open-standards/.github/workflows/wai-conformance.yml

name: WAI conformance

# WAI's load-bearing claim is bit-exact reconstruction on every machine,
# no tolerance parameter. This workflow enforces it mechanically: every
# cargo-class corpus — and the one neural decoder that can make the same
# promise, the integer-synthesis path (i64-only, no float ONNX) — is
# re-derived on a four-platform matrix (x86-64 + arm64 across Linux, macOS,
# Windows) and must reproduce its frozen BLAKE3 hashes; and the integer
# byte-equality goldens (rANS / int_synth / JPEG-AI mANS, generated once on
# a reference machine) must decode identically everywhere. All of it is pure
# Rust (no C codec chain, no ONNX Runtime), so the matrix needs no system
# dependencies.

on:
  push:
    branches: [main]
    paths:
      - 'wai/wai-rs/**'
      # wai-rs path-depends on wai-quantum, and the quantum corpora below are
      # verified through it: a change there changes what this workflow checks.
      - 'wai/wai-quantum/**'
      - 'wai/*-conformance/**'
      - 'wai/byte-exact-conformance/**'
      - '.github/workflows/wai-conformance.yml'
  pull_request:
    paths:
      - 'wai/wai-rs/**'
      # wai-rs path-depends on wai-quantum, and the quantum corpora below are
      # verified through it: a change there changes what this workflow checks.
      - 'wai/wai-quantum/**'
      - 'wai/*-conformance/**'
      - 'wai/byte-exact-conformance/**'
      - '.github/workflows/wai-conformance.yml'

# No global `shell:` default. The Windows runner is a host executor whose
# `bash` is the WSL stub (a different OS), so shell-specific steps below pick
# their shell per-OS: bash on unix, pwsh on Windows. Bare `cargo …` steps take
# each runner's native default, which runs cargo identically either way.

jobs:
  cargo-classes:
    name: cargo-class corpora · ${{ matrix.os }}
    runs-on: ${{ matrix.os }}
    strategy:
      fail-fast: false
      matrix:
        # Self-hosted fleet labels. Both architectures across all three OS
        # families: x86_64 by linux + windows, arm64 by linux-arm64 + macos —
        # what matters for the Fx integer floor is the arch, not the OS. No
        # Intel-Mac runner exists in the fleet, and it would add no arch
        # coverage anyway.
        os: [linux, linux-arm64, macos, windows]
    steps:
      - uses: actions/checkout@v5

      # unix: the docker runners (linux / linux-arm64) ship no Rust, so rustup
      # is bootstrapped; the macOS `host` runner has it but launchd's PATH omits
      # ~/.cargo/bin. dtolnay/rust-toolchain is unusable — it calls a `rustup`
      # the docker image lacks.
      - name: Install Rust 1.98.0 (unix)
        if: matrix.os != 'windows'
        shell: bash
        run: |
          export PATH="$HOME/.cargo/bin:$PATH"
          command -v rustup >/dev/null 2>&1 || \
            curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y --profile minimal --default-toolchain 1.98.0
          rustup toolchain install 1.98.0 --profile minimal
          rustup default 1.98.0
          rustup component add --toolchain 1.98.0 clippy
          echo "$HOME/.cargo/bin" >> "$GITHUB_PATH"

      # Windows: a host executor with rustup already installed. Only ensure the
      # pinned toolchain is present + default. pwsh, never bash (that is WSL here).
      - name: Install Rust 1.98.0 (Windows)
        if: matrix.os == 'windows'
        shell: pwsh
        run: |
          rustup toolchain install 1.98.0 --profile minimal
          rustup default 1.98.0

      - name: Build the cargo classes (pure Rust, no C codec chain)
        working-directory: wai/wai-rs
        # neural_int joins the pure-Rust set: the integer-synthesis decoder is
        # i64-only (no ONNX, no float), so it carries the same bit-exact promise
        # and is checkable on this matrix with no ONNX Runtime present.
        run: cargo build --no-default-features --features world,audio_scene,haptic,splat4d,score,provenance,neural_int,film
      - name: Library unit tests (the fixed-point determinism floors)
        working-directory: wai/wai-rs
        run: cargo test --lib --no-default-features --features world,audio_scene,haptic,splat4d,score,provenance,neural_int,film,oer,receipt_claims
      # The JPEG AI exact decoder reproduces across platforms only if no platform math
      # library is on its path. Two guards, each failing the step: clippy's
      # disallowed-methods (wai-rs/clippy.toml), which each module on the decode path
      # denies, rejects a listed float method however it is spelled; and
      # tests/jpegai_decode_symbols.rs lists the linked decoder's symbols with `nm` and
      # rejects any C math-library function it imports, whatever made the call, and any
      # it defines other than the round-to-integral ones (see its doc). `-A clippy::all`
      # keeps the output to that guard. Unix legs only: `nm` is not on the Windows
      # runner, and for clippy the unix legs already cover both architectures.
      - name: JPEG AI exact decode calls no platform math library (unix)
        if: matrix.os != 'windows'
        shell: bash
        working-directory: wai/wai-rs
        run: |
          set -euo pipefail
          cargo clippy --lib --no-default-features --features neural_int -- -A clippy::all
          cargo test --no-default-features --features world,audio_scene,haptic,splat4d,score,provenance,neural_int,film --test jpegai_decode_symbols

      # The integer decoders run fast kernels (src/int_kernels.rs) that must give
      # the same integer as the scalar reference kernels for every output element.
      # The feature list above leaves out audio_int, video_int, world_int and
      # splat_int, so without these two steps int_audio's kernel-equivalence tests
      # and tests/int_kernels_vectors.rs (every committed integer-decoder vector
      # decoded under the reference and under each fast policy, against its frozen
      # hash) would run nowhere. `--lib` also carries the other kernels' equivalence
      # modules; tests/int_kernels_geometry.rs runs edge geometry (strides up to
      # usize::MAX, padding far beyond the kernel) through the public decoders
      # under every policy. Both profiles, because the overflow-parity tests assert
      # different things in each: with overflow checks on (dev) that the fast path panics
      # where the reference panics, with them off (release) that it returns the
      # same integers. All pure Rust, so every leg runs them.
      # tests/int_mlicpp.rs decodes the committed wai.neural.int_mlicpp vector against
      # its pinned hashes (image, every scale bucket, every symbol) under every kernel
      # policy and thread count, and checks that damaged payloads are refused rather than
      # overflowing — which is why it runs in both profiles too.
      # tests/int_mlicpp_preflight.rs (its own binary: it installs a counting global
      # allocator) checks that a header the model cannot fit is refused before the
      # decode allocates anything sized from it.
      - name: Fast integer kernels ≡ reference kernels (overflow checks on)
        working-directory: wai/wai-rs
        run: cargo test --no-default-features --features neural_int,video_int,audio_int,world_int,splat_int --lib --test int_kernels_vectors --test int_kernels_geometry --test int_mlicpp --test int_mlicpp_preflight
      - name: Fast integer kernels ≡ reference kernels (release, overflow checks off)
        working-directory: wai/wai-rs
        run: cargo test --release --no-default-features --features neural_int,video_int,audio_int,world_int,splat_int --lib --test int_kernels_vectors --test int_kernels_geometry --test int_mlicpp --test int_mlicpp_preflight
      - name: Carriage conformance — every registered capability round-trips byte-exact
        working-directory: wai/wai-rs
        run: cargo test --no-default-features --test carriage

      # The corpus-verify loop is genuinely shell-specific (a helper that groups
      # log output and propagates the first non-zero exit), so it ships once per
      # shell family. The command it runs is identical; only the wrapper differs.
      - name: Every cargo-class corpus reproduces its frozen hashes (unix)
        if: matrix.os != 'windows'
        shell: bash
        working-directory: wai/wai-rs
        run: |
          set -euo pipefail
          F=world,audio_scene,haptic,splat4d,score,provenance,neural_int,film,quantum,quantum_ops,oer,receipt_claims
          v() { echo "::group::$1"; cargo run -q --no-default-features --features "$F" --bin "$2" -- verify "$3"; echo "::endgroup::"; }
          v worlds        wai_world_vectors      ../world-conformance
          v spatial-audio wai_audio_vectors      ../audio-conformance
          v haptics       wai_haptic_vectors     ../haptic-conformance
          v state-feeds   wai_feed_vectors       ../feed-conformance
          v avatars       wai_avatar_vectors     ../avatar-conformance
          v volumetric    wai_splat4d_vectors    ../volumetric-conformance
          v score         wai_score_vectors      ../score-conformance
          v int-synth     wai_int_synth_vectors  ../neural-conformance
          v film          wai_film_vectors       ../film-conformance
          v quantum       wai_quantum_vectors    ../quantum-conformance
          v quantum-ops   wai_quantum_ops_vectors ../quantum-ops-conformance
          v oer           wai_oer_vectors        ../oer-conformance
          v claims        wai_claim_vectors      ../claim-conformance
      # OER's second implementation is portable C, not Rust. Rebuild it from source and
      # require its output to match the pinned corpus byte for byte: the kernel is
      # deterministic, and wai_oer_vectors above has already verified the pinned copy,
      # so equality carries that verification to what the source produces today.
      # `cc` with the Makefile's flags, so the step needs no `make` on the runner.
      - name: OER C kernel reproduces its pinned receipts (unix)
        if: matrix.os != 'windows'
        shell: bash
        working-directory: wai/oer-conformance/settle
        run: |
          set -euo pipefail
          cc -O2 -std=c99 -Wall -Wextra -Wno-unused-parameter -Wno-sign-compare \
            -o "$RUNNER_TEMP/settle_host" host_main.c settle.c sha256.c tweetnacl.c -lm
          "$RUNNER_TEMP/settle_host" "$RUNNER_TEMP/firmware_vectors.json" 400
          cmp "$RUNNER_TEMP/firmware_vectors.json" ../firmware_vectors.json
      - name: Every cargo-class corpus reproduces its frozen hashes (Windows)
        if: matrix.os == 'windows'
        shell: pwsh
        working-directory: wai/wai-rs
        run: |
          $F = "world,audio_scene,haptic,splat4d,score,provenance,neural_int,film,quantum,quantum_ops,oer,receipt_claims"
          function v($grp, $bin, $path) {
            Write-Output "::group::$grp"
            cargo run -q --no-default-features --features $F --bin $bin -- verify $path
            if ($LASTEXITCODE -ne 0) { exit $LASTEXITCODE }
            Write-Output "::endgroup::"
          }
          v worlds        wai_world_vectors      ../world-conformance
          v spatial-audio wai_audio_vectors      ../audio-conformance
          v haptics       wai_haptic_vectors     ../haptic-conformance
          v state-feeds   wai_feed_vectors       ../feed-conformance
          v avatars       wai_avatar_vectors     ../avatar-conformance
          v volumetric    wai_splat4d_vectors    ../volumetric-conformance
          v score         wai_score_vectors      ../score-conformance
          v int-synth     wai_int_synth_vectors  ../neural-conformance
          v film          wai_film_vectors       ../film-conformance
          v quantum       wai_quantum_vectors    ../quantum-conformance
          v quantum-ops   wai_quantum_ops_vectors ../quantum-ops-conformance
          v oer           wai_oer_vectors        ../oer-conformance
          v claims        wai_claim_vectors      ../claim-conformance

  byte-exact:
    name: cross-arch byte-equality · ${{ matrix.os }}
    runs-on: ${{ matrix.os }}
    strategy:
      fail-fast: false
      matrix:
        # Self-hosted fleet labels. Both architectures across all three OS
        # families: x86_64 by linux + windows, arm64 by linux-arm64 + macos —
        # what matters for the Fx integer floor is the arch, not the OS. No
        # Intel-Mac runner exists in the fleet, and it would add no arch
        # coverage anyway.
        os: [linux, linux-arm64, macos, windows]
    steps:
      - uses: actions/checkout@v5

      - name: Install Rust 1.98.0 (unix)
        if: matrix.os != 'windows'
        shell: bash
        run: |
          export PATH="$HOME/.cargo/bin:$PATH"
          command -v rustup >/dev/null 2>&1 || \
            curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y --profile minimal --default-toolchain 1.98.0
          rustup toolchain install 1.98.0 --profile minimal
          rustup default 1.98.0
          echo "$HOME/.cargo/bin" >> "$GITHUB_PATH"
      - name: Install Rust 1.98.0 (Windows)
        if: matrix.os == 'windows'
        shell: pwsh
        run: |
          rustup toolchain install 1.98.0 --profile minimal
          rustup default 1.98.0

      - name: Integer rANS + int_synth + JPEG-AI mANS goldens (one machine → every machine)
        working-directory: wai/byte-exact-conformance
        run: cargo test --release

  # The C2PA emitter is gated behind an off-by-default feature, so the four-platform
  # job above — which builds a fixed feature list — never compiles it. A feature no
  # workflow enables has no CI coverage at all: the suite passes, and the tests in it
  # simply do not run.
  #
  # One platform, deliberately. Everything the matrix above exists to prove is about
  # byte-identical reconstruction across architectures; a signed manifest makes no
  # such claim and would not become more true for being emitted on four machines. What
  # this job proves is that the manifest is still well-formed, still carries the three
  # assertions, and still validates — which is architecture-independent.
  c2pa-emitter:
    name: C2PA manifest emission · signed, read back, validated
    runs-on: linux
    steps:
      - uses: actions/checkout@v4

      - name: Install Rust 1.98.0
        shell: bash
        run: |
          export PATH="$HOME/.cargo/bin:$PATH"
          command -v rustup >/dev/null 2>&1 || \
            curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y --profile minimal --default-toolchain 1.98.0
          rustup toolchain install 1.98.0 --profile minimal
          rustup default 1.98.0
          echo "$HOME/.cargo/bin" >> "$GITHUB_PATH"

      # `rust_native_crypto` rather than the SDK's default openssl backend, so this
      # job needs no system libraries — the same property the matrix above depends on.
      # Asserted, not assumed: an openssl-linked build would still pass the tests here
      # while quietly making the crate undistributable to a bare runner.
      - name: The emitter pulls in no system crypto
        working-directory: wai/wai-rs
        shell: bash
        run: |
          for dep in openssl openssl-sys; do
            if cargo tree --no-default-features --features c2pa_emit -i "$dep" >/dev/null 2>&1; then
              echo "::error::$dep reached the c2pa_emit dependency graph"
              exit 1
            fi
          done
          echo "no system crypto in the graph"

      - name: Emit, sign, read back, validate
        working-directory: wai/wai-rs
        run: cargo test --lib --no-default-features --features c2pa_emit c2pa_emit

      # The feature must stay optional. If `provenance` ever pulls c2pa in, every
      # consumer that wanted only the binding types inherits a large dependency tree
      # they did not ask for, and the "off by default" claim in the extension is false.
      - name: The feature stays optional
        working-directory: wai/wai-rs
        shell: bash
        run: |
          if cargo tree --no-default-features --features provenance -i c2pa >/dev/null 2>&1; then
            echo "::error::c2pa is reachable from the provenance feature; it must stay opt-in"
            exit 1
          fi
          cargo build --no-default-features --features provenance
          echo "provenance builds without c2pa"