name: Deploy

on:
  push:
    branches: [master]
  workflow_dispatch:

# One deploy at a time; if you push twice quickly, cancel the older run so the
# newest commit is what lands on the server.
concurrency:
  group: deploy
  cancel-in-progress: true

jobs:
  build-deploy:
    # `flatpak` is a privileged runner registered to this repository alone
    # (/etc/forgejo-runner/config-catcrafts.yaml on the box, label
    # flatpak:docker://archlinux:latest): crafter-build is the Crafter.Build
    # Flatpak, and bubblewrap inside the job container needs the privilege.
    # The same runner carries the two deploy bind mounts (/deploy, /deploy-app)
    # and a docker volume at /var/lib/flatpak that keeps the installed Flatpak
    # and its Sdk between runs. Same route as imsd's package job; see
    # deploy/README.md "Runner configuration".
    runs-on: flatpak
    env:
      # Which crafter-build. `stable` is the branch Crafter.Build publishes
      # from its release tags, `beta` follows every push to its master (the
      # same thing the rolling `latest` tarball this job used to download
      # was). This site tracks Crafter.Graphics master, which since 2026-10-06
      # uses the Shader `defines`/`variant` fields that only Crafter.Build
      # master has: the 2.1.0 release predates them by three days, and run
      # 533 failed on exactly that. Flip to //stable once a Crafter.Build
      # release carries them. imsd pins stable; it has no Crafter.Graphics.
      CRAFTER_REF: net.catcrafts.CrafterBuild//beta
    steps:
      - name: Check live payment test credentials
        # ShouldSettleEurcOnTestnet is a mandatory deploy gate, so a missing
        # secret fails HERE, before any build time is spent — never as a silent
        # skip. The secret lives in Forgejo: repo Settings -> Actions ->
        # Secrets. See deploy/README.md "Live payment suites in CI" for what it
        # is and how to keep the testnet wallet funded.
        #
        # There is ONE live suite now, not two. The bank half used to be a
        # hosted provider with a test-mode API, and that gate died with the
        # account on 2026-08-20. Its replacement is a transfer to our own
        # account, which has no counterpart to call and therefore no credential
        # to check: ShouldSettleBankTransfers drives the real rail with no
        # network at all and runs unconditionally with the other black-box
        # suites. Losing a live gate here is a fact about the rail, not a
        # weakening of the standard.
        env:
          EURC_E2E_PRIVATE_KEY: ${{ secrets.EURC_E2E_PRIVATE_KEY }}
        run: |
          set -eu
          if [ -z "${EURC_E2E_PRIVATE_KEY:-}" ]; then
            echo "ERROR: EURC_E2E_PRIVATE_KEY secret is not set (an Ethereum Sepolia key holding testnet EURC + gas)." >&2
            exit 1
          fi
          echo "Live payment test credentials are present."

      - name: Install build dependencies
        run: |
          # Same keyring bootstrap the Crafter.Build CI does: the slim
          # archlinux:latest image ships without a populated pacman keyring
          # or local master key.
          pacman-key --init
          pacman-key --populate archlinux
          pacman -Sy --noconfirm --needed archlinux-keyring
          # nodejs is required for the JS-based actions (checkout, cache) to
          # run inside this archlinux container — the runner execs them with
          # node. This shell step needs no node, so installing it here (before
          # Checkout) is enough.
          # ffmpeg/ffprobe are both used by tools/fetch-media.sh. ffprobe reads
          # the pixel dimensions of each mirrored file, which become the
          # width/height attributes that stop the posts page reflowing as 5 MB
          # recordings arrive. ffmpeg transcodes each still image into the AVIF
          # and PNG renditions the page serves it between. The
          # ShouldServePostPages suite asserts both are present — so without
          # this package the deploy fails at the test gate rather than
          # shipping a janky page.
          #
          # Nothing that compiles is installed here. clang, lld, libc++ and the
          # WASI sysroot come with the crafter-build Flatpak (next step), and
          # its Sdk has the cmake, make and perl that build msquic and quictls
          # (Crafter.Network's ExternalBuilder::CMake dependency), the OpenSSL
          # the server links, and the gpg the suites call: the suites run
          # inside the sandbox, so what this container has is beside the point.
          pacman -Syu --noconfirm --needed \
            nodejs flatpak \
            git curl tar rsync zstd gzip jq ffmpeg
          # Container runs as root; workspace may be owned by another uid.
          git config --global --add safe.directory '*'

      - name: Install crafter-build
        # crafter-build is the Crafter.Build Flatpak, the route imsd's package
        # job takes: it carries its own clang, lld, libc++ and WASI sysroot, and
        # its Sdk (org.freedesktop.Sdk, from Flathub) has the cmake, make, perl,
        # git, gpg and OpenSSL the build and the suites need. The branch is
        # CRAFTER_REF at the top of the job. Crafter.Build's CI publishes the
        # Flatpak and nothing else since 2026-10-02: the rolling `latest`
        # release's crafter-build-linux-x86_64-v2.tar.gz this step used to
        # download is gone (404), which is what failed the deploy on
        # 2026-10-12.
        #
        # /var/lib/flatpak is a docker volume the runner mounts, so the Sdk and
        # the Flatpak (about 4 GB on disk) are downloaded once; `install` is a
        # no-op after that and `update` moves the installed branch forward.
        # The install is retried: a flaky download must not fail a deploy
        # before any build time is spent.
        run: |
          set -eu
          flatpak remote-add --if-not-exists flathub https://dl.flathub.org/repo/flathub.flatpakrepo
          flatpak remote-add --if-not-exists catcrafts https://flatpak.catcrafts.net/catcrafts.flatpakrepo
          for i in 1 2 3; do
            flatpak install -y --noninteractive catcrafts "$CRAFTER_REF" && break
            if [ "$i" -eq 3 ]; then
              echo "ERROR: installing the crafter-build Flatpak failed 3 times." >&2
              exit 1
            fi
            echo "install failed (attempt $i/3), retrying in 15s..." >&2
            sleep 15
          done
          flatpak update -y --noninteractive
          # The Flatpak under the name the rest of this file uses. Inside the
          # sandbox PATH is /app/bin:/usr/bin (the Flatpak's tools, the Sdk's);
          # /opt/foundry/bin is appended because the suites run INSIDE the
          # sandbox and ShouldSettleEurcOnTestnet shells out to cast, which
          # the next step installs there: the sandbox sees the host's /opt but
          # not its /usr/local.
          cat > /usr/local/bin/crafter-build <<EOF
          #!/bin/sh
          exec flatpak run --env=PATH=/app/bin:/usr/bin:/opt/foundry/bin $CRAFTER_REF "\$@"
          EOF
          # A binary crafter-build builds for the host links the Flatpak's
          # libc++ (and the Sdk's glibc, newer than this container's), so it
          # runs inside the sandbox, not in this shell: `crafter-run <binary>
          # <args>` is how the steps below call catcrafts-server. The host
          # itself gets those libraries shipped next to the binary, see
          # "Deploy backend".
          cat > /usr/local/bin/crafter-run <<EOF
          #!/bin/sh
          cmd=\$1; shift
          case \$cmd in /*) ;; *) cmd=\$PWD/\$cmd ;; esac
          exec flatpak run --env=PATH=/app/bin:/usr/bin:/opt/foundry/bin --command="\$cmd" $CRAFTER_REF "\$@"
          EOF
          chmod +x /usr/local/bin/crafter-build /usr/local/bin/crafter-run
          flatpak info "$CRAFTER_REF" | sed -n 's/^ *\(Version\|Commit\|Date\|Branch\): */crafter-build \1: /p'
          crafter-build --version
          cat /usr/local/bin/crafter-build /usr/local/bin/crafter-run

      - name: Install cast (Foundry)
        # The ShouldSettleEurcOnTestnet gate pays its own test order with a
        # real ERC-20 transfer on Base Sepolia, and cast is the only tool a
        # stock shell has that can sign one — the server deliberately cannot,
        # that is the EURC rail's design. Extracted alone: forge/anvil/chisel
        # are dead weight here.
        #
        # Into /opt, not /usr/local/bin: the suite runs inside the crafter-build
        # sandbox, which sees the host's /opt but has the Sdk's own /usr. The
        # crafter-build wrapper above puts /opt/foundry/bin on the sandbox PATH;
        # the last line checks that is where the suite will find it.
        run: |
          set -eux
          url="https://github.com/foundry-rs/foundry/releases/download/stable/foundry_stable_linux_amd64.tar.gz"
          curl -fsSL "$url" -o /tmp/foundry.tar.gz
          mkdir -p /opt/foundry/bin
          tar -xzf /tmp/foundry.tar.gz -C /opt/foundry/bin cast
          crafter-run /usr/bin/sh -c 'command -v cast && cast --version'

      - name: Checkout
        uses: actions/checkout@v4

      - name: Cache crafter-build dependency clones
        # crafter-build's cache holds the Crafter.Graphics and Crafter.Network
        # clones, the prebuilt module cache and the msquic/quictls external
        # build. crafter-build still git-pulls each dep every run, so a stale
        # cache only means a smaller delta fetch, never a stale build.
        #
        # The path is Flatpak's: it sets XDG_CACHE_HOME to the app's own
        # directory under ~/.var/app, crafter-build honors that, and the
        # sandbox shares the host's home, so the directory is the same inside
        # and out. The "Deploy backend" step reads libmsquic out of the same
        # tree.
        uses: actions/cache@v4
        with:
          path: ~/.var/app/net.catcrafts.CrafterBuild/cache/crafter.build
          key: crafter-cache-${{ runner.os }}-flatpak-${{ hashFiles('project.cpp') }}
          restore-keys: |
            crafter-cache-${{ runner.os }}-flatpak-

      - name: Fetch ECB reference rates
        # Feeds the indicative national-currency line on order pages. Every
        # charge is in euros; this is display only, labelled with its date —
        # which is why build-time freshness is enough and no rate service is
        # ever called at page-view time. Exits 0 on failure: a stale rate
        # (or none — the page then shows only euros) must not fail a deploy.
        run: tools/fetch-rates.sh

      - name: Fetch fediverse posts
        # Build-time, not run-time: the site hosts the owner's own posts in full
        # and links out for the discussion, so there is no sync service and no
        # runtime dependency on the instance being up. The script leaves the
        # committed content/posts.json untouched and exits 0 on any failure, so a
        # fediverse outage cannot fail a deploy.
        run: tools/fetch-posts.sh

      - name: Mirror post media
        # Downloads the images and screen recordings the posts carry — both the
        # headline file and everything embedded inside the body — and rewrites
        # content/posts.json (including the body text itself) to point at our own
        # copies, so nothing the browser loads is third-party. That is what keeps
        # the privacy notice's "everything comes from catcrafts.net" true.
        #
        # Content-addressed: the name is the hash of the bytes. Note what that
        # does NOT mean — a third-party file is re-fetched on every build, because
        # the name cannot be known until the bytes are in hand; only the write is
        # skipped when the hash is already on the mount. Media we host ourselves
        # (tools/publish-media.sh) is the exception that is genuinely incremental:
        # fetch-media.sh adopts an own-origin URL without any request at all.
        #
        # Writes straight into the mount so the copies persist across deploys —
        # they are NOT always reproducible, because a source instance deleting a
        # file leaves ours as the only one.
        id: media
        run: |
          set -eu
          if [ -d /deploy-app ]; then
            mkdir -p /deploy-app/media
            dir=/deploy-app/media
          else
            echo "WARNING: /deploy-app not mounted; mirroring to a throwaway dir." >&2
            echo "Media will be re-downloaded on every build until the mount exists." >&2
            dir=media
          fi
          tools/fetch-media.sh "$dir"
          # Handed to the test step (E2E_MEDIA_DIR) so ShouldServePostPages
          # checks the files where they actually are. It used to assume
          # ./media and reported every referenced file as missing here, which
          # reads as a broken site and means a wrong path.
          echo "dir=$dir" >> "$GITHUB_OUTPUT"

      - name: Build and test the backend
        id: srv
        # The server product builds Catcrafts.Shared for the host, which is the
        # only way to actually RUN the code that generates every byte of markup
        # the site emits. `crafter-build test` is the gate: if escaping or the
        # JSON reader regress, the deploy stops here rather than shipping
        # broken pages. One suite per tests/<Name>/main.cpp — the unit suites
        # replaced the binary's --selftest flag, the Should{ServeRoutes,…}
        # black-box suites replaced tools/e2e.sh: each spawns the freshly
        # built server on its own scratch port with the fake payment rails and
        # exercises it over real HTTP. ShouldBootWasmAtDepth self-skips here —
        # no wasm bundle exists yet — and re-runs after the wasm build below.
        #
        # E2E_MEDIA_DIR points ShouldServePostPages at the mirror mount, where
        # the media files actually are.
        #
        # EURC_E2E_PRIVATE_KEY feeds the live crypto suite, which runs as part
        # of this same gate: a deploy that cannot settle a testnet EURC
        # transfer does not ship. If it fails at `cast send`, the testnet
        # wallet is empty — see deploy/README.md "Live payment suites in CI"
        # for the faucets. The bank rail's own suite needs no secret; it drives
        # the real rail with a credits file instead of a provider.
        #
        # Same refuse-to-guess rule as the wasm bundle below: a variant
        # directory embeds a config hash, so more than one match means the tree
        # is ambiguous and picking the first would deploy an arbitrary build.
        #
        # The binary is called through crafter-run: it links the Flatpak's
        # libc++, which this container does not have.
        env:
          E2E_MEDIA_DIR: ${{ steps.media.outputs.dir }}
          EURC_E2E_PRIVATE_KEY: ${{ secrets.EURC_E2E_PRIVATE_KEY }}
        run: |
          set -eux
          crafter-build --product=server
          crafter-build test --product=server
          matches=$(find bin -maxdepth 1 -type d -name 'Catcrafts.Server-*' | sort)
          count=$(printf '%s\n' "$matches" | grep -c . || true)
          if [ "$count" -ne 1 ]; then
            echo "Expected exactly one Catcrafts.Server-* directory, found $count:" >&2
            printf '%s\n' "$matches" >&2
            exit 1
          fi
          echo "srv=$matches" >> "$GITHUB_OUTPUT"
          crafter-run "$matches/catcrafts-server" --routes

      - name: Generate sitemap and Atom feed
        # Both come from the same route table and Post model the pages use, so
        # they cannot drift from what the site serves. Generated BEFORE the wasm
        # build so cfg.files picks them up into the bundle.
        env:
          SRV: ${{ steps.srv.outputs.srv }}
        run: |
          set -eux
          crafter-run "$SRV/catcrafts-server" --sitemap > sitemap.xml
          crafter-run "$SRV/catcrafts-server" --feed    > feed.xml
          head -n 4 sitemap.xml

      - name: Build (wasm bundle)
        run: crafter-build

      - name: Locate build output
        id: out
        run: |
          set -eu
          # The directory name embeds a config hash, so glob for it. Any change
          # to compile/link flags produces a NEW hash, which is why we refuse to
          # guess when more than one variant is present rather than taking
          # whichever the filesystem happened to list first.
          matches=$(find bin -maxdepth 1 -type d -name 'Catcrafts.Net-wasm32-wasip1-*' | sort)
          count=$(printf '%s\n' "$matches" | grep -c . || true)
          if [ "$count" -eq 0 ]; then
            echo "No build output directory found under bin/" >&2
            ls -la bin || true
            exit 1
          fi
          if [ "$count" -gt 1 ]; then
            echo "Ambiguous build output — $count variant directories under bin/:" >&2
            printf '%s\n' "$matches" >&2
            echo "Refusing to guess which one to deploy. Clean bin/ and rebuild." >&2
            exit 1
          fi
          dist=$matches
          echo "dist=$dist" >> "$GITHUB_OUTPUT"
          echo "Built bundle: $dist"
          ls -la "$dist"

      - name: Make the static shell depth-safe
        # Caddy serves this index.html directly when the backend is down, at
        # whatever URL was requested — including two-segment ones like
        # /demos/raytracer. Crafter.Build emits relative boot scripts and its
        # runtime.js fetches variants.json/files.json/the wasm relative to the
        # DOCUMENT, so at any depth the fallback loads nothing at all. The script
        # roots the tags and adds <base href="/">, and fails loudly rather than
        # silently no-opping. The SSR path handles itself; this is only the
        # backend-down fallback.
        env:
          DIST: ${{ steps.out.outputs.dist }}
        run: tools/fix-bundle-depth.sh "$DIST"

      - name: Wasm boot and script-shape checks
        # The bundle-sensitive slice of the black-box suites, re-run now that
        # the wasm bundle exists under bin/. The server discovers the bundle
        # and lifts its <script> tags from it, so before the wasm build there
        # is nothing to assert about wasm booting: ShouldBootWasmAtDepth
        # self-skipped in the backend step, and ShouldStayScriptFree's checks
        # around the boot tags passed vacuously because no page had scripts to
        # begin with. Re-running both against the real bundle is what makes
        # them meaningful.
        #
        # A gate, not a report: a failure here stops the deploy.
        run: crafter-build test ShouldBootWasmAtDepth ShouldStayScriptFree --product=server

      - name: Pre-compress static assets
        # Caddy's `precompressed zstd gzip` (see deploy/Caddyfile.example)
        # serves these siblings straight from disk instead of re-compressing
        # the ~700 KB wasm module on every request. Building them here also
        # buys a better ratio than on-the-fly encoding would spend CPU on.
        env:
          DIST: ${{ steps.out.outputs.dist }}
        run: |
          set -eu
          for f in "$DIST"/*.wasm "$DIST"/*.js "$DIST"/*.css "$DIST"/*.xml "$DIST"/*.svg; do
            [ -f "$f" ] || continue
            zstd -19 -q -f -k -- "$f"
            gzip -9 -f -k -- "$f"
          done
          echo "Compressed artifacts:"
          ls -la "$DIST"

      - name: Deploy to web root
        # No SSH: the job already runs on the deploy box. The server's web root
        # is bind-mounted into this container at /deploy via the runner config
        # (container.options: "-v /path/to/webroot:/deploy"). We just copy into
        # it. --delete makes the root an exact mirror of the build; Caddyfile.coi
        # is a local dev helper (its 'root' points at the build machine), so the
        # server's own Caddy config supplies the headers instead.
        env:
          DIST: ${{ steps.out.outputs.dist }}
        run: |
          set -eu
          if [ ! -d /deploy ]; then
            echo "ERROR: /deploy is not mounted into the runner container." >&2
            echo "Add '-v /your/webroot:/deploy' to the runner's container.options" >&2
            echo "in config.yaml and restart the runner." >&2
            exit 1
          fi
          rsync -a --delete --exclude 'Caddyfile.coi' "$DIST"/ /deploy/
          echo "Deployed $DIST -> /deploy (host web root)"

      - name: Guard against leaking runtime files into the web root
        # The web root is served by Caddy's file_server AND mirrored with
        # --delete, so anything runtime-owned that lands there is both published
        # and destroyed on the next push. Nothing does today; this is here so a
        # future cfg.files line cannot quietly change that once the shop has a
        # database and bank credentials.
        run: |
          set -eu
          # The inverse guard: everything that IS meant to be served must be
          # world-readable. Caddy and catcrafts-server run as their own users,
          # so a 0600 file (mktemp's default, once shipped by fetch-posts.sh)
          # deploys "successfully" and then 403s / renders an empty page.
          # (/deploy-app/content gets the same check in its own deploy step,
          # after its rsync — here it would still hold the previous run.)
          unreadable=$(find /deploy -type f ! -perm -o+r -print 2>/dev/null || true)
          if [ -n "$unreadable" ]; then
            echo "ERROR: deployed files not world-readable (host services cannot serve them):" >&2
            echo "$unreadable" >&2
            exit 1
          fi
          bad=$(find /deploy -maxdepth 1 \( -name '*.db' -o -name '*.db-*' \
                -o -name '*.pem' -o -name '*.key' -o -name '*.env' \
                -o -name '.*' ! -name '.' \) -print 2>/dev/null || true)
          if [ -n "$bad" ]; then
            echo "ERROR: runtime-owned files found in the public web root:" >&2
            printf '%s\n' "$bad" >&2
            exit 1
          fi
          echo "Web root contains no database, key or dotfile."

      - name: Deploy backend
        # Second bind mount, separate from the web root on purpose: the backend
        # binary and its content must NOT be inside a directory Caddy serves and
        # rsync --delete mirrors. Add to the runner's container.options:
        #   -v /srv/catcrafts-app:/deploy-app
        #
        # The restart is not done here. This job runs inside a container with no
        # access to the host's systemd, and the alternatives (host root for the
        # runner, or a polkit rule) grant CI far more than "restart one service"
        # needs. Instead the last step touches a marker file that
        # catcrafts-deploy.path watches — see deploy/catcrafts-deploy.path.
        # That unit smoke-checks the new binary (--render /) before cutting
        # over, so a broken deploy leaves the working server running.
        env:
          SRV: ${{ steps.srv.outputs.srv }}
        run: |
          set -eu
          if [ ! -d /deploy-app ]; then
            echo "ERROR: /deploy-app is not mounted into the runner container." >&2
            echo "Add '-v /srv/catcrafts-app:/deploy-app' to the runner's" >&2
            echo "container.options in config.yaml and restart the runner." >&2
            exit 1
          fi
          install -m 0755 "$SRV/catcrafts-server" /deploy-app/catcrafts-server.new
          mkdir -p /deploy-app/content
          # --delete on content/ is safe: it is regenerated every build. Note it
          # does NOT touch /deploy-app/media, which must survive.
          rsync -a --delete content/ /deploy-app/content/
          # The server links libmsquic as a SHARED lib out of crafter-build's
          # external cache; the RUNPATH into that cache only exists on this
          # build machine. The binary carries an rpath to /srv/catcrafts-app
          # (project.cpp), so the .so must travel with it. install dereferences
          # the libmsquic.so.2 -> .so.2.x.y symlink into a regular file.
          MSQ=$(find ~/.var/app/net.catcrafts.CrafterBuild/cache/crafter.build/external -name 'libmsquic.so.2*' -type f | head -1)
          if [ -z "$MSQ" ]; then
            echo "ERROR: libmsquic.so.2 not found in the external build cache." >&2
            exit 1
          fi
          install -m 0755 "$MSQ" /deploy-app/libmsquic.so.2.new
          mv -f /deploy-app/libmsquic.so.2.new /deploy-app/libmsquic.so.2
          # The binary also links the Flatpak's libc++, libc++abi and LLVM
          # libunwind: its rpath into /app/lib exists only inside the sandbox,
          # the host's own libc++ is an older release than the headers the
          # binary was compiled against, and the host has no libunwind.so.1 on
          # the loader path at all. They travel with the binary the way
          # libmsquic does and the same rpath finds them. Nothing in them needs
          # more than GLIBC_2.38 from the host's glibc (checked with objdump -T
          # on 2026-10-12; the Sdk's own glibc is newer than the host's, which
          # is why it matters).
          APP=$(flatpak info --show-location "$CRAFTER_REF")/files/lib
          for lib in libc++.so.1 libc++abi.so.1 libunwind.so.1; do
            if [ ! -f "$APP/$lib" ]; then
              echo "ERROR: $lib not found in the crafter-build Flatpak under $APP." >&2
              exit 1
            fi
            install -m 0755 "$APP/$lib" "/deploy-app/$lib.new"
            mv -f "/deploy-app/$lib.new" "/deploy-app/$lib"
          done
          # Same world-readable guard as the web root: the server runs as its
          # own user, and a 0600 content file renders as an empty page.
          unreadable=$(find /deploy-app/content -type f ! -perm -o+r -print 2>/dev/null || true)
          if [ -n "$unreadable" ]; then
            echo "ERROR: deployed content not world-readable:" >&2
            echo "$unreadable" >&2
            exit 1
          fi
          # Swap the binary into place atomically, so a request arriving mid-copy
          # never hits a truncated executable.
          mv -f /deploy-app/catcrafts-server.new /deploy-app/catcrafts-server
          date -u +%FT%TZ > /deploy-app/.deploy-stamp
          echo "Deployed backend -> /deploy-app; marker touched"