name: ci

# Runs on every push to main and every PR. No user-controlled inputs
# interpolate into shell commands anywhere in this file — all `run:`
# blocks use static strings only.

on:
  push:
    branches: [main]
    # Explicitly *not* tags — tag pushes are the release workflow's
    # trigger. Leaving this list out would inherit the "all refs"
    # default and double-fire on `vX.Y.Z` tags that point at HEAD,
    # wasting a full CI run on a commit the release workflow is
    # already validating end-to-end.
    tags-ignore: ['**']
  pull_request:
  workflow_dispatch:

# One concurrency group per ref: if you push twice in a row (e.g.
# CHANGELOG typo fix on main), the first run cancels instead of
# keeping two redundant jobs alive.
concurrency:
  group: ci-${{ github.ref }}
  cancel-in-progress: true

permissions:
  contents: read

env:
  CARGO_TERM_COLOR: always

jobs:
  # ─────────────────────────────────────────────────────────────────────
  # Version sync gate. VERSION (repo root) is the canonical source of
  # truth; downstream files (payload config.h, Tauri conf, Cargo.toml,
  # package manifests) must all match it. Runs first because if this
  # fails there's no point burning cycles on clippy/tests for a release
  # that would ship with mismatched version strings on the payload
  # binary vs the desktop bundle.
  # ─────────────────────────────────────────────────────────────────────
  version-sync:
    name: version sync (VERSION is canonical)
    runs-on: ubuntu-24.04
    timeout-minutes: 5
    steps:
      - uses: actions/checkout@v7
      - uses: actions/setup-node@v7
        with:
          node-version: 22
      - name: Verify downstream files match VERSION
        run: node scripts/update-version.js --check

  # ─────────────────────────────────────────────────────────────────────
  # Rust workspace (engine + core + ftx2-proto). Format, clippy, tests.
  # Gates PRs that touch any Rust code.
  # ─────────────────────────────────────────────────────────────────────
  rust-engine:
    name: rust (engine workspace)
    runs-on: ubuntu-24.04
    timeout-minutes: 30
    steps:
      - uses: actions/checkout@v7

      - name: Install Rust stable
        uses: dtolnay/rust-toolchain@stable
        with:
          components: rustfmt, clippy

      - name: Cache cargo registry + target
        uses: actions/cache@v6
        with:
          path: |
            ~/.cargo/registry
            ~/.cargo/git
            engine/target
          key: ${{ runner.os }}-engine-cargo-${{ hashFiles('engine/Cargo.lock') }}
          restore-keys: ${{ runner.os }}-engine-cargo-

      - name: cargo fmt
        working-directory: engine
        run: cargo fmt --all -- --check

      - name: cargo clippy
        working-directory: engine
        run: cargo clippy --workspace -- -D warnings

      # The webui feature (rust-embed + static serving) is only enabled
      # in the Docker webui build, not the default workspace clippy/test
      # above. Compile-check it here so a broken #[cfg(feature = "webui")]
      # branch doesn't slip through CI and only surface at Docker build
      # time (as happened in 4.1.0). Full test/lint of the feature isn't
      # needed — a type-check catches the class of bug we're guarding.
      #
      # rust-embed's #[derive(RustEmbed)] #[folder = "webui/"] requires
      # the directory to exist at compile time — without it the macro
      # generates no impl and WebAssets::get fails (E0599). The real
      # content is the Vite build output (client/dist/) and is only
      # copied in by Dockerfile.webui; here we create a stub with a
      # dummy index.html so the derive macro is satisfied.
      - name: create webui stub for rust-embed
        working-directory: engine/crates/ps5upload-engine
        run: mkdir -p webui && echo '<!DOCTYPE html><title>stub</title>' > webui/index.html
      - name: cargo check (webui feature)
        working-directory: engine
        run: cargo check -p ps5upload-engine --features webui

      - name: cargo test
        working-directory: engine
        env:
          RUST_BACKTRACE: 1
        run: cargo test --workspace -- --nocapture

  # ─────────────────────────────────────────────────────────────────────
  # Engine/core target portability. This is intentionally separate from
  # the full release matrix: PRs get a fast compile-only signal that the
  # host-side engine code still type-checks for the OS/arch combinations
  # we ship, while publish.yml remains the authoritative bundle builder.
  # ─────────────────────────────────────────────────────────────────────
  rust-engine-targets:
    name: rust engine target (${{ matrix.target }})
    strategy:
      fail-fast: false
      matrix:
        include:
          - os: ubuntu-24.04
            target: x86_64-unknown-linux-gnu
          # Native arm64 runner — cross-compiling to aarch64 from x86_64
          # ubuntu-24.04 fails because blake3's NEON SIMD path needs
          # aarch64-linux-gnu-gcc (not preinstalled, and bringing it in
          # would mean threading TARGET_CC env vars through every step).
          # Matches publish.yml's arm64 build leg.
          - os: ubuntu-24.04-arm
            target: aarch64-unknown-linux-gnu
          - os: macos-14
            target: aarch64-apple-darwin
          - os: macos-14
            target: x86_64-apple-darwin
          - os: windows-2022
            target: x86_64-pc-windows-msvc
          - os: windows-2022
            target: aarch64-pc-windows-msvc
    runs-on: ${{ matrix.os }}
    timeout-minutes: 15
    steps:
      - uses: actions/checkout@v7

      - name: Install Rust stable
        uses: dtolnay/rust-toolchain@stable
        with:
          targets: ${{ matrix.target }}

      - name: cargo check core
        working-directory: engine
        run: cargo check -p ps5upload-core --target ${{ matrix.target }}

      - name: cargo check engine
        working-directory: engine
        run: cargo check -p ps5upload-engine --target ${{ matrix.target }}

  # ─────────────────────────────────────────────────────────────────────
  # Rust desktop shell (client/src-tauri). Uses its own workspace so
  # it doesn't share target/ with the engine. Tauri pulls heavy
  # system deps on Linux (webkit/gtk), so we cache aggressively.
  # ─────────────────────────────────────────────────────────────────────
  rust-desktop:
    name: rust desktop (${{ matrix.os }})
    strategy:
      fail-fast: false
      matrix:
        os: [ubuntu-24.04, macos-14, windows-2022]
    runs-on: ${{ matrix.os }}
    timeout-minutes: 45
    steps:
      - uses: actions/checkout@v7

      - name: Install Rust stable
        uses: dtolnay/rust-toolchain@stable
        with:
          components: rustfmt, clippy

      # Formatting is platform-independent, so gate it on one runner
      # rather than paying for it three times. This job previously
      # installed clippy ONLY, so client/src-tauri was never format-
      # gated at all while the engine workspace was — six files had
      # quietly drifted by 5.1.1.
      - name: cargo fmt
        if: matrix.os == 'ubuntu-24.04'
        working-directory: client/src-tauri
        run: cargo fmt --all -- --check

      # Bounded + retried for the same reason as publish.yml's copy: a
      # stalled apt mirror once burned a whole 45-minute job budget there.
      - name: Install Tauri Linux deps
        if: runner.os == 'Linux'
        timeout-minutes: 12
        run: |
          set -euo pipefail
          # apt has NO default network timeout: if a mirror stalls mid-fetch it
          # waits forever. That is exactly what sank the v5.3.3 publish — the
          # Azure mirror went quiet after "Ign: .../noble-updates/main amd64
          # Packages" and the step sat there for 45 minutes until the job
          # timeout killed it, skipping create-release. Acquire::*::Timeout is
          # the part that actually fixes it; the retry loop is only useful
          # because of it, since without a timeout every attempt would hang
          # the same way.
          apt_retry() {
            for attempt in 1 2 3; do
              if sudo DEBIAN_FRONTEND=noninteractive apt-get \
                   -o Acquire::Retries=3 \
                   -o Acquire::http::Timeout=20 \
                   -o Acquire::https::Timeout=20 \
                   "$@"; then
                return 0
              fi
              echo "apt attempt $attempt failed; retrying in 15s" >&2
              sleep 15
            done
            echo "apt failed after 3 attempts: apt-get $*" >&2
            return 1
          }
          apt_retry update
          apt_retry install -y \
            libgtk-3-dev \
            libwebkit2gtk-4.1-dev \
            libappindicator3-dev \
            librsvg2-dev \
            libsoup-3.0-dev \
            libjavascriptcoregtk-4.1-dev \
            patchelf \
            build-essential \
            clang \
            libclang-dev

      - name: Cache cargo target
        uses: actions/cache@v6
        with:
          path: |
            ~/.cargo/registry
            ~/.cargo/git
            client/src-tauri/target
          key: ${{ matrix.os }}-desktop-cargo-${{ hashFiles('client/src-tauri/Cargo.lock') }}
          restore-keys: ${{ matrix.os }}-desktop-cargo-

      # tauri.conf.json bundles two out-of-tree resources: the PS5
      # payload (ps5upload.elf.gz) and the engine sidecar binary.
      # cargo check parses those resource paths and errors out if
      # anything is missing, long before any code compiles. Stub the
      # payload .gz and build the engine binary for real so its target
      # path exists.
      #
      # The stub is a *fake ELF* (ELF magic + 200 KB of zeros, gzipped),
      # not an empty gzip: `cargo test` runs probes.rs::
      # embedded_payload_decompresses_to_elf, which gunzips the embedded
      # payload and asserts ELF magic + >100 KB. That test guards the
      # "Send payload" decode path (an empty stub once shipped on mobile
      # and broke it). The real ~1 MB payload is built + embedded by the
      # `payload` job here and by publish.yml for releases; this fake one
      # only has to satisfy the decode plumbing.
      - name: Stub payload resource
        shell: bash
        run: |
          mkdir -p payload
          { printf '\x7fELF'; head -c 204800 /dev/zero; } > payload/ps5upload.elf
          gzip -nf payload/ps5upload.elf   # -> payload/ps5upload.elf.gz

      - name: Build engine sidecar (needed as a Tauri resource)
        working-directory: engine
        run: cargo build --release -p ps5upload-engine

      - name: cargo check
        working-directory: client/src-tauri
        run: cargo check --all-targets

      - name: cargo clippy
        working-directory: client/src-tauri
        run: cargo clippy --all-targets -- -D warnings

      - name: cargo test
        working-directory: client/src-tauri
        env:
          RUST_BACKTRACE: 1
        run: cargo test

  # ─────────────────────────────────────────────────────────────────────
  # Client frontend — TypeScript typecheck, ESLint, Vitest, Vite build.
  # ─────────────────────────────────────────────────────────────────────
  client-frontend:
    name: client (ts + vite)
    runs-on: ubuntu-24.04
    timeout-minutes: 15
    steps:
      - uses: actions/checkout@v7

      - uses: actions/setup-node@v7
        with:
          node-version: 22
          cache: npm
          cache-dependency-path: client/package-lock.json

      - name: Install deps
        working-directory: client
        run: npm ci --no-audit --no-fund

      - name: TypeScript typecheck
        working-directory: client
        run: npm run typecheck

      - name: ESLint
        working-directory: client
        run: npm run lint

      - name: Vitest
        working-directory: client
        run: npm test

      # i18n coverage gate. validate-repo runs this locally and
      # TESTING.md elevates it to a release gate; CI was the missing
      # link. Static command, no event-derived input. Reviewed for
      # GitHub Actions injection risk: safe.
      - name: i18n coverage
        run: npm run i18n:check

      - name: Vite production build
        working-directory: client
        run: npm run build:vite

      - name: Install Playwright Chromium
        working-directory: client
        run: npx playwright install --with-deps chromium

      - name: Browser workflow smoke test
        working-directory: client
        run: npm run test:e2e

  # ─────────────────────────────────────────────────────────────────────
  # Repo scripts and hygiene. Runs the same parser/audit layer exposed by
  # `make lint-scripts` / `npm run scripts:check`.
  # ─────────────────────────────────────────────────────────────────────
  repo-scripts:
    name: repo scripts
    timeout-minutes: 10
    strategy:
      fail-fast: false
      matrix:
        os: [ubuntu-24.04, macos-14, windows-2022]
    runs-on: ${{ matrix.os }}
    steps:
      - uses: actions/checkout@v7
      - uses: actions/setup-node@v7
        with:
          node-version: 22

      - name: Check script syntax
        run: npm run scripts:check

      - name: Audit scripts
        # Strict mode: fail CI if any tracked script/test utility
        # has zero references AND isn't on the manualEntrypoints
        # allowlist. Previously this only printed warnings, so a
        # renamed-without-callers script silently slipped through.
        # Static command, no event-derived input.
        run: node scripts/audit-scripts.mjs --strict

  # ─────────────────────────────────────────────────────────────────────
  # Coverage reports. This is a reporting job, not the primary gate:
  # rust-engine and client-frontend above fail fast on correctness.
  # Artifacts contain HTML + lcov for trend tracking and local inspection.
  # ─────────────────────────────────────────────────────────────────────
  coverage:
    name: coverage reports
    runs-on: ubuntu-24.04
    # `continue-on-error` below does NOT cover the timeout: a job that hits
    # `timeout-minutes` is reported as *cancelled*, which cancels the whole
    # run and leaves main red over a job that is explicitly not a gate. It
    # reached 20m17s on 2026-09-16 against a 16m02s green run the day
    # before, so the headroom here is deliberate — keep it well clear of
    # the real runtime as the test suite grows.
    timeout-minutes: 40
    # Reporting job — not a gate. cargo-llvm-cov occasionally trails
    # rustc's profraw format version (we hit "raw profile version
    # mismatch: ... expected version = 9" when rustc bumped to v10),
    # and we don't want a coverage tool's release cadence to block
    # PRs on unrelated correctness work. The artifact is for trend
    # inspection, not enforcement.
    continue-on-error: true
    steps:
      - uses: actions/checkout@v7

      - uses: actions/setup-node@v7
        with:
          node-version: 22
          cache: npm
          cache-dependency-path: client/package-lock.json

      - name: Install client deps
        working-directory: client
        run: npm ci --no-audit --no-fund

      - name: Install Rust stable
        uses: dtolnay/rust-toolchain@stable
        with:
          components: llvm-tools-preview

      - name: Install cargo-llvm-cov
        uses: taiki-e/install-action@cargo-llvm-cov

      - name: Generate coverage
        run: npm run coverage

      - name: Upload coverage artifact
        uses: actions/upload-artifact@v7
        with:
          name: coverage-reports
          path: |
            coverage/
            client/coverage/
          retention-days: 14

  # ─────────────────────────────────────────────────────────────────────
  # Payload ELF — compile against the ps5-payload-sdk so we catch any
  # breakage to the C code. Doesn't upload artifacts; publish.yml does
  # that on tagged runs. Keeps PR feedback loop tight.
  # ─────────────────────────────────────────────────────────────────────
  payload:
    name: payload (ELF)
    runs-on: ubuntu-24.04
    timeout-minutes: 20
    steps:
      - uses: actions/checkout@v7

      - name: Install payload toolchain
        timeout-minutes: 12
        run: |
          set -euo pipefail
          # apt has NO default network timeout: if a mirror stalls mid-fetch it
          # waits forever. That is exactly what sank the v5.3.3 publish — the
          # Azure mirror went quiet after "Ign: .../noble-updates/main amd64
          # Packages" and the step sat there for 45 minutes until the job
          # timeout killed it, skipping create-release. Acquire::*::Timeout is
          # the part that actually fixes it; the retry loop is only useful
          # because of it, since without a timeout every attempt would hang
          # the same way.
          apt_retry() {
            for attempt in 1 2 3; do
              if sudo DEBIAN_FRONTEND=noninteractive apt-get \
                   -o Acquire::Retries=3 \
                   -o Acquire::http::Timeout=20 \
                   -o Acquire::https::Timeout=20 \
                   "$@"; then
                return 0
              fi
              echo "apt attempt $attempt failed; retrying in 15s" >&2
              sleep 15
            done
            echo "apt failed after 3 attempts: apt-get $*" >&2
            return 1
          }
          apt_retry update
          # Add LLVM apt repo for llvm-22 (not in Ubuntu 24.04 default repos)
          curl -fsSL --retry 3 --retry-delay 5 https://apt.llvm.org/llvm.sh | sudo bash -s -- 22
          apt_retry install -y clang-22 llvm-22-dev lld-22 unzip python3

      # The shared installer reads scripts/ps5-sdk.env, verifies the official
      # release checksum, and is also used by developer bootstrap scripts and
      # publish.yml. One pin therefore controls every payload build.
      - name: Install pinned PS5 Payload SDK
        run: sudo env PS5_SDK_INSTALL_DIR=/opt/ps5-payload-sdk bash scripts/install-ps5-sdk.sh

      - name: Build payload
        env:
          PS5_PAYLOAD_SDK: /opt/ps5-payload-sdk
          # SDK v0.42's prospero-llvm-config candidate search maxes at 21,
          # so point it at the CI-installed LLVM 22 explicitly.
          LLVM_CONFIG: /usr/bin/llvm-config-22
        run: make payload

      - name: Verify ELF
        run: |
          file payload/ps5upload.elf payload/installer/ps5upload-installer.elf
          test -s payload/ps5upload.elf
          test -s payload/ps5upload.elf.gz
          test -s payload/installer/ps5upload-installer.elf
          test -s payload/installer/ps5upload-installer.elf.gz
          gzip -t payload/ps5upload.elf.gz payload/installer/ps5upload-installer.elf.gz
          # Sanity-check the payload is a real build, not a zero-byte
          # stub. Current payload is ~600 KB (runtime + takeover +
          # register + blake3 SIMD backends); drop the floor to 200 KB
          # so legitimate shrinkage from future refactors doesn't false-
          # positive while still catching a broken/empty build.
          size_bytes=$(stat -c '%s' payload/ps5upload.elf)
          echo "ps5upload.elf: ${size_bytes} bytes"
          if [ "$size_bytes" -lt 200000 ]; then
            echo "::error::payload ELF is smaller than expected — build may have failed silently"
            exit 1
          fi
          dpi_size_bytes=$(stat -c '%s' payload/installer/ps5upload-installer.elf)
          echo "ps5upload-installer.elf: ${dpi_size_bytes} bytes"
          if [ "$dpi_size_bytes" -lt 50000 ]; then
            echo "::error::PS5Upload installer ELF is smaller than expected"
            exit 1
          fi

  # ─────────────────────────────────────────────────────────────────────
  # Android APK (Tauri mobile). Verifies the whole mobile pipeline builds
  # end-to-end: the engine linked in-process, the payload embedded, the
  # Tauri lib cross-compiled to aarch64-linux-android, and Gradle
  # assembling an installable debug APK. Uses the same fake-ELF payload
  # stub as rust-desktop — this is a build-verification job, not a
  # release. publish.yml builds the real, signed APK with the actual
  # payload. The debug APK is uploaded as an artifact for manual testing.
  # ─────────────────────────────────────────────────────────────────────
  android:
    name: android (apk build)
    runs-on: ubuntu-24.04
    timeout-minutes: 40
    steps:
      - uses: actions/checkout@v7

      - uses: actions/setup-node@v7
        with:
          node-version: 22
          cache: npm
          cache-dependency-path: client/package-lock.json

      - name: Setup JDK 17
        uses: actions/setup-java@v6
        with:
          distribution: temurin
          java-version: "17"

      - name: Install Rust + Android targets
        uses: dtolnay/rust-toolchain@stable
        with:
          targets: aarch64-linux-android,armv7-linux-androideabi

      - name: Cache cargo target
        uses: actions/cache@v6
        with:
          path: |
            ~/.cargo/registry
            ~/.cargo/git
            client/src-tauri/target
          key: ${{ runner.os }}-android-cargo-${{ hashFiles('client/src-tauri/Cargo.lock') }}
          restore-keys: ${{ runner.os }}-android-cargo-

      # Gradle's dependency cache.
      #
      # v5.4.5's CI failed here: Maven Central answered 403 for
      # kotlin-stdlib and Gradle could not configure the project, so the
      # build died before compiling anything of ours. Nothing was wrong
      # with the code -- the publish workflow built the same commit
      # minutes later.
      #
      # gen/android is scaffolded fresh every run, so the Gradle project
      # is new each time, but ~/.gradle/caches persists across runs when
      # cached. With the artifacts already local, a transient Maven
      # outage stops being a build failure.
      #
      # Keyed loosely on purpose: the generated project pins its own
      # versions, so any recent cache is a useful starting point and a
      # miss only costs a download.
      - name: Cache Gradle dependencies
        uses: actions/cache@v6
        with:
          path: |
            ~/.gradle/caches
            ~/.gradle/wrapper
          key: ${{ runner.os }}-android-gradle-${{ hashFiles('client/src-tauri/tauri.conf.json') }}
          restore-keys: ${{ runner.os }}-android-gradle-

      - name: Install client deps
        working-directory: client
        run: npm ci --no-audit --no-fund

      # Fake-ELF payload stub (see rust-desktop's note): satisfies the
      # embedded-payload decode + this build without the real ~1 MB ELF.
      - name: Stub payload resource
        shell: bash
        run: |
          mkdir -p payload
          { printf '\x7fELF'; head -c 204800 /dev/zero; } > payload/ps5upload.elf
          gzip -nf payload/ps5upload.elf

      # Pin the NDK to a known-good major version rather than blindly
      # taking whatever the runner image ships. An NDK rotation on the
      # runner image can silently break the build (new clang warnings →
      # errors, libc symbol changes, etc.). If the pinned major is absent
      # we fall back to the newest available and warn.
      - name: Resolve NDK
        shell: bash
        run: |
          # Preferred NDK major (r27 = NDK 27.x). Update when CI is
          # validated against a newer one.
          PREFERRED_MAJOR="27"
          ndk=""
          for candidate in \
            "$ANDROID_HOME/ndk/${PREFERRED_MAJOR}."* \
            "$ANDROID_HOME/ndk/"*; do
            if [ -d "$candidate" ]; then
              ndk="$candidate"
              break
            fi
          done
          if [ -z "$ndk" ]; then
            echo "::error::no NDK found under $ANDROID_HOME/ndk"
            exit 1
          fi
          # Warn if we couldn't get the preferred major.
          case "$(basename "$ndk")" in
            "${PREFERRED_MAJOR}"*) ;;
            *) echo "::warning::preferred NDK r${PREFERRED_MAJOR} not found, using $(basename "$ndk")" ;;
          esac
          echo "NDK_HOME=$ndk" >> "$GITHUB_ENV"
          echo "Using NDK: $ndk"

      - name: Scaffold Android project (gen/ is gitignored)
        working-directory: client
        run: npx tauri android init

      # `android init` seeds placeholder launcher icons; re-apply ours so
      # the Android app icon matches the desktop app icon.
      - name: Apply app icon
        working-directory: client
        run: npx tauri icon src-tauri/icons/sources/macos_squircle_1024.png

      # Keep the manifest patch in the build-verification path too, so a
      # regression in android-postinit-patch.sh is caught by CI (release
      # parity; the loopback + backup config is harmless for debug).
      - name: Patch manifest (loopback cleartext + backup rules)
        run: bash scripts/release/android-postinit-patch.sh

      - name: Build debug APK (arm64 + armv7)
        working-directory: client
        run: npx tauri android build --debug --apk --target aarch64 --target armv7

      - name: Upload debug APK
        uses: actions/upload-artifact@v7
        with:
          name: ps5upload-android-debug-apk
          path: client/src-tauri/gen/android/app/build/outputs/apk/**/*.apk
          if-no-files-found: error
          retention-days: 7

  # ─────────────────────────────────────────────────────────────────────
  # Aggregate gate. One stable check that turns green only when every
  # required job above succeeded. Branch protection on `main` requires
  # THIS single context ("PR gate") instead of each matrix job by name,
  # so adding/removing a matrix leg never silently drops it from the gate.
  #
  # `coverage` is deliberately excluded — it's continue-on-error (a
  # reporting job, not a correctness gate), so a flaky coverage tool must
  # not block a PR.
  #
  # `if: always()` is load-bearing: without it a FAILED dependency would
  # SKIP this job, and GitHub counts a skipped required check as passing —
  # letting a broken PR merge. With always() + the result check below, the
  # gate runs regardless and fails when any dependency didn't succeed.
  # ─────────────────────────────────────────────────────────────────────
  pr-gate:
    name: PR gate
    if: always()
    needs:
      - version-sync
      - rust-engine
      - rust-engine-targets
      - rust-desktop
      - client-frontend
      - repo-scripts
      - payload
      - android
    runs-on: ubuntu-24.04
    timeout-minutes: 5
    steps:
      - name: Require every gating job to have succeeded
        run: |
          results="${{ join(needs.*.result, ' ') }}"
          echo "dependency results: $results"
          for r in $results; do
            if [ "$r" != "success" ]; then
              echo "::error::a required CI job did not succeed (saw '$r') — PR is not mergeable"
              exit 1
            fi
          done
          echo "✓ all required CI jobs succeeded"