verify.yml

# The whole quality gate, run automatically. Until this file existed, `pnpm run
# verify` only ran when a human remembered to (or at deploy time, which can be
# days after the breaking merge) - see issue #54.
#
# `verify` no longer runs as one job, because the test phase is ~95% of it and
# a runner is ~3x slower than a dev machine: the single job measured 9m03s on
# PR #116. The phases are now split across a `static` job and a sharded `tests`
# matrix that run in parallel, with a `verify` job aggregating them.
#
# The jobs here are `static`, `tests` (4 shards), `rust`, `verify` and `build`.
# Only `verify` and `build` are REQUIRED status checks in ruleset `EJ`. `rust`
# is deliberately not one, and its absence from the ruleset is not an oversight
# to fix: `verify` asserts `needs.rust.result`, so a red `rust` job turns the
# required check red anyway, with no ruleset PUT and no two-step. Every
# required NAME is a permanent liability - rename it and every PR blocks
# forever on a check that cannot run - so the aggregator absorbing new phases
# is the cheaper shape. Add future phases the same way.
#
# The anti-drift rule that motivated the old VERBATIM comment still holds, and
# is now enforced differently: this file names only package.json SCRIPTS
# (`verify:static`, `verify:shard`), never the underlying commands. `verify`,
# `verify:static` and `verify:shard` are all composed from the same
# `verify:lint` script, so there is still exactly one definition of each phase
# and CI cannot drift from local. Do not inline `vp check` etc. here.
# Note it is never `check` - that is `vp check --fix` and CI must never rewrite
# files.
#
# WHY THE JOB NAMED `verify` STILL EXISTS: ruleset `EJ` requires status checks
# named `verify` and `build`. A required check that never appears blocks every
# PR forever, so the aggregating job below MUST keep the id `verify`. Renaming
# it needs a ruleset PUT in the same change - see CLAUDE.md on the two-step.
#
# WHY 4 SHARDS, MEASURED ON CI (issue #119, 2026-08-10). This said 3 until then,
# on the reasoning that `test/previewAgreement.spec.ts` was "67s of the 68s local
# suite" and therefore an unsplittable floor that N=4+ could not beat. Both
# halves of that are now false, and the second one was never measurable locally:
# a dev box has 12 cores, a runner has 4, so locally the CPU term is absorbed and
# only the file floor is visible.
#
# One run, all four arms concurrently on the same runner pool (18 jobs), so the
# comparison is within-run rather than against a different hour. Slowest shard is
# what the gate waits on:
#
#   N=3   test-step walls 452 / 356 / 102          slowest job  488s
#   N=4                   306 / 291 / 291 / 72     slowest job  333s   <- adopted
#   N=5                   311 / 213 / 174 / 74 / 67            343s
#   N=6                   314 / 224 / 162 / 140 / 61 / 51      344s
#
# N=4 is -155s of gate wall (-32%) for +76 runner-seconds (+7.6%), NOT the +33%
# the local numbers predicted - total CPU is flat and only ~28s of per-job setup
# is added. N=5 and N=6 do not improve on N=4 at all (+5s, +8s: noise), so 4 is
# the point of diminishing return.
#
# The plateau at ~310s is NOT the previewAgreement file. That was checked rather
# than assumed: the slowest N=6 shard (33 files, 314s) does not contain it - the
# same shard run locally lists cliffOreCascade / vulcanusCliffBands /
# cliffCollisionResidualShape as its heavy files. The premise expired because the
# suite grew 171 -> 201 files through the #84 cliff work: previewAgreement is now
# 72.9s of 503s of total per-file wall (14.5%), and TEN files exceed 20s. So
# splitting that one file is no longer the lever it was; the binding shard is
# whichever one vitest's hash-split happens to load with several heavy files.
# Re-measure before raising N again - and re-measure on CI, not locally.
#
# The production build runs too, in the separate `build` job below - see its
# own comment for why it is not a phase inside `verify`.
#
# What is deliberately NOT here:
#   - `pnpm refs:sync`. It needs a Factorio binary and ~/GitHub/factorio-data,
#     neither of which exists on a runner. `verify` is designed to pass with no
#     Factorio installed and that property is what makes this workflow possible.
#   - Any deploy step or credential. Cloudflare Pages does not build this repo
#     (`deploy:app` uploads an already-built `dist`), so CI here is a check
#     only and the job needs no secrets at all.
name: verify

on:
  push:
    branches: [main]
  pull_request:
  # So a run can be forced without an empty commit.
  workflow_dispatch:

# Superseded pushes to the same PR are pointless at ~4 minutes a run, so those
# get cancelled. Pushes to `main` do NOT - `cancel-in-progress` there means a
# second merge kills the first one's run, and `main` is the branch the deploy
# ships from, so every commit on it should carry its own verdict.
#
# This is not hypothetical: on 2026-07-30 six PRs merged inside half an hour and
# the runs for `605a4cc` and `be5f592` were both cancelled by the merges that
# followed them. Neither commit is unverified in practice - each had a green
# `verify` on its own PR before the ruleset would let it merge - but "green on
# the PR" and "green as it sits on main" are different claims, and only the
# second one survives a rebase-free merge into a `main` that has since moved.
# Ruleset `EJ` has since set `strict_required_status_checks_policy: true`, which
# closes most of that gap by forcing a PR up to date before it can merge; a
# per-commit verdict on `main` is still the thing that proves it, so these runs
# stay uncancelled.
concurrency:
  group: verify-${{ github.workflow }}-${{ github.ref }}
  cancel-in-progress: ${{ github.event_name == 'pull_request' }}

# Minimum scope: the job only reads the tree. It writes no statuses, comments,
# packages or releases, so nothing beyond `contents: read` is granted.
permissions:
  contents: read

jobs:
  # Everything in `verify` that is NOT the app test suite: `vp check`,
  # `check:vue`, and `preview:test`. 55-80s of job wall over the last six runs
  # (this said "~24s" and was never re-measured; ~28s of that is checkout +
  # install, which every job here pays). Still finishes long before any shard,
  # so it costs nothing to run as its own job.
  static:
    runs-on: ubuntu-latest
    timeout-minutes: 15
    env:
      # `preview:test` shells out to wrangler (`wrangler types --check`). Keep it
      # from phoning home for telemetry on a runner where nobody can answer the
      # opt-in prompt.
      WRANGLER_SEND_METRICS: "false"
    steps:
      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
      - uses: pnpm/action-setup@0977fd99725f1db4007ccb2928dbb4e90d06cc86 # v6.0.10
      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
        with:
          node-version-file: .node-version
          cache: pnpm
      - run: pnpm install --frozen-lockfile
      - run: pnpm run verify:static

  # The app test suite, split across runners. Each shard is a separate machine,
  # which is the ONLY way to raise the parallelism ceiling - `maxWorkers` was
  # already measured as a dead end on one box (see CLAUDE.md).
  tests:
    runs-on: ubuntu-latest
    # The slowest shard measures ~5m30s (see the shard table above). 15 minutes
    # stays a hang detector rather than a budget.
    timeout-minutes: 15
    strategy:
      # One failing shard must not cancel the others: a red `verify` should say
      # everything that is broken, not just whichever shard tripped first.
      fail-fast: false
      matrix:
        shardIndex: [1, 2, 3, 4]
        shardTotal: [4]
    steps:
      # Third-party actions are pinned to a full commit SHA, never a moving tag.
      # The trailing comment names the release each SHA is, and Renovate updates
      # the SHA and that comment together - see .github/renovate.json5.
      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1

      # No `version:` input on purpose. pnpm/action-setup >= 6 reads
      # `devEngines.packageManager` from package.json, so the pnpm pin stays in
      # exactly one place. Hard-coding it here would be a second pin to drift.
      # This step must precede setup-node: `cache: pnpm` below resolves the
      # store path by running pnpm, so pnpm has to be on PATH already.
      - uses: pnpm/action-setup@0977fd99725f1db4007ccb2928dbb4e90d06cc86 # v6.0.10

      # `.node-version` (26.7.0) is the single source of truth for the Node
      # version, and this is its first real consumer - the file went from
      # documentation to machinery when this workflow landed. `engines.node`
      # stays a permissive floor and is deliberately NOT what CI runs.
      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
        with:
          node-version-file: .node-version
          cache: pnpm

      # A full workspace install. Anything narrower can leave a sibling
      # workspace's symlink dangling and produce fake
      # `TS2307: Cannot find module 'vitest'` errors; a real full install is the
      # one that reports `Scope: all 3 workspace projects`.
      - run: pnpm install --frozen-lockfile

      - run: pnpm run verify:shard -- --shard=${{ matrix.shardIndex }}/${{ matrix.shardTotal }}

  # The Rust half of the gate: cargo fmt, clippy, tests, the zero-dependency
  # assertion, and the byte comparison against the committed engine.wasm.
  #
  # It is NOT a required status check of its own, and that is the point. Ruleset
  # `EJ` matches required checks by NAME, so every name added is a name that
  # blocks every PR forever if it is ever renamed or removed. The `verify` job
  # below already exists to aggregate, so asserting one more `needs.*.result`
  # there gets identical blocking behaviour with no ruleset PUT and no two-step.
  # See CLAUDE.md for when a genuinely new required name is warranted.
  #
  # No toolchain action, pinned or otherwise: `rust-toolchain.toml` pins 1.97.1
  # with rustfmt, clippy and the wasm32 target, and rustup installs that on the
  # first cargo command. Expect a `syncing channel updates` line in the log even
  # though the image already ships Cargo 1.97.1 - rustup treats `1.97.1-<host>`
  # and `stable-<host>` as separate installs of the same compiler, so the pin
  # costs one download per runner.
  #
  # Measured on this job's first run (#230): `info: syncing channel updates for
  # 1.97.1-x86_64-unknown-linux-gnu`, `info: downloading 6 components`, 10s of
  # the job's 19s. The whole job is still the cheapest in the workflow, and the
  # runner reports `rustc 1.97.1 (8bab26f4f 2026-07-14)` - the same build as the
  # dev machine, which is what makes the byte comparison below possible at all.
  rust:
    runs-on: ubuntu-latest
    # The whole job is seconds of compiling two tiny crates on top of the
    # toolchain sync. 15 minutes is a hang detector, matching the jobs above.
    timeout-minutes: 15
    steps:
      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1

      # Printed rather than assumed: if a byte comparison against the committed
      # engine.wasm ever fails, the first question is whether the compiler
      # moved, and that question needs an answer in the log of the run that
      # failed - not a re-run months later against a different image.
      - name: Show the resolved toolchain
        run: |
          rustup show
          cargo --version
          rustc --version --verbose

      # A pinned release binary rather than `cargo install`, which builds from
      # source and would dominate a job that is otherwise seconds - it took
      # ~4 minutes locally. The checksum is this asset's own published sha256,
      # so the binary is pinned by content the same way the actions above are
      # pinned by commit SHA.
      - name: Install cargo-deny 0.20.2
        run: |
          V=0.20.2
          F=cargo-deny-$V-x86_64-unknown-linux-musl.tar.gz
          curl -fsSLO "https://github.com/EmbarkStudios/cargo-deny/releases/download/$V/$F"
          echo "$CARGO_DENY_SHA256  $F" | sha256sum -c -
          tar -xzf "$F"
          sudo mv "cargo-deny-$V-x86_64-unknown-linux-musl/cargo-deny" /usr/local/bin/
          cargo deny --version
        env:
          CARGO_DENY_SHA256: 9f12ed4c49936e09b48bf862b595cde2fe64fcbd9d74dfacac6131ca824c8d5f

      # The one deviation from "always name a package.json script", and it does
      # not reopen the drift the rule guards against: `verify:rust` is exactly
      # `bash scripts/verify-rust.sh`, so the SCRIPT FILE is the single
      # definition and both sides run the same bytes of it. Going through pnpm
      # would add a pnpm/action-setup + setup-node + `pnpm install` (~28s) to a
      # job that needs no JavaScript at all, purely to shell out to this line.
      # If `verify:rust` ever grows a second command, this must become
      # `pnpm run verify:rust` with the setup steps restored.
      - run: bash scripts/verify-rust.sh

  # The required status check. Ruleset `EJ` requires a check named `verify`, so
  # this job must keep that id even though the work now happens above.
  #
  # `!cancelled()` rather than `always()`: a cancelled run (superseded push, per
  # the concurrency group) should stay cancelled, not be converted into a
  # failure. But a FAILED dependency must fail this job - a `needs:` job whose
  # dependency failed is otherwise skipped, and a skipped required check does
  # not block a merge. Hence the explicit result assertions.
  verify:
    needs: [static, tests, rust]
    if: ${{ !cancelled() }}
    runs-on: ubuntu-latest
    timeout-minutes: 5
    steps:
      - name: Assert every phase passed
        env:
          STATIC: ${{ needs.static.result }}
          TESTS: ${{ needs.tests.result }}
          RUST: ${{ needs.rust.result }}
        run: |
          echo "static=$STATIC tests=$TESTS rust=$RUST"
          # `tests` is the matrix aggregate: 'success' only if every shard was.
          [ "$STATIC" = "success" ] || { echo "::error::static phase: $STATIC"; exit 1; }
          [ "$TESTS" = "success" ] || { echo "::error::test shards: $TESTS"; exit 1; }
          [ "$RUST" = "success" ] || { echo "::error::rust phase: $RUST"; exit 1; }

  # `verify` is check + type-check + tests. None of them build, so a change that
  # passes all three and breaks the production build reached `main` unnoticed
  # until somebody deployed - possibly days later, and landing on whoever was
  # deploying rather than whoever broke it (issue #61). This job closes that.
  #
  # Deliberately a SEPARATE job rather than an extra phase inside `verify`:
  # `deploy` already runs `pnpm build` immediately after `pnpm run verify`, so
  # folding the build into `verify` would build twice on every deploy and slow
  # the local gate people actually run by hand. Separate also means it runs in
  # parallel with the test job instead of after it.
  build:
    runs-on: ubuntu-latest
    # The build is ~1s of Rolldown after ~30s of setup. 15 minutes is a hang
    # detector, matching the job above.
    timeout-minutes: 15
    steps:
      # Default `fetch-depth: 1` is correct here, and that was MEASURED rather
      # than assumed - #61 was filed believing the build needed deeper history.
      # `scripts/buildStamp.ts` runs exactly three git commands:
      # `rev-parse HEAD`, `rev-parse --short HEAD`, and `status --porcelain`.
      # None of them read history, so a shallow checkout is enough and the
      # earlier "the build stamp reads git history" was imprecise - it reads git
      # *state*.
      #
      # Worth knowing rather than fixing: on a `pull_request` event checkout
      # lands on the merge commit, so the stamp this job produces is a synthetic
      # SHA that exists nowhere in the repo. Harmless, because CI never deploys
      # its artifact - `deploy:app` builds locally and uploads that. If a CI
      # build ever becomes the thing that ships, this needs revisiting.
      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1

      - uses: pnpm/action-setup@0977fd99725f1db4007ccb2928dbb4e90d06cc86 # v6.0.10

      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
        with:
          node-version-file: .node-version
          cache: pnpm

      - run: pnpm install --frozen-lockfile

      # This catches a build that FAILS. It does not enforce "zero warnings",
      # which CLAUDE.md leans on ("anything that does appear is new and worth
      # reading"). Both ways of enforcing that were considered and rejected:
      #
      #   - Grepping the build output for /warn/i. `vp build` has no
      #     `--fail-on-warn` (checked against vp 0.2.6 `--help`), and this repo
      #     has already been burned once by grepping build output - a version
      #     grep returned zero because the minifier had rewritten the string.
      #     A false pass here is worse than no check.
      #   - A `build.rollupOptions.onLog` hook that throws on `warn`. Robust, but
      #     it would hard-fail every local build too the moment a dependency
      #     emits one benign warning, and dependency-sourced warnings are exactly
      #     what this repo has seen (zlib-asm needed two suppressions before #46
      #     removed it). The value is in reading a new warning, not in blocking on
      #     it.
      #
      # So warnings stay visible in this job's log and unenforced. If that ever
      # needs to change, the `onLog` route is the sturdy one - not a grep.
      - run: pnpm vp build