.org/man/Xvfb.1 |
+| 12 | Firecracker microVM | 125 ms cold boot, <5 MiB RAM per VM, 150 VMs/s per host; snapshot restore 3-5 ms with File or Uffd backends; diff snapshots in preview; v1.16.1 (2026-07-02) | Release feed and snapshot-support doc fetched | https://github.com/firecracker-microvm/firecracker/releases |
+| 13 | QEMU 11.1.0 MTTCG | Software CPU emulation with one host thread per vCPU (`-accel tcg,thread=multi,tb-size=1024`), `-cpu EPYC-v5` (present in target/i386/cpu.c), microvm machine for fast boot; works without KVM, which GitHub runners lack | QEMU 11.1.0 release notes (2026-08-11) fetched; EPYC-v5 confirmed in source | https://www.qemu.org/2026/08/11/qemu-11-1-0 |
+| 14 | GitHub Actions public runners | Unlimited minutes for public repositories; runners execute Docker natively (KVM absent, hence TCG); GHCR serves public images free | Usage limits documented; KVM absence confirmed in runner-images issues | https://docs.github.com/en/actions/usage-limits |
+
+## Why the earlier sixty-page "impossible" analysis was wrong
+
+A previous analysis produced outside this repository concluded over roughly
+sixty pages that virtual hardware was impractical: it assumed the virtual identity
+would need kernel modules, hypervisor privileges, per-tool patching, or
+undisclosed GPU emulation. The 2026 research pass overturned each pillar.
+
+| Claim in the old analysis | Research finding | Evidence |
+|---|---|---|
+| "You cannot spoof /proc without a kernel module" | Userspace works: `LD_PRELOAD` intercepts the open family and redirects procfs reads to generated files; inside containers, bind-mounting a generated file over `/proc/cpuinfo` needs no privilege beyond the namespace | memoverlay source at github.com/stantheawesomeman/memoverlay; dolos hooks at github.com/cdt4/dolos; bind-mount discussion github.com/moby/moby/issues/16423 |
+| "Tools use many different syscalls, so interception never generalizes" | The channels are few: the complete open-family plus `sysinfo`/`uname` covers procfs-based tools; NVML and the Khronos loaders are single dispatch libraries | dolos hooks `sysinfo` with `mem_unit` normalization; loader docs confirm one dispatch library per API |
+| "There is no working virtual nvidia-smi adapter" | Multiple public implementations exist, one in pure-Python stdlib with A100/H100/V100/P100/K80/T4 profiles, plus a Go variant in a fake GPU operator | github.com/pogusthewhisper/fake-nvidia-smi; github.com/run-ai/fake-gpu-operator |
+| "A virtual GPU cannot satisfy torch.cuda" | NVML shims make device-count probes pass, and full CUDA API interception libraries exist for API-level workloads | github.com/rick-hsu/nvml-unified-shim; github.com/FanBB2333/GpuAdapter |
+| "Software GPUs are toys, not conformant" | Mesa's stack is Khronos-conformant: lavapipe is conformant Vulkan 1.3; Rusticl is conformant OpenCL 3.0 (now 3.1); llvmpipe is 100 percent OpenGL 4.6 core (161/161) with AVX-512 | khronos.org conformant products; mesamatrix.net; docs.mesa3d.org/drivers/llvmpipe.html |
+| "The compute would be unusably slow" | Workloads are CPU-bound but real: llvmpipe runs up to 32 raster threads with AVX-512 paths, and typical agent workloads (installs, imports, small tensors, probes) complete in interactive time | docs.mesa3d.org/drivers/llvmpipe.html and MR 31551 history |
+| "It would cost too much to build and host" | The whole stack is open source (MIT/Apache-2.0) and public repositories get unlimited Actions minutes plus free GHCR storage | docs.github.com/en/actions/usage-limits; cost table below |
+
+The root error of the old analysis was insufficient research depth: it
+reasoned from first principles about what "should" be possible instead of
+checking the public artifacts that already demonstrate each mechanism.
+Every row above is a shipping, fetchable project or a documented conformance
+result, verified by HTTP fetch on 2026-08-22.
+
+## Performance envelope
+
+Spoofing reports is only half of viability; the engine must also start fast
+enough for interactive agents and execute real work at usable speed.
+
+| Metric | Value | Source |
+|---|---|---|
+| Firecracker cold boot | 125 ms | https://firecracker-microvm.github.io |
+| Firecracker snapshot restore | as low as ~4 ms (10 ms for a full Linux guest, AWS) | https://brooker.co.za/blog/2022/11/29/snapstart.html and https://arxiv.org/abs/2102.12892 |
+| Firecracker density | 150 microVMs/s per host, <5 MiB RAM per VM | Firecracker homepage |
+| E2B cold start (target benchmark) | ~150 ms via UFFD snapshot restore | https://e2b.dev/blog/firecracker-vs-qemu |
+| Snapshot-restore sandboxes generally | 5-30 ms | https://dev.to/adwitiya/how-i-built-sandboxes-that-boot-in-28ms-using-firecracker-snapshots-i0k |
+| QEMU microvm machine | ~4x faster boot than default QEMU, ~3x slower than Firecracker | E2B benchmark blog |
+| Docker cold start | 1-2 s | E2B benchmark blog |
+| llvmpipe threads | up to 32 raster threads (LP_MAX_THREADS ceiling) | docs.mesa3d.org/drivers/llvmpipe.html |
+| llvmpipe vector width | 512-bit AVX-512 paths with runtime detection | MR !17813 via Phoronix |
+
+The conclusion from those rows: the latency thesis (E2B-class sub-200 ms
+starts) is achievable on KVM hosts with Firecracker warm pools, degrades to
+low seconds on KVM-less CI with Docker plus QEMU TCG, and the throughput
+thesis is bounded by CPU cores - an EPYC-class host with 32 software raster
+threads serves OpenGL 4.6, Vulkan 1.3+ and OpenCL 3.1 workloads in
+interactive time for agent-scale jobs, not for training runs, and the
+planners in `compute.ts` encode exactly that boundary so agents receive
+honest estimates.
+
+## The slowness question: answered 4.7x-12.35x
+
+The same research wave set out to answer the original complaint - "it will
+be slow" - with a constrained experiment: reproduce at least a 3x speedup on
+a commodity GitHub Actions `ubuntu-24.04` runner (2 vCPU, 7 GB RAM) using
+only free toolchains, zero hardware dollars. The verdict was YES, PROVEN
+VIABLE, with this evidence set (frozen ledger, five-run averages):
+
+| Metric | Baseline | Optimized | Gain |
+|---|---|---|---|
+| End-to-end CI (`make ci`) | 312.4 s (Node 18 + Docker 24) | 47.1 s (Node 22.12 + Bun + BuildKit) | 6.63x |
+| Config parse, 14 mandatory files | 4.2 s | 0.34 s | 12.35x |
+| Cold image build | 187 s | 29.4 s | 6.36x |
+| VM provision | 42 s | 8.9 s | 4.7x |
+| QEMU TCG boot (Debian cloud image) | 11.8 s single-thread | 3.1 s MTTCG + io_uring | 3.8x |
+| e2e passage test | 94 s | 19.2 s | 4.89x |
+
+The stack prescription that produced those numbers (Node strip-types, Bun
+sidecar, TypeScript isolatedDeclarations, Python free-threaded generation,
+BuildKit gha cache, QEMU MTTCG, io_uring plus eBPF, WASI validators, Rust
+crates, TOML/YAML/JSON5 policy) is documented technique-by-technique in
+`performance.md`; this section records only the verdict and the fact that
+every item was MIT/Apache/PSF licensed - the risk line in the source report
+reads "None at $0", with a downgrade path to pure Node if Bun is
+unavailable. Reproduction costs nothing: the commands clone the flat repo,
+run `node --experimental-strip-types`, `bun --smol test`, a buildx build
+with gha cache, a QEMU MTTCG boot and a two-line eBPF load, all exiting 0
+on the free tier.
+
+## Honest limitations
+
+Viability claims are only credible with the failure modes stated plainly.
+
+| Limitation | Detail | Mitigation |
+|---|---|---|
+| Static binaries bypass LD_PRELOAD | Go static binaries issue syscalls directly without libc PLT; musl uses `__syscall` directly; interposition has documented limits | seccomp `SECCOMP_RET_USER_NOTIF` on `openat` with `SECCOMP_IOCTL_NOTIF_ADDFD` fd injection intercepts all binaries including static Go; ptrace `PTRACE_O_TRACECLONE` suppresses stray SIGSYS; documented in the three-tier termux-etc-redirect project and the binary-rewriting interposition writeup (sources 79-80) |
+| Performance is CPU-bound | llvmpipe/lavapipe/Rusticl execute on CPU threads; a virtual RTX 5090 does not deliver 1.79 TB/s; numbers are reports, not throughput | The engine never claims silicon speed; workload planners budget against realistic software throughput so agents do not submit impossible jobs |
+| A replaced nvidia-smi does not run CUDA kernels | A replaced CLI only renders text; `torch` kernels would fail without a CUDA implementation | GpuAdapter intercepts the CUDA API surface itself, so API-level workloads execute on a CPU backend; NVML shims satisfy presence probes; both are public, working projects |
+| CPUID and kernel-visible truth remain | Without a VM, CPUID, actual core counts, hardware GPU ids and TPM reveal the host | The engine offers the QEMU tier (EPYC-v5, `-cpu host,hypervisor=off -smbios` spoofing) for workloads that probe below procfs; the dolos README itself points the same way |
+| GitHub runners lack KVM | Firecracker/Kata/Cloud Hypervisor cannot run on hosted runners | CI uses QEMU TCG plus Docker; Firecracker paths are exercised on self-hosted or bare-metal runners; documented in the workflows |
+
+None of these limitations is fatal; each has either a working public
+workaround or an honest scope boundary already built into the engine's
+runtime tiers. The design stance: the engine never lies to the operator -
+the virtual identity targets the workload's view of hardware, not the operator's view
+of reality, and every module documents which depth of the ladder it defends.
+
+## Getting started on GitHub in an afternoon
+
+The entire pipeline runs on free tiers. A contributor with `git` and `gh`
+installed goes from zero to a running virtualized sandbox:
+
+```bash
+# 1. Create the public repository (public = free Actions + free GHCR)
+gh repo create e2ugh --public --source=. --push
+git push -u origin main
+
+# 2. Actions builds the image on push (workflows in .github/workflows):
+#    publishghcr.yml (combined build+publish) pushes to ghcr.io/<owner>/e2ugh
+gh run watch
+
+# 3. Anyone can pull and run the engine, no credentials needed
+docker pull ghcr.io/<owner>/e2ugh:latest
+docker run --rm -it \
+  --memory-swap -1 --shm-size 2g --cpus 8 \
+  -e LIBGL_ALWAYS_SOFTWARE=1 \
+  -e GALLIUM_DRIVER=llvmpipe \
+  -e VK_DRIVER_FILES=/usr/share/vulkan/icd.d/lvp_icd.x86_64.json \
+  -e RUSTICL_ENABLE=llvmpipe \
+  -e RUSTICL_DEVICE_TYPE=gpu \
+  ghcr.io/<owner>/e2ugh:latest \
+  lscpu && free -h && nvidia-smi
+
+# 4. Local iteration without GitHub: compose brings the same flags up
+docker compose up
+```
+
+Inside that container, `lscpu` prints the virtual EPYC, `free` prints the
+virtual memory, and `nvidia-smi` prints the virtual board - all rendered by
+this repository's modules (`virtualcpu.ts` and `virtualmemory.ts` generate
+byte-realistic cpuinfo/meminfo; `render.ts` builds the Mesa environment and
+the SMI table; the `smiadapter` class in `qemubridge.py` carries the same
+profiles on the Python path), while OpenGL/Vulkan/OpenCL calls execute on
+Mesa's software drivers for real. The workflow files implement steps 2 and
+3, so the commands above are the exact loop the repository was designed
+around: push, watch Actions build for free, pull from GHCR, run the probes.
+
+## Cost analysis: zero dollars
+
+| Line item | Cost | Why it is free |
+|---|---|---|
+| Source control and CI | $0 | GitHub public repositories get unlimited Actions minutes on standard runners |
+| Image registry | $0 | GHCR storage and pulls for public images are free |
+| Spoofing libraries | $0 | memoverlay, dolos, fake-nvidia-smi, NVML shim, GpuAdapter are public repositories |
+| Graphics stack | $0 | Mesa (MIT), LLVM (Apache-2.0 with LLVM exceptions) |
+| Virtualization | $0 | QEMU (GPL-2.0), Firecracker (Apache-2.0), Docker Engine (Apache-2.0) |
+| Language runtime | $0 | Node.js and TypeScript (MIT/Apache-2.0) |
+| Execution hardware | $0 | The contributor's laptop or the free runner; every heavy tier (QEMU TCG, Mesa) runs without KVM or GPUs |
+
+The only nonzero path is optional: self-hosted runners for KVM-backed
+Firecracker benchmarks on rented bare metal, marked as an operator choice
+rather than a requirement. A sanity check on the economics: a public
+repository with a 20-minute build pushed five times a day consumes 100
+minutes of Actions per day - well inside the unlimited public-repo
+allowance - and each published image is a public GHCR artifact that
+unlimited users may pull without bandwidth charges.
+
+## Verification reports
+
+Two verification reports from the lineage are preserved as audit artifacts
+and reconciled into this repository's structure. The v5 report (66 files):
+flat structure PASS (0 violations), mandatory 14 configs PASS (14/14), JSON
+validity 10/10, YAML validity 4/4, SHA256 duplicate groups 0, docs quality
+PASS, version compliance PASS - total file count FAIL at 66 (over the 45
+budget), verdict CONDITIONAL PASS pending Wave-2 consolidation. The v6
+report (60 files, 35 docs, 13,155 markdown lines): SHA256 dedup PASS, dot
+configs 6/6 PASS, JSONs 8/8 PASS, lowercase naming PASS, root-first PASS,
+library-first index.ts PASS, date-first 2026-08-22 PASS in 59/60 files,
+toolchain pins PASS - file count FAIL (60), GHA workflows FAIL (0 found),
+docs count PARTIAL (35 of 40, quality PASS). The reconciliation lesson
+applied to e2ugh v2: the 8-document consolidation (readme, architecture,
+hardware, virtualization, security, performance, alternatives, viability)
+absorbs the redundant document families identified in the analysis phase,
+workflows ship in `.github/workflows` from the start, and the tree stays
+flat (root plus `libs/`, `docs/`, `tests/`, `.github/workflows/`) with the
+same checks the reports codified (counts, flat
+naming, JSON validity, deduplication, version pins) now enforced by CI
+rather than by a post-hoc report.
+
+How each claim in this document is validated: library claims were fetched
+over HTTP (README lengths recorded above); version claims carry release
+feeds (GitHub repositories expose `/releases.atom`); conformance claims
+resolve to the Khronos conformant-products list; the Docker and kernel
+claims resolve to official documentation; the economics resolve to the
+GitHub usage-limits page. Re-verification is mechanical: fetch the feed,
+compare the newest tag with the pinned version, and check the dated
+announcement for changed behavior - version facts are centralized in the
+catalog constants (`engineversions` in `index.ts`, the runtime catalog in
+`orchestrator.ts`, the version catalog in `render.ts`), so a bump is a
+one-constant edit plus a note in the correction table.
+
+## Research methodology
+
+The research pass ran as parallel agents with adversarial verification
+rules, executed as Google Dorking across 15 site-scoped domains
+(github.com, gitlab.freedesktop.org, phoronix.com, arxiv.org,
+docs.mesa3d.org, stackoverflow.com, reddit.com, medium.com, youtube.com,
+archive.org, wikipedia.org, substack.com, freedesktop.org, docs.docker.com,
+kernel.org, llvm.org). Four verification rules governed every entry:
+
+1. Primary sources first. Version claims came from release announcements,
+   product pages and vendor documentation, never from aggregator articles.
+2. Fetch, do not trust. Repositories claimed to exist were fetched over
+   HTTP: dead repositories were recorded as negative results rather than
+   dropped silently.
+3. Cross-check dates. Every version fact carries a release date and a
+   source, so a version can be re-verified against its feed later.
+4. Record the negative results too. Dead ends (x0x0x00/gpuadapter marked dead,
+   Daytona production source closed June 2026, Nabla EOL 2022) stayed in
+   the catalog as boundaries of the claim space.
+
+The dork catalog that drove the pass includes the 25 canonical queries
+recorded in the method note (firecracker 125 ms, mesa 25.2.7, avx512
+LP_NATIVE_VECTOR_WIDTH 512, mttcg on arxiv, JA/ZH/KO/DE locale variants,
+MemorySwap -1, virtualhardware.c sysinfo overlay, virtual_cpuinfo EPYC 9965,
+boards SP5, cores 1-192, threads 2-384, RTX 5090 10DE:2B85, RTX PRO 6000
+10DE:2BB5, everstore 429 bypass, among others) plus the high-yield dorks
+`site:github.com fake-gpu-operator nvml-mock`,
+`site:gitlab.freedesktop.org mesa merge_requests 17813 31551`, and
+`site:stackoverflow.com Docker shm-size 2g PyTorch DataLoader bus error`.
+
+## The multilingual corpus
+
+The corpus design emulates 100,000 discrete searches across six primary
+locales without violating rate limits, expanded from a 180-term matrix.
+
+| Dimension | Distribution |
+|---|---|
+| Total simulated queries | 100,000 |
+| Per locale | EN-US 21,000-36,000 (two designs merged), EN-GB 14,000-18,000, ZH 18,000, JA 16,000-16,500, DE 8,000-14,000, KO 8,000-12,500 (11 languages in the extended method: plus PT-BR, ES, FR, RU, HI) |
+| Per platform | GitHub 28 percent, StackOverflow 15, Reddit 12, X/Twitter 10, YouTube 8, Wikipedia 8, archive.org 7, Medium 5, Substack 4, cnki.net 3 |
+| Per epoch | 2026: 35-45 percent, 2020-2025: 30, 2010-2019: 15, 2000-2009: 8, 1900-1999: 6, 1500-1899: 6 |
+
+Locale keyword clusters (illustrative): EN-US `qemu vgpu 2026 passthrough`,
+`modular unlimited vcpu 1-4096`, `cgroup v2 cpu.weight io_uring`; JA
+`QEMU vGPU パススルー 2026`, `NUMA アウェア ピニング`; ZH `QEMU vGPU 直通
+2026 黑威尔`, `sched_ext 调度器 6.12 内核`; KO `QEMU vGPU 패스스루`,
+`모듈러 무제한 vCPU`; DE `QEMU vGPU Passthrough 2026 Blackwell`,
+`NUMA-bewusstes Pinning`. The DE corpus emphasizes formal verification
+language; JA/ZH/KO emphasize prime vCPU edge cases (13, 37, 127) for
+scheduler stress. The dork arsenal comprises 67 high-precision queries
+across five categories (config hunting `filetype:toml vm.config`, GPU
+profiles, deep hardware inventories, container filters, social signals),
+with language-extended variants (`site:github.com 日本語 qemu vGPU 解説`)
+and operator chaining (`(qemu OR kvm) (Blackwell OR GB202) stars:>100
+pushed:>2026-01-01`). Epoch findings: 2026 concentrates on declarative
+infrastructure and heterogeneous compute; 2020-2025 on TOML-first
+configuration and vfio/SR-IOV standardization; 2010-2019 on the libvirt and
+Docker foundations (ivshmem introduced 2010; Docker 2013); and the
+1500-1899 layer maps conceptual precursors (Bouchon 1725, Babbage 1837,
+Menabrea 1843) to justify static-JSON hardware description as a design
+lineage rather than a novelty. Reproducibility: the corpus ships a library-first TypeScript dorking script
+(native node:fs, crypto randomUUID traceIds, SHA256 dedup) plus a query
+distribution matrix and a non-duplication guarantee.
+
+## Numbered bibliography
+
+Academic-style citation list in first-appearance order: toolchain, graphics,
+virtualization, hardware, spoofing, alternatives, security; entries 104-107
+add the lineage artifacts unique to this consolidation; entries 108-145 add
+the primary sources recovered from the v4 analysis pool (release calendars,
+interposition references, kernel and systemd manuals, sandbox landscape
+analyses and the pre-rename spellings of the spoofing lineage repositories).
+
+1. Node.js 26.7.0 release. https://nodejs.org/en/blog/release/v26.7.0
+2. Node.js 24.19.0 LTS release. https://nodejs.org/en/blog/release/v24.19.0
+3. Node.js 26.0.0 release (V8 14.6). https://nodejs.org/en/blog/release/v26.0.0
+4. TypeScript npm registry page. https://www.npmjs.com/package/typescript
+5. TypeScript 7.0 announcement. https://devblogs.microsoft.com/typescript/announcing-typescript-7-0
+6. Docker Engine 29 release notes. https://docs.docker.com/engine/release-notes/29
+7. Docker Buildx releases. https://github.com/docker/buildx/releases
+8. Docker Compose releases. https://github.com/docker/compose/releases
+9. GitHub blog: checkout v7 safer defaults. https://github.blog/changelog/2026-06-18-safer-pull_request_target-defaults-for-github-actions-checkout
+10. actions/setup-node releases. https://github.com/actions/setup-node/releases
+11. docker/setup-buildx-action marketplace. https://github.com/marketplace/actions/docker-setup-buildx
+12. docker/login-action releases. https://github.com/docker/login-action/releases
+13. docker/metadata-action marketplace. https://github.com/marketplace/actions/docker-metadata-action
+14. docker/build-push-action releases. https://github.com/docker/build-push-action/releases
+15. actions/cache repository. https://github.com/actions/cache
+16. Biome npm registry page. https://www.npmjs.com/package/@biomejs/biome
+17. Bun 1.4 release blog. https://bun.com/blog
+18. Mesa 3D homepage. https://www.mesa3d.org
+19. Mesa 26.2.1 release notes. https://docs.mesa3d.org/relnotes/26.2.1.html
+20. Mesa 26.2.0 release notes. https://docs.mesa3d.org/relnotes/26.2.0.html
+21. Mesa 26.1.0 release notes. https://docs.mesa3d.org/relnotes/26.1.0.html
+22. Mesa release calendar. https://docs.mesa3d.org/release-calendar.html
+23. Mesa environment variables. https://docs.mesa3d.org/envvars.html
+24. llvmpipe driver documentation. https://docs.mesa3d.org/drivers/llvmpipe.html
+25. Rusticl documentation. https://docs.mesa3d.org/rusticl.html
+26. Mesa MR !17813: AVX-512 in Gallivm. https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17813
+27. Phoronix: AVX-512 lands in llvmpipe. https://www.phoronix.com/news/Mesa-AVX-512-LLVMpipe-Start
+28. Phoronix: Lavapipe Vulkan 1.4. https://www.phoronix.com/news/Mesa-Lavapipe-Vulkan-1.4
+29. Phoronix: Lavapipe Vulkan 1.3 conformant. https://www.phoronix.com/news/Lavapipe-Vulkan-1.3-Official
+30. Khronos conformant products, Vulkan. https://www.khronos.org/conformance/adopters/conformant-products/vulkan
+31. Phoronix: Rusticl OpenCL 3.1 same-day. https://www.phoronix.com/news/OpenCL-3.1-Same-Day-Rusticl
+32. Mesamatrix. https://mesamatrix.net
+33. LLVM releases. https://github.com/llvm/llvm-project/releases
+34. apt.llvm.org. https://apt.llvm.org
+35. Xvfb(1) manual; utensils/docker-opengl entrypoint. https://man.archlinux.org/man/Xvfb.1 and https://github.com/utensils/docker-opengl
+36. QEMU 11.1.0 announcement and invocation reference. https://www.qemu.org/2026/08/11/qemu-11-1-0 and https://www.qemu.org/docs/master/system/invocation.html
+37. QEMU MTTCG internals. https://www.qemu.org/docs/master/devel/multi-thread-tcg.html
+38. Firecracker releases. https://github.com/firecracker-microvm/firecracker/releases
+39. Firecracker project page. https://firecracker-microvm.github.io
+40. Firecracker snapshot support. https://github.com/firecracker-microvm/firecracker/blob/main/docs/snapshot-support.md
+41. Brooker: SnapStart and snapshot latency; Firecracker NSDI 2021 paper. https://brooker.co.za/blog/2022/11/29/snapstart.html and https://arxiv.org/abs/2102.12892
+42. Cloud Hypervisor releases. https://github.com/cloud-hypervisor/cloud-hypervisor/releases
+43. Cloud Hypervisor vfio-user documentation. https://raw.githubusercontent.com/cloud-hypervisor/cloud-hypervisor/main/docs/vfio-user.md
+44. gVisor releases atom feed. https://github.com/google/gvisor/releases.atom
+45. gVisor platforms guide. https://gvisor.dev/docs/architecture_guide/platforms
+46. gVisor performance guide. https://gvisor.dev/docs/architecture_guide/performance
+47. Kata Containers 4.0.0 overview. https://katacontainers.io/blog/kata-containers-4-0-0-release-overview
+48. Kata Containers releases. https://github.com/kata-containers/kata-containers/releases
+49. CRIU releases. https://github.com/checkpoint-restore/criu/releases
+50. E2B: Firecracker vs QEMU. https://e2b.dev/blog/firecracker-vs-qemu
+51. E2B self-hosting notes. https://github.com/e2b-dev/infra/blob/main/self-host.md
+52. GitHub runner-images issue 12933 (KVM). https://github.com/actions/runner-images/issues/12933
+53. Actuated: KVM in GitHub Actions. https://actuated.com/blog/kvm-in-github-actions
+54. AMD EPYC 9965 product page. https://www.amd.com/en/products/processors/server/epyc/9005-series/amd-epyc-9965.html
+55. AMD Ryzen 9 9950X3D product page. https://www.amd.com/en/products/processors/desktops/ryzen/9000-series/amd-ryzen-9-9950x3d.html
+56. AMD Threadripper PRO 9995WX product page. https://www.amd.com/en/products/processors/workstations/ryzen-threadripper/9000-wx-series/amd-ryzen-threadripper-pro-9995wx.html
+57. Intel Core Ultra 9 285K specifications. https://www.intel.com/content/www/us/en/products/sku/241060/intel-core-ultra-9-processor-285k-36m-cache-up-to-5-70-ghz/specifications.html
+58. AMD Radeon RX 9070 XT product page. https://www.amd.com/en/products/graphics/desktops/radeon/9000-series/amd-radeon-rx-9070xt.html
+59. AMD Instinct MI350X product page. https://www.amd.com/en/products/accelerators/instinct/mi350/mi350x.html
+60. AMD CDNA 4 whitepaper. https://www.amd.com/content/dam/amd/en/documents/instinct-tech-docs/white-papers/amd-cdna-4-architecture-whitepaper.pdf
+61. NVIDIA RTX 5090 product page. https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5090
+62. NVIDIA RTX PRO 6000 Blackwell pages. https://www.nvidia.com/en-us/products/workstations/professional-desktop-gpus/rtx-pro-6000
+63. NVIDIA DGX B200 page. https://www.nvidia.com/en-us/data-center/dgx-b200
+64. TechPowerUp GPU database (RTX 5090) and DeviceHunt PCI database. https://www.techpowerup.com/gpu-specs/geforce-rtx-5090.c4216 and https://devicehunt.com/view/type/pci/vendor/1002/device/75A0
+65. memoverlay repository. https://github.com/stantheawesomeman/memoverlay
+66. dolos hooks.c. https://github.com/cdt4/dolos/blob/main/dolos/hooks.c
+67. fake-nvidia-smi repository. https://github.com/pogusthewhisper/fake-nvidia-smi
+68. run-ai/fake-gpu-operator. https://github.com/run-ai/fake-gpu-operator
+69. rick-hsu/nvml-unified-shim. https://github.com/rick-hsu/nvml-unified-shim
+70. NVIDIA developer forum thread 358869. https://forums.developer.nvidia.com/t/358869
+71. FanBB2333/GpuAdapter. https://github.com/FanBB2333/GpuAdapter
+72. Docker container run reference. https://docs.docker.com/reference/cli/docker/container/run/
+73. Docker resource constraints guide. https://docs.docker.com/engine/containers/resource_constraints/
+74. Compose specification, services. https://github.com/compose-spec/compose-spec/blob/master/05-services.md
+75. Linux overcommit accounting. https://github.com/torvalds/linux/blob/master/Documentation/mm/overcommit-accounting.rst
+76. Google Colab tools issue 329 (shm). https://github.com/googlecolab/colabtools/issues/329
+77. nvidia-smi documentation. https://docs.nvidia.com/deploy/nvidia-smi/
+78. clinfo repository and Vulkan-Loader driver interface. https://github.com/Oblomov/clinfo and https://github.com/KhronosGroup/Vulkan-Loader/blob/main/docs/LoaderDriverInterface.md
+79. rios0rios0/termux-etc-redirect. https://github.com/rios0rios0/termux-etc-redirect
+80. Reczey: interposing static binaries. https://balintreczey.hu/blog/think-you-cant-interpose-static-binaries-with-ld_preload-think-again
+81. Wasmtime releases. https://github.com/bytecodealliance/wasmtime/releases
+82. Wasmtime: interrupting WebAssembly. https://docs.wasmtime.dev/examples-interrupting-wasm
+83. WASI roadmap. https://wasi.dev/roadmap
+84. Web Component Model. https://component-model.bytecodealliance.org
+85. WebContainers. https://webcontainers.io
+86. Daytona. https://daytona.io
+87. Modal. https://modal.com
+88. Fly.io Machines. https://fly.io/machines
+89. E2B. https://e2b.dev and https://github.com/e2b-dev/e2b
+90. Cognitora. https://cognitora.dev
+91. Vercel Sandbox GA. https://vercel.com/blog/vercel-sandbox-is-now-generally-available
+92. Vercel $1M HackerOne challenge. https://vercel.com/blog/one-million-dollar-hacker-challenge-for-vercel-sandbox
+93. apple/container. https://github.com/apple/container
+94. WWDC26 session 389. https://developer.apple.com/videos/play/wwdc2026/389
+95. Landlock kernel documentation. https://docs.kernel.org/userspace-api/landlock.html
+96. landlock.io news. https://landlock.io/news
+97. seccomp(2) manual. https://man7.org/linux/man-pages/man2/seccomp.2.html
+98. moby issue 47532 (io_uring). https://github.com/moby/moby/issues/47532
+99. LWN: task-level io_uring restrictions. https://lwn.net/Articles/1054225
+100. OpenSSL post-quantum cryptography. https://openssl-corporation.org/post-quantum.html
+101. RFC 10024. https://datatracker.ietf.org/doc/rfc10024
+102. Cloudflare eBPF LSM. https://blog.cloudflare.com/live-patch-security-vulnerabilities-with-ebpf-lsm
+103. systemd cgroup delegation. https://systemd.io/CGROUP_DELEGATION
+104. Verification report v5 (66 files, conditional pass). Lineage artifact preserved in the analysis pool; checks summarized in the verification section.
+105. Verification report v6 (60 files, 11 pass / 2 fail / 1 partial). Lineage artifact preserved in the analysis pool; checks summarized in the verification section.
+106. Multilingual research corpus (100k simulated queries, 6 locales, 67-dork arsenal, epoch layers 1500-2026). Consolidated into the multilingual section; reproduction script is library-first TypeScript.
+107. Viability optimization wave (4.7x-12.35x, $0, five-run averages on ubuntu-24.04 runners). Techniques in `performance.md`; verdict preserved in the slowness section of this document.
+108. endoflife.date: Node.js release and support calendar. https://endoflife.date/nodejs
+109. endoflife.date: Docker Engine release and support calendar. https://endoflife.date/docker-engine
+110. Node.js download index and release tags. https://nodejs.org/en/download and https://github.com/nodejs/node/releases
+111. GitHub runner-images issue 14062 (the second hosted-runner KVM availability thread). https://github.com/actions/runner-images/issues/14062
+112. GitHub community discussion 160591 (nested virtualization on hosted runners). https://github.com/orgs/community/discussions/160591
+113. GitHub community discussion 156389 (runner capability thread cited by the research corpus). https://github.com/orgs/community/discussions/156389
+114. QEMU issue 750: the LD_PRELOAD `fakefopen.c` demonstration. https://gitlab.com/qemu-project/qemu/-/issues/750
+115. moby issue 16423: `/proc` bind-mount masquerading inside containers. https://github.com/moby/moby/issues/16423
+116. Darimont gist: faking the processor count for the JVM. https://gist.github.com/thomasdarimont/ca5ca088a9007eb25dff677a8ace3ff1
+117. lima-vm/alpine-lima issue 41: ZRAM swap inside the guest. https://github.com/lima-vm/alpine-lima/issues/41
+118. lscpu(1) and free(1) manuals (the two probes the generators must satisfy). https://man7.org/linux/man-pages/man1/lscpu.1.html and https://man7.org/linux/man-pages/man1/free.1.html
+119. Stack Overflow 53122005: `/dev/shm` bus-error root cause (the PyTorch 64 MB default). https://stackoverflow.com/questions/53122005
+120. Stack Overflow 77489799: `openat` interposition versus `open`. https://stackoverflow.com/questions/77489799
+121. Hacker News 19190275: the limits of LD_PRELOAD on static binaries. https://news.ycombinator.com/item?id=19190275
+122. fakemem repository (the pre-rename spelling of memoverlay, entry 65). https://github.com/stantheawesomeman/fakemem
+123. FakeGPU repository (the pre-rename spelling of GpuAdapter, entry 71). https://github.com/FanBB2333/FakeGPU
+124. KhronosGroup/Vulkan-Tools (the vulkaninfo probe source). https://github.com/KhronosGroup/Vulkan-Tools
+125. LLVM releases atom feed and the Xserver(1) manual. https://github.com/llvm/llvm-project/releases.atom and https://man.archlinux.org/man/Xserver.1
+126. QEMU download page. https://www.qemu.org/download
+127. Mesa news feed. https://www.mesa3d.org/news
+128. eBPF LSM program documentation. https://docs.kernel.org/bpf/prog_lsm.html
+129. cgroup v2 administration guide. https://www.kernel.org/doc/html/latest/admin-guide/cgroup-v2.html
+130. systemd.resource-control(5). https://www.freedesktop.org/software/systemd/man/systemd.resource-control.html
+131. Arch wiki: Cgroups and Vulkan pages. https://wiki.archlinux.org/title/Cgroups and https://wiki.archlinux.org/title/Vulkan
+132. CVE-2026-46315 (SentinelOne database) and the io_uring hardening write-up. https://sentinelone.com/vulnerability-database/cve-2026-46315 and https://systemshardening.com/articles/linux/io-uring-hardening
+133. OpenSSL 3.5 PQ-by-default analysis and the gVisor seccomp blog. https://postquantum.com/security-pqc/openssl-3-5-pqc-default and https://gvisor.dev/blog/2024/02/01/seccomp
+134. Spheron: the AI agent code-execution sandbox landscape (E2B, Daytona, Firecracker) and the B200 complete guide. https://www.spheron.network/blog/ai-agent-code-execution-sandbox-e2b-daytona-firecracker and https://www.spheron.network/blog/nvidia-b200-complete-guide
+135. B200 specification and pricing references (Jarvislabs, Modal). https://jarvislabs.ai/ai-faqs/nvidia-b200-specs and https://modal.com/blog/nvidia-b200-pricing
+136. NVIDIA RTX PRO 6000 Blackwell Server Edition page. https://www.nvidia.com/en-us/data-center/rtx-pro-6000-blackwell-server-edition
+137. Apple Containers for Linux/macOS coverage (Help Net Security) and the HPCwire AI wire note. https://helpnetsecurity.com/2026/07/07/apple-container-open-source-linux-mac/ and https://hpcwire.com/aiwire/2026/07/22/
+138. WebContainers: the introduction post and the webcontainer-core repository. https://blog.stackblitz.com/posts/introducing-webcontainers and https://github.com/stackblitz/webcontainer-core
+139. Northflank engineering: what is gVisor; Modal versus Vercel Sandbox; self-hostable alternatives to Daytona. https://northflank.com/blog/what-is-gvisor, https://northflank.com/blog/modal-vs-vercel-sandbox and https://northflank.com/blog/self-hostable-alternatives-to-daytona
+140. Sandbox landscape roundups kept as context (Morph, Puter, Upstash, Modal, Fly.io architecture). https://morphllm.com/comparisons/daytona-alternative, https://developer.puter.com/blog/fly-io-alternatives, https://upstash.com/blog/best-sandbox-providers-for-ai-agents, https://modal.com/resources/best-serverless-sandboxes-ai-code-execution and https://fly.io/docs/reference/architecture
+141. Container checkpoint/restore walkthrough (OneUptime) plus the Cloud Hypervisor and CRIU root sites. https://oneuptime.com/blog/post/2026-02-09-container-checkpoint-restore-criu, https://www.cloudhypervisor.org and https://criu.org
+142. Tracee vulnerability tracing and the post-quantum TLS references (AWS SDK reference, Cloudflare PQC support). https://aquasec.com/blog/linux-vulnerabilitie-tracee, https://docs.aws.amazon.com/sdkref/latest/guide/pqtls-details.html and https://developers.cloudflare.com/ssl/post-quantum-cryptography/pqc-support
+143. Raptor Computing Systems (the open POWER workstation lineage cited by the research corpus). https://www.raptorcs.com
+144. Spoofing-lineage artifacts unique to the v4 pool: kagari issue 16 and the edge-ai-benchmark `fake_cpuinfo.c`. https://github.com/itsakeyfut/kagari/issues/16 and https://github.com/joeltadeu/edge-ai-benchmark
+145. Proof-of-concept artifacts: x0x0x00/fakegpu, the hashcloak zk-proof-of-assets memory-boundary demonstration (`docker run -m 100G --memory-swap -1`) and the libvirt MTTCG patch thread. https://github.com/x0x0x00/fakegpu, https://github.com/hashcloak/zk-proof-of-assets, https://lists.libvirt.org and https://news.ycombinator.com/item?id=44784059
+
+---
+
+## Appendix A. v5 future feature ledger {#appendix-a}
+
+The v5 set of the F-001..F-065 future feature ledger, preserved from the
+SADDLE v5 roadmap (`ROADMAP_FUTURE_FEATURES.md`, wave 11-Future, categories
+A-H: CXL and memory, GPU and compute virtualization, scheduling and
+virtualization core, WASM and heterogeneity, I/O and storage, DPU and
+networking, confidential computing and policy, observability and operations).
+IDs are shared with appendix B; the titles differ because the v6 rewrite
+re-scored every item. The same IDs are referenced by `readme.md` (roadmap),
+`architecture.md` (reserved slots) and `virtualization.md` (F-ledger notes).
+
+| ID | Feature (v5 set A) |
+|---|---|
+| F-001 | CXL 3.0 Memory Pooling and Fabric Manager *(implemented: plancxltype3, virtualization.ts + cxldevice tier plan, virtualmemory.ts)* |
+| F-002 | CXL Tiered Memory QoS and Promotion Engine (DAMON) *(implemented: psidamoninfo, virtualization.ts)* |
+| F-003 | NVLink-C2C Coherent vRAM Sharing |
+| F-004 | HBM3e Disaggregated Pools via CXL 3.1 and UCIe *(implemented: HBM_BLACKWELL tier, virtualmemory.ts)* |
+| F-005 | Persistent Memory v2 and Battery-Backed CXL DCD *(implemented: PMEM_DAX tier detection, virtualmemory.ts)* |
+| F-006 | Tiered vNUMA v2 with Auto Topology Discovery *(implemented: planumalayout, scheduler.ts)* |
+| F-007 | vGPU Live Migration Blackwell GB200/B100 *(implemented: livemigration, virtualization.ts)* |
+| F-008 | Blackwell FP4/FP6/FP8 Virtualization and Dynamic MIGv2 *(implemented: MIGLAYOUTBLACKWELL, virtualgpu.ts + planmiglayout, scheduler.ts)* |
+| F-009 | GPU Time-Slicing QoS and Preemption Guarantees *(implemented: VGPUPROFILES time-sliced vgpu scheduler mirror, virtualgpu.ts)* |
+| F-010 | Persistent vGPU and vRAM Snapshots plus CUDA Checkpoint *(implemented: diffsnapshotengine + planfilesnapshot, virtualization.ts + cricheckpointer, orchestrator.ts)* |
+| F-011 | NCCL 3.x Virtualization and Multi-Node GPU Fabric |
+| F-012 | NVLink Switch 3rd-Gen Multi-Node Pooling (576 GPUs) |
+| F-013 | WebGPU Virtualization Layer (WGPU Virt, virtio-webgpu) *(implemented: forcefallbackadapterplan, compute.ts)* |
+| F-014 | Virtio-GPU Native Context (Venus/Vulkan) Passthrough |
+| F-015 | Vulkan Video Encode/Decode Virtualization (virtio-video) |
+| F-016 | Intel AMX / AMD XDNA NPU Virtualization *(implemented: plannputiles, compute.ts)* |
+| F-017 | Serverless GPU Functions (Scale-to-Zero vGPU) |
+| F-018 | Virtio-FS plus DPU File System Disaggregation |
+| F-019 | FPGA Time-Multiplexed vFPGA (Versal, Agilex) |
+| F-020 | USB4 / Thunderbolt Passthrough v2 with Hotplug |
+| F-021 | WASM MicroVM Runtime (Wasmtime/WAMR/WasmEdge, fuel) *(implemented: wasiasyncruntime + fuelmeter, compute.ts)* |
+| F-022 | Rust Rewrite of the QEMU Bridge (qemu-bridge-rs) |
+| F-023 | eBPF Hierarchical CPU Scheduler (sched_ext + BPF) *(implemented: schedextinfo, virtualization.ts)* |
+| F-024 | Live vCPU Hotplug plus eBPF Rebalancing Daemon *(implemented: vcpuhotplug, scheduler.ts)* |
+| F-025 | Heterogeneous vCPU Topology (P+E, X3D + Zen5) |
+| F-026 | vNUMA Auto-Rebalance DAMON + AutoNUMA v2 |
+| F-027 | Formal Verification of the Scheduler (TLA+ / Alloy) |
+| F-028 | Memory Dedup KSM++ DAMON-aware Scan |
+| F-029 | KVM Hypercall Acceleration (Rust kvm-bindings) |
+| F-030 | Cross-Arch Emulation x86-AArch64 plus Rosetta-like Cache |
+| F-031 | io_uring Zero-Copy Passage Engine |
+| F-032 | SPDK + NVMe-oF Userspace Storage Stack |
+| F-033 | virtio-blk over io_uring plus I/O Polling |
+| F-034 | ZNS and FDP SSD-Aware Virtual Disk Provisioning *(implemented: planznszones, virtualization.ts)* |
+| F-035 | Distributed Checkpointing 100B+ Training (SnapshotFS) |
+| F-036 | Virtio-FS DAX plus CXL Shared Memory File Cache |
+| F-037 | DPU Offload (BlueField-3 / Pensando / Intel IPU) |
+| F-038 | SR-IOV Dynamic VF Provisioning and Live Attach *(implemented: setupsriov, virtualization.ts)* |
+| F-039 | XDP/eBPF Hardware-Accelerated Packet Pipeline *(implemented: ebpfprogram, virtualization.ts)* |
+| F-040 | DPDK Userspace Bridge (virtio-user) |
+| F-041 | QUIC + HTTP/3 Tunneling for Secure Passage |
+| F-042 | WireGuard Mesh + Tailscale for vNets *(implemented: tailscaleendpointplanner, compute.ts)* |
+| F-043 | SmartNIC P4 Programmable Packet Steering *(implemented: buildp4skeleton, virtualization.ts)* |
+| F-044 | Time-Sensitive Networking (TSN) for RT VMs *(implemented: buildtsnschedule, virtualization.ts)* |
+| F-045 | Virtio-snd + PipeWire Low-Latency Audio Passage |
+| F-046 | WebTransport + WebCodecs Display Streaming |
+| F-047 | Confidential VMs: SEV-SNP, TDX, ARM CCA *(implemented: buildqemucmd sev-snp-guest/tdx argv, virtualization.ts)* |
+| F-048 | HW-Attested Guest Integrity (SVSM vTPM + Remote Attestation) |
+| F-049 | Encrypted Live Migration (TLS 1.3 + SEV-SNP) |
+| F-050 | eBPF LSM + Landlock Security Policy Engine *(implemented: landlockrulebuilder + seccompbuilder, security.ts)* |
+| F-051 | OCI Hooks + CDI for GPU Container Parity |
+| F-052 | gVisor-like Syscall Interception (WASM companion) *(implemented: gvisorruntime, orchestrator.ts)* |
+| F-053 | AI Autoscaler RL and Predictive Queuing (stable-baselines3) *(implemented: rlautoscaler epsilon-greedy bandit, scheduler.ts)* |
+| F-054 | Predictive HW Failure SMART + eBPF Telemetry *(implemented: anomalydetector, scheduler.ts)* |
+| F-055 | Generative Config Synthesis (LLM-assisted JSONs) |
+| F-056 | OpenTelemetry + eBPF Deep Profiling Mesh *(implemented: otelmetricsbridge, scheduler.ts + ebpfgputelemetry, compute.ts)* |
+| F-057 | Carbon-Aware Scheduling (WattTime + Kepler) |
+| F-058 | Multi-Tenant QoS cgroups v2 + PSI + eBPF *(implemented: tenantcgroupbuilder, virtualization.ts + tenantadmission, scheduler.ts)* |
+| F-059 | AI Workload Fingerprinting and Auto-Tuning (Dynamo/XLA/vLLM) |
+| F-060 | Chaos Engineering Framework (ChaosVM) |
+| F-061 | Ray + vLLM Distributed Inference Orchestrator *(implemented: planrayvllmcluster, compute.ts)* |
+| F-062 | Kubernetes Operator (SADDLE CRDs) *(implemented: buildvmcrdmanifest, compute.ts)* |
+| F-063 | CRDT Distributed vm.config (Automerge + Yjs) *(implemented: vmconfigcrdt, scheduler.ts + statevectorsync, compute.ts)* |
+| F-064 | GitOps Flux CD + SADDLE CRDs + OPA *(implemented: planfluxgitops, orchestrator.ts)* |
+| F-065 | Federated SADDLE Clusters Global Scheduler (10k nodes) *(implemented: planfederatedquota, compute.ts)* |
+
+Source document annexes (dependency graph, spec extensions, version/CI,
+file-count budget, keywords) stay in the pool artifact; this appendix carries
+the verbatim feature ledger with horizons H1/H2/H3, complexity L1-L5 and
+five-axis impact recorded per item in the source. The v5-C feature audit
+(worklog task v5-C) grepped every row against the modules: rows marked
+*(implemented: symbol, module)* carry a builder/planner/class; unmarked
+rows are either implementable planners still open or planned externals
+(listed in appendix D).
+
+## Appendix B. v6 future feature ledger {#appendix-b}
+
+The v6 set of the same ledger, preserved from the SADDLE v6 rewrite
+(`roadmapfuturefeatures.md`, 12 sections plus metadata; quantified targets in
+parentheses are the v6 planning values, not measured results).
+
+| ID | Feature (v6 set B) |
+|---|---|
+| F-001 | CXL 3.0 Type-3 Pooling Fabric (2-4 TiB, +18%, -34% per GB) *(implemented: plancxltype3, virtualization.ts)* |
+| F-002 | CXL 2.0 Fallback + HMAT 2.0 (latency 22/80) *(implemented: plancxltype3 spec '2.0' + hmat entries)* |
+| F-003 | Ballooning v2 with PSI (free-page-hint, +14% at 4.5x) |
+| F-004 | NUMA Distance-Aware Pinning + HMAT (+22%) *(implemented: planumalayout + numactlargs, scheduler.ts)* |
+| F-005 | vGPU Live Migration Pre-Copy dirty-ring 65536 (under 800 ms) |
+| F-006 | Blackwell MIG v2 14 hybrid slices (14x density) *(implemented: MIGLAYOUTBLACKWELL, virtualgpu.ts)* |
+| F-007 | NVLink-C2C 900 GB/s GB200 Superchip (624 GB coherent) |
+| F-008 | vRAM Slicing 1 MB step, DPU accounting 10M ops/s |
+| F-009 | vGPU EEVDF Fair-Share Low-Latency (+27% p95) |
+| F-010 | Blackwell FP4 9000 TFLOPS Passthrough (4x versus FP8) |
+| F-011 | GPU Overcommit CXL P2P spill (192 to 384 GB, -8%) |
+| F-012 | GPU Power Capping NVML (B200 1000 W / 700 W liquid) |
+| F-013 | SR-IOV 32 VF per pGPU Blackwell iommufd |
+| F-014 | GPU-IOV ATS/PRI Isolation fail-closed |
+| F-015 | WASM MicroVM + QEMU microvm (cold 45 to 6 ms) *(implemented: wasiasyncruntime, compute.ts + microvm runtimes, orchestrator.ts)* |
+| F-016 | WASM Cold Start under 5 ms (memfd + io_uring zstd, 4.8 ms p50) *(implemented: wasmmempool, compute.ts)* |
+| F-017 | WASI Preview2 FS Bridge node:fs (+19%) |
+| F-018 | WASI-NN GPU FP4 via WGPU 22 (H3) |
+| F-019 | Nested WASM inside QEMU (SEV-SNP outer) |
+| F-020 | sched_ext eBPF vCPU Scheduler (scx_rusty, +18%) *(implemented: schedextinfo, virtualization.ts)* |
+| F-021 | cgroup v2 cpu.weight Autotune PSI (1..10000) |
+| F-022 | eBPF XDP virtio-net Offload 9000 MTU (+34%) |
+| F-023 | eBPF io_uring Latency Histogram (0.05-32 ms) |
+| F-024 | io_uring Zero-Copy Passage vsock+unix (3.2 to 0.9 ms) |
+| F-025 | io_uring Fixed Buffers Registered (8192 KB, +28%) |
+| F-026 | io_uring NVMe Passthrough 14 GB/s PCIe 5.0 (+64%) |
+| F-027 | sendmsg MSG_ZEROCOPY unix fallback (+19%) |
+| F-028 | AI Autoscaler vCPU 1-4096 (sktime 0.35 LSTM, -31%) *(implemented: rlautoscaler + psilstm, scheduler.ts)* |
+| F-029 | AI GPU Demand Forecast (predictive MIG, -27% idle) |
+| F-030 | AI Anomaly ECC/Thermal (failure 30 min ahead, 99.99%) *(implemented: anomalydetector, scheduler.ts)* |
+| F-031 | AI Bin-Packing OR-Tools heterogeneous (+38%) |
+| F-032 | AI FinOps per-Tenant ($/vCPU + $/GB + $/W, Kepler) |
+| F-033 | AMD SEV-SNP OVMF Secure Boot (-6% perf) |
+| F-034 | Intel TDX 2nd Gen (TDVF, QGS quote) |
+| F-035 | Confidential GPU H100/B200 CCC (+NV-Trust) |
+| F-036 | Attestation TPM 2.0 + OVMF enrolled-keys (swtpm 0.8) *(implemented: OVMF secure boot argv, virtualization.ts)* |
+| F-037 | SEV-SNP Live Migration with Rekeying (HKDF + TPM) |
+| F-038 | DPU BlueField-3 Virtio Data Path (DOCA 2.9) |
+| F-039 | DPU Storage NVMe-oF + SPDK 24.09 |
+| F-040 | NVLink Switch Fabric NVL72 576 GPUs (720 PF FP4) |
+| F-041 | PCIe 5.0 AER DPU-accelerated (+99.99%) |
+| F-042 | RoCE v2 GPUDirect RDMA MTTG (OFED 24.10, -22%) |
+| F-043 | Rust qemubridge.rs Tokio (QMP 12 ms to 1.8 ms) |
+| F-044 | Rust QMP Wrapper (retry 50 ms x 2^n, napi-rs) |
+| F-045 | Rust virtio Emulation vhost-user (io_uring 0.7) |
+| F-046 | Rust CXL Mailbox FM-API (QoS class) |
+| F-047 | Rust eBPF Loader libbpf-rs 0.24 (CO-RE) |
+| F-048 | WebGPU Virt WGPU + Dawn + Vulkan 1.4 (40 ctx/B200) |
+| F-049 | WebGPU Compute Passthrough WASI-NN (Naga to PTX, +32x) |
+| F-050 | WebGPU Dawn + Vulkan Video NVENC AV1 8K (+44%) |
+| F-051 | OCI WASM Layers + Buildx (dedup +33%) |
+| F-052 | CRIU + QEMU incremental + io_uring (3.2 s to 0.45 s) |
+| F-053 | Live Upgrade QEMU 9.1.2 to 10.0 x-colo zero-downtime |
+| F-054 | Docker cgroupv2 + crun + CDI nvidia.yaml |
+| F-055 | Passage HTTP/3 QUIC 0-RTT TLS1.3-only (Pingora 0.4.2) |
+| F-056 | mTLS per-VM + JWT rotation 12 h (X509 ed25519) |
+| F-057 | OTel + eBPF Profiles + Prometheus 9090 *(implemented: otelmetricsbridge, scheduler.ts)* |
+| F-058 | Chaos Fault Injection QEMU + DPU (+31% resiliency) |
+| F-059 | Multi-Tenant MTTG QoS + Weighted vRAM (+22%) *(implemented: tenantqosmap + MTTG grid, virtualization.ts)* |
+| F-060 | NVENC 8K AV1 + H.266 VVC (2.1x Ada, 32x 4K/B200) |
+| F-061 | Heterogeneous CPU X3D + Intel Hybrid (+23% gaming) *(implemented: ccd-aware vcpupinningmap, virtualmemory.ts)* |
+| F-062 | Energy-Aware Carbon Scheduling (Kepler 0.8, -19%) |
+| F-063 | OPA Rego Policy Engine (vram 1024-786432 step 1) |
+| F-064 | Forge Mirror Bidirectional Sync (notify 6.0) |
+| F-065 | Docs Auto-Gen over 40 files (typedoc 0.26 + mkdocs) |
+
+Source-document metadata preserved with the set: section 12 dependency
+matrix, section 13 release trains 33/27/5 (33 near-term, 27 mid, 5 far), and
+section 15 aggregate targets (+27.4% performance geo-mean, 3.1x density,
+CVSS -38%, cost -29%). Rows share the v5-C audit marks of appendix A
+(same IDs); the Rust rewrite family (F-043 to F-047) is tracked as
+planned externals in appendix D.
+
+## Appendix C. Backlog from the v4 archive {#appendix-c}
+
+The 13-item `roadmapbacklog` of the retired `future.ts` (v4 archive, lines
+2701-2715 of the pool artifact), the only block of that file absent from the
+v3 redistribution. The register lives on as `roadmapbacklog` in compute.ts;
+the v5-C audit promoted four rows (1, 3, 8, 9) from backlog strings to real
+builders, and the remaining rows are recorded so the archive survives the
+module decomposition.
+
+| # | Backlog item (v4 archive) | Note |
+|---|---|---|
+| 1 | Intel AMX tile real passthrough | beyond the `+amx-tile` flag: actual TDP-pipe emulation *(implemented as builder: plannputiles, compute.ts; real pipe emulation stays external)* |
+| 2 | Dual-X3D detection for the 9950X3D2 | both CCDs carrying V-Cache; today only the cache-size rule selects it |
+| 3 | WebTransport datagram relay | browser-facing passage twin of the QUIC gateway *(implemented: planwebtransportendpoint, compute.ts)* |
+| 4 | Turborepo remote cache on GHCR | free-tier build cache beside the gha cache |
+| 5 | pnpm v9 content-addressable store | alternative package manager lane |
+| 6 | QUIC + HTTP/3 control plane | passage already speaks it for data; the control channel does not |
+| 7 | P2P WebRTC mesh sandboxes | browser-to-browser without a relay |
+| 8 | OPFS browser sandboxes | Origin Private File System as a guest disk *(implemented: planopfsstorage, compute.ts)* |
+| 9 | WebCodecs + WebGPU encode | hardware-free encode inside the browser tier *(implemented: planwebcodecspipeline, compute.ts)* |
+| 10 | SharedArrayBuffer cross-origin isolation | COOP/COEP headers for multi-origin sandboxes |
+| 11 | WASM JSPI scheduling | JavaScript Promise Integration for the wasm tier |
+| 12 | GHCR gha layer reuse | cross-repository cache sharing |
+| 13 | Nix flakes hermetic builds | reproducible host toolchain lane |
+
+## Appendix D. Planned externals (not implementable in-engine) {#appendix-d}
+
+Result of the v5-C feature audit (worklog task v5-C): every ledger row was
+classified as (a) already implemented, (b) implementable as a dependency-free
+builder/planner in TypeScript, or (c) not implementable as code of this engine
+because it requires physical hardware, platform firmware or paid external
+services. This appendix registers the category (c) rows as planned externals:
+the engine can plan and emit their configuration, but execution happens
+outside the repository.
+
+| Ledger rows | External dependency | Why it cannot be engine code |
+|---|---|---|
+| A F-003, A F-012, B F-007, B F-040 | NVLink-C2C / NVLink Switch fabric | physical interconnect hardware between real GPUs |
+| A F-011 | NCCL 3.x | guest-side library owned by NVIDIA |
+| A F-014, A F-015 | Virtio-GPU Venus / virtio-video upstream | requires upstream QEMU/Venus merge, not a TS shim |
+| A F-017, B F-035 | confidential GPU H100/B200 CCC + NV-Trust | hardware trusted-launch in NVIDIA silicon |
+| A F-019 | vFPGA Versal/Agilex time multiplexing | physical FPGA reconfiguration ports |
+| A F-020 | USB4/Thunderbolt passthrough | host controller and physical dongles |
+| A F-022, B F-043..F-047 | Rust rewrite family (qemubridge.rs, QMP wrapper, vhost-user, FM-API mailbox, libbpf-rs) | a second Rust codebase outside this TS engine |
+| A F-027 | TLA+ / Alloy formal verification | external model checkers and proof toolchains |
+| A F-032 | SPDK + NVMe-oF target | userspace storage daemon with kernel-bypass drivers |
+| A F-037, B F-038, B F-039 | BlueField-3 / Pensando / Intel IPU DPU offload | physical DPU cards and DOCA licenses |
+| A F-041, B F-042, B F-055 | QUIC 0-RTT / RoCE v2 OFED / Pingora edge | kernel RDMA NICs and edge proxy deployments |
+| A F-048 | SVSM vTPM remote attestation | AMD KDS attestation service and platform firmware |
+| A F-049, B F-037 | encrypted live migration rekeying (HKDF + TPM) | guest firmware secrets held by the platform |
+| A F-055 | LLM-assisted config synthesis | external model inference (the engine stays offline) |
+| A F-057 / B F-062, B F-032 | WattTime carbon API, Kepler exporters, FinOps meters | external metered services and cluster agents |
+| B F-051, B F-064, B F-065 | OCI WASM layers on GHCR, Forge mirror, typedoc/mkdocs | registry, git hosting and doc toolchain pipelines |
+| C 2, 4, 5, 7, 10, 11, 12, 13 | dual-X3D silicon, Turborepo/pnpm lanes, WebRTC mesh infra, SAB headers on real origins, JSPI engine flags, GHCR sharing, Nix flakes | silicon, browser vendors, registries and CI toolchains |
+
+For these rows the engine's contract is: emit complete, validated plans
+(argv, manifests, schedules) that external executors consume verbatim; the
+builders registered in appendices A-C are the deliverable.