W0 probe evidence and handoff

Date: 2026-09-30. Branch: codex/webgpu-w0. Base: 29436c9 (current remote main when implementation began). Status: in progress; hardware acceptance is pending. No browser engine backend or itch.io game has been released.

Implemented

• Standalone exception-free C++23/Emscripten WebGPU C API probe with async adapter/device requests, animated clear submission, failure status, and per-frame handle release.
• SDK/port version and checksum lock, explicit web toolchain policy, ADR 0009.
• Build, format/tidy, and licensed ZIP package commands; sandboxed iframe harness and a dedicated compile/package workflow.
• Persistent stage plan and acceptance/handoff tracking for subsequent PRs.

Toolchain observed

Component	Version / identity
Emscripten	4.0.23; compiler driver revision 7a5d93b50f6a3a35e85a0d2fc9e667b8498e6aed
SDK build	aaa43392544d695232b70eda706d751f18980c2a
LLVM bundled with SDK	22.0.0git; 5243501cca02d7e54294d4bb5de0a85b06c40b7d
SDK libc++	_LIBCPP_VERSION=200100
Emdawnwebgpu	v20251002.162335; downloaded ZIP SHA512 matches config/web_toolchain.json
Repository analysis	clang-format 18 and clang-tidy 18

The browser reached the adapter-result callback after the runtime feature checks, demonstrating execution of std::expected, floating std::to_chars, and std::format with production C++ exceptions disabled. This is a targeted feature probe, not a claim that every engine dependency is portable.

Commands and results

Commands below ran from the isolated worktree. The native source was unchanged.

Check	Command	Result
Web Debug compile	./scripts/webgpu-probe build	Pass; project warning set, -Werror, C++23, -fno-exceptions
Web Debug format/tidy	./scripts/webgpu-probe check	Pass with version 18 against the wasm source/sysroot/port headers
Web Release compile	./scripts/webgpu-probe build --configuration Release	Pass
Web Release format/tidy	./scripts/webgpu-probe check --configuration Release	Pass
ZIP packaging	./scripts/webgpu-probe package and --configuration Release	Pass; root HTML/JS/wasm, harness, and licenses inspected
Script syntax	python3 -m py_compile scripts/webgpu-probe	Pass
Workflow syntax	YAML parse and Python compile of setup heredoc	Pass; hosted workflow execution still pending
Native unit suite	./scripts/test linux-clang-development	20 registered tests, zero failures; 1 Wayland display test skipped
Native format/tidy	./scripts/check linux-clang-development --all	Pass
Native sanitizers	CI=true ./scripts/build linux-clang-asan-ubsan; ./scripts/test linux-clang-asan-ubsan	Pass; 18 registered tests, zero failures, 1 Wayland display test skipped
Installed SDK	./scripts/install-sdk linux-clang-development	Pass; external consumer links/runs and policy probe reports as expected
Diff hygiene	git diff --check	Pass

Native warning-as-error validation uses explicit CMake configuration:

out/host-tools/venv/bin/cmake --preset linux-clang-development -DLUDUS_WARNINGS_AS_ERRORS=ON
out/host-tools/venv/bin/cmake --build --preset linux-clang-development --parallel 2

The development compile database contains -Werror. Do not pass a -D flag to scripts/build: its extra arguments are build-command arguments, not configure arguments. That attempted invocation rejected the flag and was replaced by the explicit configure/build above.

Native socket/signal diagnostic tests and the SDK consumer initially failed inside the restricted execution sandbox. They passed when rerun with those OS facilities available. These restrictions were not worked around in engine code. System/third-party tidy warnings remain filtered by repository settings; no project warning suppression was added.

Browser/package observations

Browser: Codex In-app Browser in this Linux session; its exact version was not available through the browser inspection API. No experimental flags were added by the agent. No /dev/dri GPU device is exposed to the shell, and the browser adapter request returned no usable adapter. The separately listed Edge browser was unavailable to the control tool. No real-GPU rendering result is claimed.

Release artifact: out/packages/webgpu-probe-release.zip, approximately 180 KiB. SHA256 at this session's packaging: 0061d74d9002bb84e3d0b7acdc7d2e7df0d1f2ff2e0e1b155b7d4d3203b94a0c. ZIP timestamps are not normalized; rebuilding/repackaging can change its hash.

Scenario	Observation
Release app on localhost	Adapter callback displays “No WebGPU adapter is available”; data-state=failed, data-frames=0; no validation errors observed
Sandboxed iframe	Same readable adapter-unavailable result inside allow-scripts allow-same-origin harness
Extracted Release ZIP	HTML, JS, and wasm load from an independently extracted directory; same adapter callback result; no errors for that URL
Extracted iframe	Same result; demonstrates package paths work inside the local harness
JS missing	Shell displays “Could not load index.js. Check the package contents.”
wasm missing	Shell displays a request-failure message; console records wasm 404/instantiation failure as expected

The missing-file cases used temporary fixture directories under out/, not changes to the packaged artifact. Browser snapshots and status attributes were inspected through the browser tool. No screenshot artifact or hosted HTTPS/ itch.io evidence was collected. The shell's absent-WebGPU branch and the device request failure/loss branches are implemented but not yet runtime-validated.

Next session

1. Check out the PR branch, read AGENTS.md and the plan, and build/extract the Release ZIP using tools/webgpu-probe/README.md.
2. Run it in a real GPU-enabled browser without experimental flags. Record browser version, OS, adapter/backend, increasing frame count, visible color animation, and console/screenshot evidence. Exercise the device failure/loss cases with a suitable test harness. If the browser exposes no adapter, diagnose platform support rather than marking the rendering test passed.
3. Complete HTTPS iframe and an authorized itch.io draft embed check, recording embed settings and the tested artifact hash. No itch.io upload was authorized or performed by this PR session; the Release package is ready for that test.
4. Resolve any CI/probe issues, then update W0 acceptance. W1 remains gated on real-device local feasibility. Actual itch.io acceptance is mandatory for W8 even if hosted access is deferred while independent stages proceed.

Future RHI work must preserve callback lifetimes, avoid wgpuSurfacePresent, and keep normal logging separate from FoundationBase failures. W0 startup handles live until reload; W4 must implement explicit shutdown/cancellation.

Compatibility retry follow-up

The probe now explicitly requests core WebGPU, retries once with compatibility feature level on adapter failure, and shows bounded callback messages/statuses and browser version in the page. Device loss and uncaptured errors also retain the browser message. This does not establish a compatibility renderer policy for the eventual engine; feature/limit negotiation remains later-stage work.

Debug/Release builds and format/tidy checks passed for this follow-up. Local in-app Chromium 154 returned null for both adapter requests on the standalone page and sandboxed iframe; both failures appeared in Diagnostics. Real-GPU rendering and the actual itch.io embed remain pending with this updated package.

Temporary browser request stubs also verified exactly one core→compatibility retry, no retry after core adapter success, and propagation of a controlled device-request rejection plus the port's device-creation loss message. These fixtures were generated outside source/package and do not prove GPU rendering. Updated local Release ZIP SHA256: ceb6e6daaf4ed28bb22cfac01991c216c63194c44fc871c869fc7abb56a724bf.

User-hosted itch.io render, 2026-09-30

The user's updated screenshot shows the real itch.io draft embed reporting WebGPU is ready (core request) with a rendered purple clear frame and no visible WebGPU errors. It follows enabling Vulkan in Edge on the previously reported Linux/Wayland Intel UHD 620 laptop. The preceding screenshot showed null adapters in both core and compatibility mode. This is evidence of adapter, device, surface and at least one submitted frame in the actual embed; a still screenshot does not independently verify continuous animation. The successful configuration uses an experimental Vulkan flag, so the flag-free W0/W8 support gate remains open. There is no evidence requiring an external GPU.

The user then explicitly requested continuing implementation. Proceed with W1's independent CPU-only Foundation port using the validated toolchain, while retaining flag-free GPU support as an unresolved deployment gate.