validate.yml

name: validate

on:
  pull_request:
    # mcpp.toml and index.toml carry the workspace member list, the inherited
    # [indices] redirect and the client version floor — a change to any of them
    # can break every member, so they gate the run like the descriptors do.
    paths: ["pkgs/**/*.lua", "tests/**", "README.md", "README.zh-CN.md", "mcpp.toml", "index.toml", ".github/workflows/validate.yml"]
  push:
    branches: [main]
  schedule:
    # WEEKLY full regression — every workspace member, regardless of diff.
    #
    # This is the safety net for the one thing selective testing structurally
    # cannot see: two changes that are each green on their own and break each
    # other once both have landed. Neither PR's diff names the member that
    # breaks, so neither PR selects it.
    #
    # Weekly rather than nightly because a full run is no longer cheap — with
    # the second linux toolchain leg it is ~11 hours of runner time — and the
    # net catches the same interactions whether it is cast every day or every
    # seven. Sunday 06:00 UTC.
    - cron: "0 6 * * 0"
  workflow_dispatch:
    inputs:
      cache:
        description: "Package build cache — 'local' rebuilds every dependency per member, which is what the timing table should be read against when comparing"
        type: choice
        options: [global, local]
        default: global

env:
  # 2026.8.5.4 carries two things this workflow depends on:
  #   .5.1  `tools = [...]` — how a consumer asks for a dependency's
  #         `kind = "bin"` target, which tests/examples/protobuf-protoc is built
  #         on. Before it: "tools must be a string, inline dep table, or nested
  #         table".
  #   .5.4  windows links with lld. link.exe caps a response-file LINE at
  #         128 KiB and opencv-module / opencv-module-dnn went past it —
  #             fatal error LNK1170: line in command file contains 135135 or
  #             more characters
  #         after 795s / 1166s of compiling. .5.3 newline-separated OUR response
  #         file, which was necessary but not sufficient: clang, acting as the
  #         driver, writes a SECOND one for the linker that we do not control.
  #         lld's response-file parser has no per-line limit at all.
  #         Together with re-enabling the global package cache below, this is
  #         what makes a green FULL run possible again: .5.3 removes the
  #         windows link failure, the cache removes the 150-minute timeout.
  #
  # Neither of them moves index.toml's min_mcpp: exposing compat.protobuf's `protoc`
  # target is additive, and 2026.8.3.3 still parses that descriptor with an
  # empty unknown_keys. The floor an index publishes decides whether older
  # clients keep working at all (mcpp#349), so it moves only when a descriptor
  # genuinely stops being readable — which is not the case here.
  # 2026.8.3.1: on macOS, a global object that touches std::cout during static
  # init crashes on sight (mcpp#336). Mach-O has no priority-ordered init
  # section and libc++'s <iostream> carries no ios_base::Init guard of its own,
  # so the streams are still all-zero when an archive member's initializer
  # runs. mcpp now links a generated object FIRST whose constructor brings them
  # up. It is not fixable package-side — std::ios_base::Init is only
  # forward-declared in libc++'s <ios> — so boost-ext.ut genuinely requires
  # this floor on macOS, and min_mcpp/latest_mcpp move with the pin as they
  # always have (a consumer below the floor would get a segfault with no
  # diagnostic, which is worse than E0006).
  # 2026.8.3.3 is the pin rather than .3.1: .3.2/.3.3 are cross-compilation
  # fixes (PE artifact naming, -static host-vs-target) that no leg of this
  # matrix exercises, so taking the newest of the train costs nothing.
  # 0.0.109: a bare dependency's wire address takes BOTH halves from the
  # descriptor the identity gate accepted (mcpp#286). This is the client-side
  # other half of the SPEC-001 migration below: mcpp used to take the NAME from
  # the descriptor and the NAMESPACE from the request, so a bare `gtest =
  # "1.15.2"` addressed `mcpplibs:gtest` — a key no index has. It only ever
  # worked because the pre-migration literal `package.name` read
  # "compat.gtest", which the hardcoded `compat.<short>` retry then caught.
  # Short names removed that coincidence and left every bare request against
  # this index broken on 0.0.108, which is why min_mcpp moves to 0.0.109.
  # Note this index cannot cover that spelling itself: the `[indices]` redirect
  # is keyed by the REQUEST's namespace, so a bare dependency resolves from the
  # published remote index rather than the checkout under test. Bare-name
  # resolution against a short-name index is upstream's e2e 165; what moving
  # the floor buys here is that consumers of THIS index get a client that can
  # address it.
  # 0.0.106: SPEC-001 package identity (mcpp#280). `package.name` is a SINGLE
  # ATOMIC SEGMENT — all hierarchy lives in `package.namespace` — and mcpp
  # addresses a package by the LITERAL name it read, so descriptors no longer
  # repeat their namespace inside `name`. This index is migrated to the short
  # form, which is why min_mcpp/latest_mcpp move in lock-step: an older client
  # re-derives `<ns>.<short>`, misses, and reports a bare E_NOT_FOUND. Bundles
  # xlings 0.4.69, which keys its index by (namespace, name) so two packages
  # sharing a short name in one index are both addressable (xlings#381) — this
  # index now has three such pairs (imgui / ffmpeg / lua under compat vs the
  # default namespace).
  # 0.0.102: windows command-line ceiling (mcpp#261 — the clang scan rule got
  # its P1689 JSON through shell redirection, which forced a `cmd /c` wrapper
  # and with it cmd.exe's 8191-char limit; clang-scan-deps -o removes both, and
  # $local_includes-carrying rules now fall back to response files). The pin
  # matters here because a package consumed FROM the registry sits under a
  # ~124-char xpkgs path instead of its own ~23-char checkout, which is what
  # pushed the vendored-opencv scan command over the line. Also: purview-include
  # depfile tracking extended to Clang (#257 — stale BMI reuse), OS-conditional
  # `[build].flags` (#258), and per-OS splices keyed on the resolved target
  # rather than the host (#254).
  # 0.0.101: per-feature per-glob flags + per-OS features (mcpp#253) — what
  # lets opencv select its dnn gemm backend per platform.
  # 0.0.99: feature dep/feat forwarding (mcpp#243 — a feature can open a
  # feature OF a dependency, e.g. opencv `dnn` forwarding compat.opencv/dnn);
  # vendored xlings 0.4.67 for the >=2 index_repo install fix (mcpp#238 /
  # openxlings/xlings#374); build.mcpp compiled program named `.exe` on
  # Windows (mcpp#230 secondary surface, after the 0.0.96 scanner crash fix).
  # 0.0.98: closes the obj-path disambiguation follow-ups that gated the
  # source-build compat.opencv unification — #240 (link inputs now follow
  # the disambiguated object names, so a dependency + consumer sharing a
  # source basename like `src/main.cpp` no longer 'obj/main.o missing') and
  # #239 (absolute/`..` dep-generated source paths sanitized component-wise
  # so objects stay under obj/). Also: `MCPP_DEP_<NAME>_DIR` build.mcpp
  # contract (#241), consumer-side `default-features = false` (#242), and a
  # loud unknown-mcpp-key warning with did-you-mean (#237, replaces the
  # silent-ignore at build time). Carried from 0.0.97: default-namespace
  # index redirect (`[indices] default = { path }`), which turned the public
  # module packages (imgui/ffmpeg/opencv/tinyhttps) into ordinary workspace
  # members and retired the per-package reseeding smoke shells + their
  # dedicated jobs; synchronous nasm bootstrap (mcpp#232 — the `mcpp index
  # update` pre-step is gone), obj-path disambiguation (#233), spacey-defines
  # quoting (#234), purview-include depfile tracking (#235). Older floors of
  # note: 0.0.96 fixed the windows scanner symlink-escape crash (mcpp#230);
  # 0.0.94 fixed feature-gated `sources` under `mcpp test` (mcpp#218); 0.0.91
  # added standard = "c++fly" to the resolver grammar, so c++fly descriptors
  # get the lint WARN below, not a hard grammar-parse rejection.
  # 2026.8.6.1 是本次身份迁移的**前置**,不是顺手升级。
  #
  # `Fetcher::install_path(ns, shortName, version)` 的 legacy 扫描此前匹配任何以
  # `-x-<shortName>` 结尾的目录、不看命名空间,于是查 `ocornut:imgui@1.92.8` 会拿到
  # `compat-x-imgui/1.92.8` —— 另一个仅仅短名相同的包。它一直够不到,是因为 module
  # 层用打包计数(imgui@0.0.6)而 compat 用上游版本(compat.imgui@1.92.8),版本永远
  # 不撞;本次把 module 层对齐到上游之后它们重合了。mcpp#364 修掉了它。
  #
  # index.toml 的 min_mcpp **不动**。两个会撞的 compat 邻居都是 Form B,所以旧客户端
  # 撞上时是响亮报错而不是静默用错包;下限是一道让整个索引对旧客户端失效的闸门
  # (mcpp#349),只该在描述符真的读不动时抬。这里读得动,差的是解析得对。
  #
  # 2026.8.8.2 曾经是本 pin 的**前置**,不是顺手升级:
  #
  # `compat.glx-runtime` 依赖 mesa,mesa 声明 `xim:glibc@>=2.38`;下限被任何更高
  # 版本满足,于是安装图形栈会在既有 glibc 旁边**再装一个**。2026.8.8.2 之前的
  # mcpp 用 `readdir` 的第一项来解析「那个 glibc payload」,编译侧与产物的
  # interpreter 因此可以指向不同版本 —— 这正是 #179 落地后 `asio-module` 和
  # `core` 变红、并导致整份改动被 #180 revert 的原因。用早于 2026.8.8.2 的 mcpp
  # 重新落地这份改动,就是在仍会犯这个错的引擎上复现事故条件;
  # tests/check_graphics_install_side_effects.sh 也需要它 —— 该测试在更旧的 mcpp
  # 上失败是**有意的**。
  #
  # 2026.8.10.1 又加了一层:2026.8.8.2 的 registry 无法解析
  # `xim:libglvnd@>=1.7.0.1` —— compat.glx-runtime@2026.08.08 在 linux 上的硬依赖,
  # 任何经 compat.glfw 的 X11 profile 触及图形栈的包(eui-neo 等)都会踩到,导致
  # linux 上安装即失败(main 的 graphics 检查因此持续变红)。
  #
  # 补充根因:这不是索引的数据缺陷,是**内带 xlings 的版本**。四段版本参与范围比较
  # 由 xlings `2026.8.9.2` 修好(N 段 semver 重写);mcpp `2026.8.8.2` 内带的是
  # `2026.8.8.1`,正好落在修复之前,所以它把一个存在的包报成 `not found`。
  # `2026.8.10.1` 起内带的 xlings 已含该修复。**`xim-pkgindex` 一个字都不用改。**
  #
  # 取 `2026.8.10.3` 而不是 `.1`,是因为它另外带两个直接影响本仓的修复:
  #   - mcpp#405 依赖 BMI 缓存命中时,被恢复的包传递依赖的 `std` 没进构建图 ——
  #     **每个图形项目的第二次构建必挂**(消费方自己不 `import std` 时现形)。
  #   - mcpp#407 `mcpp build` 会重放 `mcpp test` 留下的构建图。
  MCPP_VERSION: "2026.8.10.3"

jobs:
  lint:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - name: Install lua
        run: sudo apt-get install -y --no-install-recommends lua5.4
      - name: Lint package descriptors
        run: |
          fail=0
          for f in pkgs/*/*.lua; do
            # 1. Lua syntax check — load (= compile) without executing.
            #    `loadfile(name, 't')` rejects bytecode and parses text only.
            if ! lua5.4 -e "assert(loadfile('$f', 't'))" >/dev/null 2>&1; then
              echo "::error file=$f::lua syntax error"
              fail=1
            fi
            # 2. xpkg V1 baseline: the file has to populate `package = { ... }`
            #    with at least `spec`, `name`, and an `xpm` table. Form A vs
            #    Form B (mcpp = "<path>" / mcpp = { ... }) is descriptor-author
            #    choice and not enforced here.
            for needle in 'spec *=' 'name *=' 'xpm *='; do
              if ! grep -q "$needle" "$f"; then
                echo "::error file=$f::missing required field ($needle)"
                fail=1
              fi
            done
            # 3. Package version identifiers and dependency versions should be
            #    bare versions ("1.2.3"), not upstream tag names ("v1.2.3").
            #    Download URLs may still contain refs/tags/v* when upstream
            #    uses that tag spelling.
            if grep -nE '\["v[0-9]+|\["[^"]+"\][[:space:]]*=[[:space:]]*"v[0-9]+' "$f"; then
              echo "::error file=$f::version identifiers must not use a leading v"
              fail=1
            fi
            # 4. Mirror table sanity: when a download `url` is written as a
            #    { GLOBAL=..., CN=... } table, both regions must be present and
            #    the CN entry must point at the gitcode mcpp-res mirror.
            if ! lua5.4 tests/check_mirror_urls.lua "$f"; then
              fail=1
            fi
            # 5. `name` must be a SINGLE ATOMIC SEGMENT; hierarchy belongs in
            #    `namespace` (mcpp SPEC-001 §3.2). The legacy fully-qualified
            #    spelling stays accepted. Cheap second gate: it runs before the
            #    pinned mcpp is even downloaded, and mcpp >= 0.0.106 enforces
            #    the same rule inside `mcpp xpkg parse`.
            if ! lua5.4 tests/check_package_name.lua "$f"; then
              fail=1
            fi
            # 6. c++fly admission policy (mcpp design 2026-07-14 §11-Q2, v1):
            #    c++fly means "toolchain's latest level + every experimental
            #    gate" — deliberately toolchain-dependent, so a published
            #    package built with it is not reproducible for consumers.
            #    Policy: WARN (never fail) and observe ecosystem usage before
            #    deciding whether to tighten. Two spellings: `language = ` is
            #    the descriptor's inline mcpp-segment key; `standard = ` covers
            #    mcpp.toml content embedded in heredoc/generated_files blocks.
            if grep -nE '\b(language|standard)[[:space:]]*=[[:space:]]*"c\+\+fly"' "$f" >/dev/null; then
              echo "::warning file=$f::declares C++ standard \"c++fly\" (experimental playground mode) — toolchain-dependent and non-reproducible for consumers; published packages should pin a concrete standard (c++23/c++26)"
            fi
          done
          [ $fail -eq 0 ] && echo "All package files valid."
          exit $fail
      # ── Whole-repository check (needs every descriptor at once) ──────
      # An install() hook addressing a sibling package does so by
      # `<namespace>:<literal package.name>`, and a miss returns nil rather
      # than raising — so a stale spelling surfaces far downstream (a broken
      # libxcb showed up as a libX11 link error). Verified across the repo
      # because it needs the full set of declared identities.
      - name: Lint cross-package references
        run: lua5.4 tests/check_cross_package_refs.lua pkgs/*/*.lua
      # ── Partial version bumps ────────────────────────────────────────
      # A bump is a one-line-looking edit that has to land in N platform
      # sections. Editing `xpm.linux` and reading the file back gives a file
      # that CONTAINS the new version, so the author -- and any whole-file
      # grep -- sees success, while the other platforms are left behind. The
      # failure then surfaces as `<pkg>@<ver> not found` on a platform,
      # against a file that literally contains that version string.
      #
      # Measured 2026-08-06: xpkg 0.0.52 and 0.0.53 were both added to
      # `xpm.linux` alone; linux CI went green twice while macOS and Windows
      # failed, and eight checks made from the outside all came back correct
      # because each asked "is it in the index?" instead of "is it in THIS
      # platform's section?". Whole-repo, because a partial bump is only
      # visible by comparing sections against each other.
      - name: Lint platform version parity
        run: lua5.4 tests/check_platform_version_parity.lua pkgs/*/*.lua
      # ── Single-source-of-truth grammar check ─────────────────────────
      # `mcpp xpkg parse` uses EXACTLY the resolver's parser, so what
      # passes here is what builds for users of the pinned MCPP_VERSION.
      # Strict by default: unknown mcpp-segment keys fail (they would be
      # silently ignored at build time). This also mechanically enforces
      # the rollout rule "floor first, new grammar after": descriptors
      # needing a newer grammar cannot pass a lint pinned to an older mcpp.
      - name: Download pinned mcpp
        run: |
          curl -L -fsS -o mcpp.tar.gz \
            "https://github.com/mcpp-community/mcpp/releases/download/v${MCPP_VERSION}/mcpp-${MCPP_VERSION}-linux-x86_64.tar.gz"
          tar -xzf mcpp.tar.gz
          echo "MCPP=$PWD/mcpp-${MCPP_VERSION}-linux-x86_64/bin/mcpp" >> "$GITHUB_ENV"
      - name: Parse descriptors with the resolver grammar (mcpp xpkg parse)
        run: |
          fail=0
          for f in pkgs/*/*.lua; do
            if ! "$MCPP" xpkg parse "$f" > /dev/null; then
              echo "::error file=$f::mcpp xpkg parse failed (resolver grammar)"
              fail=1
            fi
          done
          [ $fail -eq 0 ] && echo "All descriptors parse with mcpp ${MCPP_VERSION}."
          exit $fail

  mirror-cn-reachable:
    # Closed-loop guard for the CN mirror: every CN url referenced by a
    # descriptor must be a live, downloadable gitcode release asset.
    #
    # ── Scoped to the CHANGED descriptors ─────────────────────────────────
    # This used to walk all 87 descriptors on every run — 105 unique urls,
    # each fetched in full (`curl -o /dev/null` downloads the body; that is
    # the point, a gitcode release page can 200 while the asset behind it is
    # gone) — while a PR mirrors one package's assets at a time. Everything
    # else it re-downloaded was mirrored by an earlier PR that already
    # checked it, and a url that was live then can only have gone dark
    # SINCE, which is decay over time rather than something this PR did.
    #
    # Decay is real, and it is what the weekly cron below is for: `select`
    # answers `__ALL__` for schedule/dispatch, so the full 105-url sweep
    # still happens — once a week, and on demand — instead of on every push.
    # It also still happens on any PR that touches this workflow file or a
    # shared test script (including tests/list_cn_urls.lua, the extractor
    # this job reads the urls with), because those widen `select` to a full
    # run for the same reason they widen the workspace matrix.
    #
    # `needs: select` rather than a second diff here: the "can I trust this
    # diff" question is already answered there, in one place.
    needs: select
    # Empty when the diff names no descriptor at all — a docs-only or
    # tools-only PR mirrors nothing.
    if: needs.select.outputs.pkgs != ''
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - name: Install lua
        run: sudo apt-get install -y --no-install-recommends lua5.4
      - name: Check CN mirror assets are reachable
        env:
          PKGS: ${{ needs.select.outputs.pkgs }}
        run: |
          fail=0
          if [ "$PKGS" = "__ALL__" ]; then
            files=$(printf '%s\n' pkgs/*/*.lua)
            echo "full sweep: every descriptor"
          else
            files=$PKGS
            echo "changed descriptor(s): $files"
          fi
          # collect unique CN urls across the selected descriptors
          : > /tmp/cn.tsv
          for f in $files; do
            lua5.4 tests/list_cn_urls.lua "$f" >> /tmp/cn.tsv || true
          done
          sort -u /tmp/cn.tsv -o /tmp/cn.tsv
          total=$(grep -c . /tmp/cn.tsv || true)
          echo "checking $total CN mirror url(s)"
          while IFS=$'\t' read -r url sha; do
            [ -z "$url" ] && continue
            # follow redirects; gitcode release assets resolve to object storage
            code=$(curl -fsSL -o /dev/null -w '%{http_code}' --retry 2 --max-time 60 "$url" || echo "000")
            if [ "$code" != "200" ]; then
              echo "::error::CN mirror unreachable ($code): $url"
              fail=1
            else
              echo "ok: $url"
            fi
          done < /tmp/cn.tsv
          [ $fail -eq 0 ] && echo "All CN mirror urls reachable."
          exit $fail

  # ── The whole test surface, as a mcpp workspace ───────────────────────
  # mcpp-index is a mcpp [workspace]; every per-library test project under
  # tests/examples/ is a member. `mcpp test --workspace` builds + runs each
  # member's tests/ (behavioral assertions) on each OS — members self-gate by
  # `[target.'cfg(...)']` (e.g. the X11/glfw stack is linux-only, openblas is
  # windows-only), so one command covers the matrix with no shell driver.
  # The ~/.mcpp/registry cache carries the built compat packages (xpkgs) across
  # runs, so repeat builds are fast.
  #
  # timeout-minutes is sized for the COLD build, not the cached path. The opencv
  # module package carries a from-source OpenCV 5 build, and each feature variant
  # re-keys the store into a full recompile, so a full run (forced whenever this
  # workflow file changes — e.g. a version bump) serially builds three OpenCV
  # variants on one runner: the opencv-module base member plus the `unifont` and
  # `dnn` feature members. The registry cache (restore-keys prefix below)
  # amortizes those across subsequent runs. 150 covers the one-time cold full
  # build with headroom; it is a ceiling, not a target.
  # ── The plan, computed ONCE ───────────────────────────────────────────
  # Was inlined in every workspace job — three runners each re-deriving the
  # same answer. It now also has to be decided BEFORE the matrix exists,
  # because the matrix's shard dimension depends on it: a full run fans out,
  # a selective one does not.
  select:
    runs-on: ubuntu-latest
    outputs:
      members: ${{ steps.fanout.outputs.members }}
      matrix:  ${{ steps.fanout.outputs.matrix }}
      plan:    ${{ steps.plan_shards.outputs.plan }}
      # The descriptors this change actually touches, for `mirror-cn-reachable`.
      # `__ALL__` on a full run, empty when the diff names no descriptor at all.
      pkgs:    ${{ steps.plan.outputs.pkgs }}
    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0
      - name: Install lua
        run: sudo apt-get install -y --no-install-recommends lua5.4
      # ── Selective member testing ──────────────────────────────────────
      # `mcpp test --workspace` builds every member (opencv, ffmpeg, …) and
      # dominates CI wall-clock, while a PR almost always touches one
      # package. Map changed files → affected members and test only those:
      #   pkgs/<x>/<lib>.lua    → members whose mcpp.toml references <lib>
      #   tests/examples/<m>/** → member <m>
      # A push to main is mapped the same way — the merge's own diff — so a
      # merge costs what its PR cost. Run the FULL workspace when the change
      # can affect everything: the weekly cron and manual dispatch (which mean
      # "check everything" by definition), this workflow file (it carries the
      # mcpp version pins, so a version bump always re-validates every
      # package), a non-member edit to the workspace manifest, or shared test
      # scripts. Docs-only and tools/-only changes select nothing.
      # Note: bash 3.2 on macOS runners — no associative arrays here.
      #
      # The same walk also collects the CHANGED DESCRIPTORS (`pkgs` output) for
      # `mirror-cn-reachable`. Two selections, one classification: every reason
      # to distrust the diff — this workflow file, a shared test script, an
      # unclassified path, a push with no predecessor, cron/dispatch — has to
      # widen both, and writing that decision twice is how the two drift.
      - name: Select affected workspace members
        id: plan
        shell: bash
        run: |
          full() {
            echo "MEMBERS=__ALL__" >> "$GITHUB_ENV"
            echo "pkgs=__ALL__" >> "$GITHUB_OUTPUT"
            echo "full run: $1"
            exit 0
          }

          # A push to main has a diff too — it was just never asked for.
          #
          # This used to be `event != pull_request -> full`, so every merge
          # re-tested all 67 members on all three platforms: ~11 hours of
          # runner time to re-confirm what the PR had already gone green on
          # minutes earlier. The premise was that a push has no base to diff
          # against, and that is not true: merges here are squashes, so
          # `github.event.before` is the previous main and
          # `before..HEAD` reproduces exactly the file list the PR saw
          # (verified on 698b95ee — same ten paths).
          #
          # schedule and workflow_dispatch stay full. They are not "a change
          # landed", they are "check everything", which is the whole point of
          # the weekly net above.
          #
          # Two-dot for push, three-dot for pull_request, deliberately: a PR
          # wants its own commits against the merge base, while a push wants
          # what actually landed on this branch.
          case "${{ github.event_name }}" in
            pull_request)
              base="origin/${{ github.base_ref }}"; range="$base...HEAD" ;;
            push)
              base="${{ github.event.before }}"
              # All-zero on branch creation; absent object after a
              # force-push that dropped it. Either way there is nothing to
              # diff against, and guessing is worse than re-testing.
              case "$base" in
                ""|0000000000000000000000000000000000000000)
                  full "push with no predecessor" ;;
              esac
              git cat-file -e "$base^{commit}" 2>/dev/null \
                || full "push predecessor $base not in history"
              range="$base..HEAD" ;;
            *)
              full "event=${{ github.event_name }}" ;;
          esac
          changed=$(git diff --name-only $range)
          printf 'changed files vs %s:\n%s\n' "$base" "$changed"
          sel=""; pkgsel=""
          add() { case " $sel " in *" $1 "*) ;; *) sel="$sel $1" ;; esac; }
          while IFS= read -r f; do
            [ -n "$f" ] || continue
            case "$f" in
              .github/workflows/validate.yml|tests/*.sh) full "$f" ;;
              mcpp.toml)
                # Workspace manifest. Every new-package PR appends to the
                # members list, so that alone must NOT force a full run:
                # select the added members; anything else in this file
                # (indices, settings) affects everyone → full.
                if ! diff -q <(git show "$base:mcpp.toml" | grep -v 'tests/examples/') \
                             <(grep -v 'tests/examples/' mcpp.toml) >/dev/null; then
                  full "mcpp.toml non-member change"
                fi
                for p in $(comm -13 <(git show "$base:mcpp.toml" | grep -o 'tests/examples/[A-Za-z0-9._-]*' | sort -u) \
                                    <(grep -o 'tests/examples/[A-Za-z0-9._-]*' mcpp.toml | sort -u)); do
                  add "${p#tests/examples/}"
                done ;;
              tests/examples/*)
                m=${f#tests/examples/}; m=${m%%/*}
                # A deleted/renamed member dir implies a mcpp.toml edit,
                # which already forces a full run above.
                [ -d "tests/examples/$m" ] && add "$m" ;;
              pkgs/*.lua|pkgs/*/*.lua)
                # A DELETED descriptor is in the diff but not on disk, and its
                # CN urls are gone with it — nothing left to reach.
                if [ -f "$f" ]; then pkgsel="$pkgsel $f"; fi
                lib=$(basename "$f" .lua); lib=${lib#compat.}
                hit=0
                for mt in tests/examples/*/mcpp.toml; do
                  if grep -q "$lib" "$mt"; then add "$(basename "$(dirname "$mt")")"; hit=1; fi
                done
                [ "$hit" = 1 ] || echo "note: no workspace member exercises $f" ;;
              # tools/ holds OFFLINE descriptor-generation and publishing
              # helpers (tools/compat-*/, tools/gtc/, publish_mcpp_index.sh).
              # Nothing under it is consumed by a package build: when one of
              # them actually changes a package, the generated pkgs/*.lua
              # changes with it and the rule above selects the right members.
              # So a tools/ edit alone selects nothing rather than forcing a
              # full workspace rebuild.
              # The timing table decides how work is DISTRIBUTED across
              # shards, never what gets built — no member's result can change
              # because a measured number moved. It used to fall through to
              # the `*)` catch-all below and force a full run, which is the
              # most expensive way in this workflow to test nothing: the next
              # full run reads the new numbers anyway.
              #
              # Note this makes the file invisible to CI. A malformed row is
              # then silent, and `plan_shards` prices anything it cannot parse
              # at the median — the exact failure mode that overflowed a shard
              # here. Guard it in `lint` if that ever bites.
              tests/member-timings.tsv) : ;;
              *.md|docs/*|.agents/*|.github/*|tools/*) : ;;
              *) full "unclassified change: $f" ;;
            esac
          done <<EOF
          $changed
          EOF
          sel=${sel# }
          echo "MEMBERS=$sel" >> "$GITHUB_ENV"
          echo "selected members: ${sel:-<none>}"
          pkgsel=${pkgsel# }
          echo "pkgs=$pkgsel" >> "$GITHUB_OUTPUT"
          echo "changed descriptors: ${pkgsel:-<none>}"
      # Sharding is for the FULL run only, and the shard count per platform is
      # that platform's RUNNER CONCURRENCY — not a round number.
      #
      # Measured on this repo (24 jobs queued, 6 running):
      #     macos 1 · linux 3 · windows 2
      #
      # That measurement is what makes over-sharding a real cost rather than a
      # theoretical one: at concurrency 1, eight macOS shards run BACK TO BACK
      # and each pays its own checkout + mcpp download + cache restore, so the
      # split is strictly slower than not splitting. Wall-clock is
      # ceil(shards / concurrency) x slowest-shard; shards beyond the
      # concurrency only add fixed cost.
      #
      # Re-measure with:
      #   gh api repos/<owner>/<repo>/actions/runs/<id>/jobs --paginate \
      #     --jq '[.jobs[]|select(.status=="in_progress")]|length'
      - name: Decide the fan-out
        id: fanout
        shell: bash
        run: |
          full=0; [ "$MEMBERS" = "__ALL__" ] && full=1
          emit() {   # platform os suffix ext mcpp xlings toolchain shards
            for i in $(seq 0 $(( $8 - 1 ))); do
              printf '{"platform":"%s","os":"%s","suffix":"%s","ext":"%s","mcpp":"%s","xlings":"%s","toolchain":"%s","shard":%d,"shards":%d},' \
                     "$1" "$2" "$3" "$4" "$5" "$6" "$7" "$i" "$8"
            done
          }
          # Shard count follows the WORK, per platform, measured.
          #
          # It used to be binary — full run: linux 3 / macos 1 / windows 2,
          # anything else: 1 each — which asserts two things that are not
          # true. "A partial run is small": touching a widely consumed
          # descriptor selects every member that consumes it, and this PR's
          # `pkgs/c/compat.openssl.lua` selected four; one linux shard took
          # grpc-codegen (3563s) and grpc-module (1701s) back to back and was
          # cancelled at exactly 1h30m, the job cap. "A full macOS run fits in
          # one shard": at 6922s measured it does not — it fit only while the
          # total sat just under the cap, and the run that added grpc-codegen
          # (1715s on macOS) pushed it to 1h30m20s.
          #
          # Sharding is not only a wall-clock lever. Each shard is its own job
          # with its own `timeout-minutes`, so it is also how the work is made
          # to FIT. That is the half the "macOS concurrency is 1, so extra
          # shards run back to back" note above left out: back to back is fine
          # when the alternative is not finishing.
          #
          # tests/member-timings.tsv already holds every member's measured
          # cost and plan_shards.lua already packs by it (LPT), so the fan-out
          # comes from the same table: sum this run's members for that
          # platform and take one shard per ~70 minutes, which leaves ~20
          # minutes of headroom under the 90-minute cap for a cold cache.
          #
          # The caps are where runner concurrency comes back in: linux 4,
          # macOS 2, windows 2. Past those, more shards buy fit that is
          # already there and pay another checkout + mcpp download + cache
          # restore.
          #
          # linux was 3 — "the measured concurrency, a 4th shard would queue".
          # That cap outlived its premise twice over: the workspace grew, and
          # the second toolchain leg means linux emits 2 x N jobs, so it
          # queues at any N. It is now 4, which is what the formula asks for.
          #
          # Full run, from the table: linux 15891s over 4 -> ~74min slowest
          # shard, macOS 6922s over 2 -> ~57, windows 8043s over 2 -> ~67.
          #
          # linux is still the tight one, and the number to watch is the
          # SLOWEST shard against the 90-minute job cap, not the average.
          # Simulated on the current table: 3-way 98min, 4-way 74, 5-way 60.
          # The levers, in order: raise the linux cap to 5, then the job
          # timeout.
          #
          # Keep tests/member-timings.tsv fresh — that is not housekeeping. An
          # untimed member is priced at the MEDIAN, so a heavy newcomer packs
          # like a trivial one. `mysql-connector-cpp` (881s) and
          # `libmysqlclient` (329s) landed on shard 0 next to grpc-codegen
          # exactly that way and pushed it past the cap, and the stale table
          # then hid the overflow by under-counting the total.
          # `plan_shards.lua <platform> 0 0` reports the platform's total cost.
          #
          # This used to sum the timing table's rows inline, which was the same
          # answer only while the table listed every member. It now lists just
          # the heavy ones, so summing rows measures a fraction of the work:
          # 10965s instead of 16095s on linux, i.e. 3 shards where 4 are
          # needed, i.e. a slowest shard of 107 minutes against a 90-minute
          # cap. Asking plan_shards keeps ONE definition of both "which members
          # are there" and "what does an untimed one cost".
          #
          # The estimate is tuned on linux (16095s vs 15891s measured, +1.3%).
          # macOS and windows members are cheaper on average than the default,
          # so their totals come out high — harmless here because both are
          # already at their cap, and erring toward more shards is the safe
          # direction. Revisit if either cap moves.
          shards_for() {   # platform cap -> shard count
            local secs sel=""
            [ "$MEMBERS" = "__ALL__" ] || sel="$MEMBERS"
            secs=$(lua5.4 tests/plan_shards.lua "$1" 0 0 $sel)
            local n=$(( secs / 4200 + 1 ))
            [ "$n" -gt "$2" ] && n="$2"
            echo "$1 work: ${secs}s (measured) -> $n shard(s)" >&2
            echo "$n"
          }
          # linux 3 -> 4. The cap was the binding constraint, not the ~70min
          # target: `secs / 4200 + 1` asks for 4 on the refreshed table
          # (15891s) and has been getting clamped back to 3. Simulated on that
          # table, slowest shard: 3-way 98min (OVER the 90min job cap — and
          # `linux default 0/3` was in fact cancelled at 90min on run
          # 31266814148), 4-way 74min, 5-way 60min. 4 is the formula's own
          # answer and leaves 16 minutes for a cold cache; 5 is there if that
          # stops being enough.
          #
          # This does cross the concurrency line the old comment drew: with two
          # toolchain legs linux now emits 8 jobs against a measured runner
          # concurrency of 3, so shards queue. Queuing is the right trade —
          # back-to-back shards still finish, a shard over the cap does not.
          ln=$(shards_for linux 4)
          mn=$(shards_for macos 2)
          wn=$(shards_for windows 2)
          {
            printf '{"include":['
            # linux runs TWICE: once on mcpp's default toolchain and once on
            # llvm. Not redundancy — the two differ in a way that decides
            # whether a descriptor is correct. mcpp's linux default (gcc)
            # reaches the compiler through --sysroot into a clean xlings
            # subos, so the HOST /usr/include is not on the search path;
            # llvm has no sysroot and the host's headers are. compat.ffmpeg
            # put its vendored source root on -idirafter (below the system
            # dirs) and was green here for months while being broken on any
            # developer box with libavutil-dev installed — the gcc leg
            # structurally cannot see that class of bug.
            #
            # Cost, measured: this doubles linux to 6 jobs (2 x 3 shards) at
            # ~75min each, against a measured linux runner concurrency of 3 —
            # so the second leg QUEUES behind the first rather than running
            # beside it, and a full run's linux wall-clock roughly doubles.
            # That is the price of the coverage; the levers if it hurts are
            # in the shards_for comment above.
            emit linux   ubuntu-latest  linux-x86_64   tar.gz bin/mcpp     registry/bin/xlings     default "$ln"
            emit linux   ubuntu-latest  linux-x86_64   tar.gz bin/mcpp     registry/bin/xlings     llvm    "$ln"
            emit macos   macos-15       macosx-arm64   tar.gz bin/mcpp     registry/bin/xlings     default "$mn"
            emit windows windows-latest windows-x86_64 zip    bin/mcpp.exe registry/bin/xlings.exe default "$wn"
            printf ']}'
          } | sed 's/,]}/]}/' > /tmp/matrix.json
          echo "matrix=$(cat /tmp/matrix.json)" >> "$GITHUB_OUTPUT"
          echo "members=$MEMBERS" >> "$GITHUB_OUTPUT"
          cat /tmp/matrix.json

      # The split is computed ONCE, here, and shipped to the runners as data.
      # It used to run on each runner, which needed lua5.4 on all three
      # platforms — windows has no apt or brew, and macOS's brew installs
      # `lua`, not `lua5.4`, so every non-linux shard died with
      # `lua5.4: command not found` after 16 seconds. Deciding once is also
      # simply correct: one plan, not three runners each re-deriving it.
      - name: Plan the shards
        id: plan_shards
        shell: bash
        run: |
          plan='${{ steps.fanout.outputs.members }}'
          [ "$plan" = "__ALL__" ] && plan=""
          {
            printf '{'
            first=1
            # `.shards`, NOT the number of matrix entries.
            #
            # These stopped being the same number the moment linux grew a
            # second toolchain leg: the platform emits 2 x 3 = 6 entries while
            # the split is still 3 ways, and every entry carries shard 0..2.
            # Planning 6 ways and consuming three of them dropped 36 of 65
            # members on the floor — silently, because a member that is never
            # assigned is indistinguishable from one that passed. `ffmpeg`,
            # `opencv-module*`, `catch2-v2` and `openssl` were among them, on
            # the very run that added the leg meant to test them.
            #
            # `.shards` is the value emit() already wrote per entry and the
            # same one the job reads for its own `matrix.shards`, so plan and
            # consumer now read one number instead of two that agreed by
            # accident.
            for spec in linux:$(jq -r 'first(.include[]|select(.platform=="linux")).shards' /tmp/matrix.json) \
                        macos:$(jq -r 'first(.include[]|select(.platform=="macos")).shards' /tmp/matrix.json) \
                        windows:$(jq -r 'first(.include[]|select(.platform=="windows")).shards' /tmp/matrix.json); do
              p=${spec%%:*}; n=${spec##*:}
              [ "$first" = 1 ] || printf ','
              first=0
              printf '"%s":{' "$p"
              for i in $(seq 0 $((n - 1))); do
                [ "$i" = 0 ] || printf ','
                m=$(lua5.4 tests/plan_shards.lua "$p" "$i" "$n" $plan)
                printf '"%s":"%s"' "$i" "$m"
              done
              printf '}'
            done
            printf '}'
          } > /tmp/plan.json
          echo "plan=$(cat /tmp/plan.json)" >> "$GITHUB_OUTPUT"
          jq . /tmp/plan.json

  # Installing the graphics stack must not change what UNRELATED members link
  # against. This is the check #179 did not have: every test in this repo asks
  # about its own package, so nothing asked whether installing one changes
  # things for everyone else -- and `asio-module` / `core` went red on a
  # change that never touched them.
  #
  # Its own job, on Linux, because it needs a real toolchain and a home it can
  # watch payloads appear in. It reports INCONCLUSIVE (and fails) rather than
  # green whenever its preconditions do not hold; see the script.
  graphics-side-effects:
    name: "graphics install: no side effects on unrelated members"
    runs-on: ubuntu-latest
    timeout-minutes: 60
    steps:
      - uses: actions/checkout@v4
      - name: Download mcpp
        shell: bash
        env:
          MCPP_ARCHIVE: mcpp-${{ env.MCPP_VERSION }}-linux-x86_64.tar.gz
          MCPP_ROOT: mcpp-${{ env.MCPP_VERSION }}-linux-x86_64
        run: |
          curl -L -fsS -o "$MCPP_ARCHIVE" \
            "https://github.com/mcpp-community/mcpp/releases/download/v${MCPP_VERSION}/${MCPP_ARCHIVE}"
          tar -xzf "$MCPP_ARCHIVE"
          root="$PWD/$MCPP_ROOT"
          # MCPP_HOME is the tarball root, not ~/.mcpp.
          #
          # A released mcpp is self-contained: it resolves its registry from
          # beside its own executable. Copying the payload tree into ~/.mcpp
          # and pointing MCPP_HOME there therefore watches one home while the
          # build uses another -- which this check detects and refuses to call
          # green (it did, on the first run). Aligning them is the fix; the
          # check was right.
          echo "MCPP=$root/bin/mcpp" >> "$GITHUB_ENV"
          echo "MCPP_HOME=$root" >> "$GITHUB_ENV"
          echo "MCPP_VENDORED_XLINGS=$root/registry/bin/xlings" >> "$GITHUB_ENV"
      - name: Side-effect check
        shell: bash
        run: |
          "$MCPP" self config --mirror GLOBAL 2>/dev/null || true
          bash tests/check_graphics_install_side_effects.sh

  workspace:
    # The shard suffix appears only when the platform is actually split, and
    # the toolchain only when it is not mcpp's default — so the existing job
    # names are unchanged and only the new llvm leg is labelled.
    name: workspace (${{ matrix.platform }}${{ matrix.toolchain == 'default' && '' || format(' {0}', matrix.toolchain) }}${{ matrix.shards == 1 && '' || format(' {0}/{1}', matrix.shard, matrix.shards) }})
    needs: select
    if: needs.select.outputs.members != ''
    runs-on: ${{ matrix.os }}
    # One shard is a fraction of the work, so this is a real ceiling rather
    # than the thing that decides whether the job finishes (a full linux run
    # used to hit 150 exactly and get cancelled).
    timeout-minutes: 90
    strategy:
      fail-fast: false
      # Whole matrix from `select`: the shard count is per-platform, because it
      # tracks that platform's runner concurrency.
      matrix: ${{ fromJSON(needs.select.outputs.matrix) }}
    steps:
      # Full history: the member-selection step below diffs against the PR
      # base to decide which workspace members to test.
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0
      # The cache key is computed ONCE, here, instead of inline in the cache
      # step. `hashFiles()` globs the WORKING TREE, and actions/cache
      # re-evaluates its `key` in the post (save) step — i.e. AFTER the build,
      # when `tests/**` no longer matches 80-odd tracked sources but tens of
      # thousands of build-output files under tests/examples/*/target and
      # .mcpp (multi-GB; .gitignore does not apply to hashFiles). Hashing that
      # tree blew past the runner's 120s template-evaluation cap on windows
      # and failed an otherwise all-green job:
      #   "hashFiles('pkgs/**/*.lua, tests/**, .github/workflows/validate.yml')
      #    couldn't finish within 120 seconds"
      # `git ls-files -s` reads the INDEX, so it sees exactly the tracked
      # inputs, never build output, and reports blob SHAs git already has —
      # no file content is read at all. Freezing the result in the job env
      # also guarantees the save step keys on the same string the restore
      # step used, no matter what the build left behind.
      - name: Compute registry cache key
        shell: bash
        run: |
          # git hash-object rather than sha256sum/cut: git is already a hard
          # requirement here (checkout ran), coreutils on the windows leg is
          # only a Git-Bash convenience.
          h=$(git ls-files -s -- 'pkgs/**/*.lua' 'tests/**' '.github/workflows/validate.yml' \
              | git hash-object --stdin)
          # matrix.toolchain is in the key: both linux legs are runner.os
          # Linux, and the cache holds TOOLCHAINS and built compat packages.
          # Sharing one entry would let the gcc leg's objects answer for the
          # llvm leg — exactly the confusion this second leg exists to catch.
          echo "REGISTRY_CACHE_KEY=mcpp-registry-${{ runner.os }}-${{ matrix.toolchain }}-${{ env.MCPP_VERSION }}-$h" >> "$GITHUB_ENV"
      - name: Restore mcpp registry cache
        uses: actions/cache@v4
        with:
          # Holds toolchains AND the built compat packages (data/xpkgs), so a
          # repeat `mcpp test` rebuilds little.
          path: ~/.mcpp/registry
          key: ${{ env.REGISTRY_CACHE_KEY }}
          restore-keys: |
            mcpp-registry-${{ runner.os }}-${{ matrix.toolchain }}-${{ env.MCPP_VERSION }}-
      # Host tools built from source (protoc, grpc_cpp_plugin, …).
      #
      # Measured on the run that added grpc-codegen: 636s to build protoc and
      # 660s to build grpc_cpp_plugin — 1296s of a 3363s member, repeated in
      # every job of every run, because the registry cache above covers
      # ~/.mcpp/registry and the tool store lives in ~/.mcpp/build-cache.
      # `protobuf-protoc` and `grpc-module` build the SAME protoc, so a full
      # run pays for it several times over.
      #
      # A coarse rolling key is safe here, which is why this is cheap. The
      # store is content-addressed by `<pkg>@<version>/<hash>` and mcpp
      # validates an entry FIELD BY FIELD against the recorded entry.json
      # (epoch, target, host triple, compiler identity, profile, features,
      # transitive dependency closure). A stale entry is not used; it is
      # rebuilt. The worst case of a cache miss-match is the status quo.
      #
      # Deliberately NOT build-cache/v1/pkg: 6.0 GB measured locally against
      # Actions' 10 GB per-repo budget, which would evict the registry cache
      # this job needs more. The tool store is 116 MB.
      - name: Cache host tool store
        uses: actions/cache@v4
        with:
          path: ~/.mcpp/build-cache/v1/tool
          key: mcpp-toolstore-${{ runner.os }}-${{ matrix.toolchain }}-${{ env.MCPP_VERSION }}-${{ github.run_id }}-${{ matrix.platform }}-${{ matrix.shard }}
          restore-keys: |
            mcpp-toolstore-${{ runner.os }}-${{ matrix.toolchain }}-${{ env.MCPP_VERSION }}-
      - name: Download mcpp
        shell: bash
        env:
          MCPP_ARCHIVE: mcpp-${{ env.MCPP_VERSION }}-${{ matrix.suffix }}.${{ matrix.ext }}
          MCPP_ROOT: mcpp-${{ env.MCPP_VERSION }}-${{ matrix.suffix }}
        run: |
          curl -L -fsS -o "$MCPP_ARCHIVE" \
            "https://github.com/mcpp-community/mcpp/releases/download/v${MCPP_VERSION}/${MCPP_ARCHIVE}"
          case "$MCPP_ARCHIVE" in
            *.zip) powershell -NoProfile -Command "Expand-Archive -Force -Path '${MCPP_ARCHIVE}' -DestinationPath '.'" ;;
            *)     tar -xzf "$MCPP_ARCHIVE" ;;
          esac
          root="$PWD/$MCPP_ROOT"
          mkdir -p "$HOME/.mcpp/registry"
          cp -a "$root/registry/." "$HOME/.mcpp/registry/"
          # Point mcpp AT the home the caches above cover.
          #
          # Without this the copy above is one-way scenery: a released mcpp is
          # self-contained and resolves its registry from beside its own
          # executable, so every build used <tarball>/registry while the cache
          # restored, and saved, ~/.mcpp/registry. Proof from the run that
          # added grpc-codegen: abseil compiled from
          # `<tarball>/registry/data/xpkgs/compat-x-abseil/...`, and
          # xim:glibc@2.44 / xim:python@3.13.12 were downloaded again on a job
          # that reported a registry cache HIT.
          #
          # With MCPP_HOME set, the restored registry is the one in play — so
          # payloads stop being re-downloaded per job, and the tool store cache
          # below lands where the build looks for it.
          echo "MCPP_HOME=$HOME/.mcpp" >> "$GITHUB_ENV"
          if [[ "$RUNNER_OS" == "Windows" ]]; then
            echo "MCPP=$(cygpath -m "$root/${{ matrix.mcpp }}")" >> "$GITHUB_ENV"
            echo "MCPP_VENDORED_XLINGS=$(cygpath -m "$root/${{ matrix.xlings }}")" >> "$GITHUB_ENV"
            echo "$(cygpath -m "$root/bin")" >> "$GITHUB_PATH"
          else
            echo "MCPP=$root/${{ matrix.mcpp }}" >> "$GITHUB_ENV"
            echo "MCPP_VENDORED_XLINGS=$root/${{ matrix.xlings }}" >> "$GITHUB_ENV"
            echo "$root/bin" >> "$GITHUB_PATH"
          fi
      # compat.ffmpeg / compat.opencv5 carry NASM .asm sources. No host
      # install and no index-refresh pre-step needed: mcpp >= 0.0.97
      # resolves nasm itself through the same synchronous gate as the
      # toolchain (index refresh + install + payload check BEFORE the build
      # plans, mcpp#232). The sandbox copy lands in ~/.mcpp/registry, so
      # the cache carries it across runs.

      # ── Make the llvm leg look like a developer's box ─────────────────
      # A bare GitHub runner has no libav* under /usr/include, so a
      # descriptor that loses its vendored headers to the host's would still
      # be green here. That is precisely how compat.ffmpeg's -idirafter hole
      # survived: nothing in CI had a system ffmpeg to lose to. Installing
      # the dev headers is what turns this leg into a real test of include
      # ORDER rather than of include EXISTENCE.
      #
      # Deliberately only on the llvm leg. The default leg reaches its
      # compiler through --sysroot, so it is immune by construction and
      # installing there would change an existing job's environment for no
      # coverage. Keeping the blast radius to the new leg means a red here
      # is always about the thing this leg was added to watch.
      - name: Install host ffmpeg dev headers (llvm leg only)
        if: matrix.platform == 'linux' && matrix.toolchain == 'llvm'
        shell: bash
        run: |
          sudo apt-get update -qq
          sudo apt-get install -y -qq \
            libavutil-dev libavcodec-dev libavformat-dev \
            libavfilter-dev libavdevice-dev libswscale-dev libswresample-dev
          dpkg -l | grep -E 'libav(util|codec)-dev|libswscale-dev'

      # ── Select the toolchain for this leg ─────────────────────────────
      # No version pin: `llvm` resolves to whatever the toolchain family's
      # current release is for the PINNED MCPP_VERSION, which is the version
      # that mcpp itself vouches for. Pinning a concrete llvm here would fail
      # hard the day that mcpp release stops shipping it, and the point of
      # this leg is the ABSENCE of a sysroot, not a particular llvm.
      - name: Select toolchain
        if: matrix.toolchain != 'default'
        shell: bash
        run: |
          "$MCPP" toolchain install ${{ matrix.toolchain }}
          "$MCPP" toolchain default ${{ matrix.toolchain }}
          "$MCPP" toolchain list

      # ── This shard's slice of the plan ────────────────────────────────
      # `select` decided WHAT runs; this decides which part of it runs HERE.
      # Round-robin by position, which is what spreads the expensive members:
      # opencv-module / -dnn / -unifont are adjacent in the list, so `% N`
      # necessarily puts them on three different runners. A single job that
      # builds all three spends 45+ minutes on opencv alone.
      # ── This shard's slice ────────────────────────────────────────────
      # Already decided by `select` (measured-time bin packing, see
      # tests/plan_shards.lua). Arrives as data, so a runner needs no lua.
      - name: Take this shard's members
        shell: bash
        run: |
          mine='${{ fromJSON(needs.select.outputs.plan)[matrix.platform][format('{0}', matrix.shard)] }}'
          echo "MEMBERS=$mine" >> "$GITHUB_ENV"
          echo "shard ${{ matrix.shard }}/${{ matrix.shards }}: ${mine:-<none>}"

      # ── Refresh the PUBLISHED index before testing ────────────────────
      # Most members resolve everything from this checkout, but a member that
      # redirects a namespace other than `compat` gets the REST from the
      # published index — and nothing here ever refreshed it. The snapshot in
      # play is whatever the pinned mcpp release vendored (the Download step
      # `cp -a`s the release's registry/ over ~/.mcpp/registry, on top of the
      # restored cache), so it is by construction older than main, and it
      # never moves: the cache is saved with that same stale copy inside it.
      #
      # mcpp does refresh on a miss for a DIRECT dependency, which is why this
      # went unnoticed — the gap is a Form-A package's TRANSITIVE dependency.
      # tests/examples/godot-cpp-module hit it head-on: the module package's
      # own compat.godot-cpp dep resolved against a snapshot predating the
      # commit that added it, and failed with `index: local index <sha> (never
      # refreshed)` even though the artifact had already been republished.
      # (Older members never noticed: their compat packages have been in the
      # index far longer than any snapshot.)
      - name: Refresh the published package index
        shell: bash
        env:
          MCPP_INDEX_MIRROR: GLOBAL
        run: |
          "$MCPP" index update

      - name: mcpp test (workspace or affected members)
        shell: bash
        env:
          MCPP_INDEX_MIRROR: GLOBAL
          # The GLOBAL package build cache is on (mcpp >= 2026.7.30.2), which
          # is the default — this step used to set `MCPP_BUILD_CACHE: local`
          # and no longer does.
          #
          # That bypass existed for mcpp#344: object-path disambiguation fires
          # on basename collisions across the WHOLE build dir — i.e. on what
          # the CONSUMER pulls in — while the cache key covered only the
          # dependency, so one entry could hold two layouts and ninja died at
          # graph time with "missing and no known rule to make it". #344
          # landed in 2026.8.3.4 with per-package Merkle keys that cover the
          # consumer-dependent layout, so the reason is gone.
          #
          # Keeping it cost real time, and the full run is where it showed:
          # with `local`, EVERY member recompiles EVERY dependency from
          # scratch. 59 members that mostly share abseil / protobuf / opencv
          # meant the same sources were built over and over —
          #
          #   linux    2h30m -> cancelled at the 150-minute timeout
          #   windows  2h20m
          #   macos    1h26m
          #
          # — and a workspace cannot be validated by a job that cannot finish.
          # With the cache on, a given (package, version, features, toolchain)
          # is built once per run and every later member hits it.
        run: |
          "$MCPP" --version
          # No `timeout` wrapper: absent on macOS runners; job-level timeout-minutes bounds it.
          # One code path: the shard step above already expanded `__ALL__`
          # into this runner's actual member names, so `mcpp test --workspace`
          # — which would ignore the sharding and rebuild everything here — is
          # gone.
          #
          # tests/run_members.sh is the SAME script you run locally. A timing
          # table that only exists in CI cannot be used while deciding what to
          # optimise, and a local harness that differs from CI measures
          # something else.
          if [ -z "$MEMBERS" ]; then
            echo "No workspace member affected by this change — nothing to test."
          else
            MCPP_TIMINGS="$PWD/timings.tsv" bash tests/run_members.sh $MEMBERS
          fi

      # Per-shard timings, merged by the `timings` job below. `always()`: a
      # run that failed is exactly when knowing where the time went matters.
      - name: Upload this shard's timings
        if: always() && hashFiles('timings.tsv') != ''
        uses: actions/upload-artifact@v4
        with:
          # The toolchain is part of the name because upload-artifact@v4
          # refuses duplicates: without it both linux legs would race to
          # upload `timings-linux-0` and the second one would fail the job.
          name: timings-${{ matrix.platform }}-${{ matrix.toolchain }}-${{ matrix.shard }}
          path: timings.tsv
          retention-days: 14

      # install()-driven packages (openssl, openblas) build through their own
      # Make/Configure system, whose output xim's interface mode swallows; a
      # failed hook surfaces only as `E_INTERNAL: [<pkg>] failed:`. Each writes
      # a log into its install prefix, so on failure surface those — otherwise
      # diagnosing a platform-specific build break costs a full CI round-trip
      # per guess.
      - name: Dump install() build logs on failure
        if: failure()
        shell: bash
        run: |
          found=0
          while IFS= read -r log; do
            found=1
            echo "::group::$log"
            tail -80 "$log"
            echo "::endgroup::"
          done < <(find tests/examples "$HOME/.mcpp/registry" -name 'mcpp_*_build.log' 2>/dev/null)
          [ "$found" = 1 ] || echo "no install() build logs found"

  # ── Where the time went ───────────────────────────────────────────────
  # Sharding hides the cost: eight runners each report their own slice, and
  # nobody can see which members actually dominate. This merges them into one
  # ranking per platform, in the run summary, so the next optimisation starts
  # from measurement instead of a guess.
  #
  # `always()` — a failed run is exactly when this is worth reading.
  timings:
    needs: [select, workspace]
    if: always() && needs.select.outputs.members != ''
    runs-on: ubuntu-latest
    steps:
      - uses: actions/download-artifact@v4
        with:
          pattern: timings-*
          path: timings
        continue-on-error: true
      - name: Rank members by wall-clock
        shell: bash
        run: |
          shopt -s nullglob
          files=(timings/*/timings.tsv)
          if [ ${#files[@]} -eq 0 ]; then
            echo "no timing data (every shard skipped or failed before testing)" \
              >> "$GITHUB_STEP_SUMMARY"
            exit 0
          fi

          # Artifact name carries platform AND toolchain leg:
          # timings-<platform>-<toolchain>-<shard>. Ranking per LEG, not per
          # platform: the two linux legs are different builds and averaging
          # them would describe neither.
          for leg in linux-default linux-llvm macos-default windows-default; do
            rows=$(mktemp)
            for f in timings/timings-$leg-*/timings.tsv; do
              [ -f "$f" ] && cat "$f" >> "$rows"
            done
            [ -s "$rows" ] || { rm -f "$rows"; continue; }

            total=$(awk -F'\t' '{s += $1} END {print s+0}' "$rows")
            count=$(wc -l < "$rows")
            {
              echo "### $leg — ${count} member(s), ${total}s of member wall-clock"
              echo
              echo "| rank | seconds | share | member | result |"
              echo "|---:|---:|---:|---|---|"
              sort -rn "$rows" | awk -F'\t' -v tot="$total" '
                { pct = tot > 0 ? ($1 * 100 / tot) : 0
                  printf "| %d | %s | %.1f%% | `%s` | %s |\n", NR, $1, pct, $2, $3 }'
              echo
            } >> "$GITHUB_STEP_SUMMARY"
            rm -f "$rows"
          done

          echo "_Total is the SUM across shards; wall-clock is the slowest shard._" \
            >> "$GITHUB_STEP_SUMMARY"

          # The table that feeds the NEXT run's sharding. Emitted as an
          # artifact rather than committed automatically: a number that
          # rewrites itself on every run would make every diff noisy and would
          # silently absorb a one-off slow runner. Refresh it deliberately —
          # download this artifact and replace tests/member-timings.tsv when
          # the numbers have actually moved.
          # Top 10 per platform, not every member.
          #
          # plan_shards prices anything absent at a fixed default, and the
          # packing decisions are made by the heavy members anyway — on linux
          # the top 10 of 67 are 69% of the total, and dropping the other 57
          # moves the slowest shard by 0 minutes. A table that lists everyone
          # goes stale the moment someone adds a test; one that lists ten only
          # goes stale when the heavy set actually changes. That is the whole
          # point: adding a member must not require touching this file.
          {
            echo "# <platform>\t<member>\t<seconds> — the HEAVY members only, top 10"
            echo "# per platform, from run ${{ github.run_id }}. Everything else is"
            echo "# priced at plan_shards.lua's fixed default; adding a test does NOT"
            echo "# require touching this file. Refresh when a member becomes heavy"
            echo "# enough to enter the top 10, or one of these numbers moves."
            # ONLY the default-toolchain leg feeds this table. The llvm leg
            # runs the SAME members again, so globbing every leg would put two
            # rows per (platform, member) into the file — `sort -u` keeps both,
            # since the seconds differ. Both legs are planned from these
            # numbers, so the default leg is the right single baseline.
            #
            # Then: heaviest first, keep 10. `sort -u` before that so a member
            # reported by two shards cannot occupy two of the ten slots.
            for plat in linux macos windows; do
              for f in timings/timings-$plat-default-*/timings.tsv; do
                [ -f "$f" ] || continue
                awk -F'\t' -v p="$plat" '{ printf "%s\t%s\t%s\n", p, $2, $1 }' "$f"
              done | sort -u | sort -t"$(printf '\t')" -k3,3nr | head -10
            done
          } > member-timings.tsv
          echo "wrote member-timings.tsv ($(grep -vc '^#' member-timings.tsv) rows, top 10/platform)"

      - name: Upload the timing table for the next run's sharding
        if: always() && hashFiles('member-timings.tsv') != ''
        uses: actions/upload-artifact@v4
        with:
          name: member-timings
          path: member-timings.tsv
          retention-days: 90

  # ─────────────────────────────────────────────────────────────────────────
  # A full sweep that fails must be SEEN.
  #
  # Selective CI is structurally unable to notice a published package that
  # broke because something UPSTREAM changed: no PR's diff names it, so no
  # PR selects it. The weekly cron exists for exactly that, and it already
  # ran — but a red cron blocks nothing and pages nobody, so it may as well
  # not have run. Measured: eight graphics members failed the same way for
  # weeks while `main` showed green, because the job that builds them was
  # `skipped` on every push.
  #
  # `main` is green and the packages install are two different claims. This
  # job turns the second one into an issue somebody sees.
  #
  # Only for full sweeps (schedule / manual dispatch). A PR's failure is
  # already visible on the PR.
  sweep-alert:
    name: full sweep failed — open an issue
    needs: [workspace]
    if: >-
      always()
      && needs.workspace.result == 'failure'
      && (github.event_name == 'schedule' || github.event_name == 'workflow_dispatch')
    runs-on: ubuntu-latest
    permissions:
      issues: write
    steps:
      - name: Open or update the sweep-failure issue
        env:
          GH_TOKEN: ${{ github.token }}
          RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
        run: |
          set -euo pipefail
          TITLE="full workspace sweep is failing"
          # One issue, reopened and appended to — not one per run. A tracker
          # that produces a new issue every week is a tracker people mute.
          NUM=$(gh issue list --state all --search "$TITLE in:title" \
                  --json number,title,state \
                  --jq "[.[] | select(.title == \"$TITLE\")] | first | .number // empty")
          BODY=$(printf '%s\n' \
            "The weekly full sweep failed: $RUN_URL" \
            "" \
            "This is the only signal that catches a published package broken by an" \
            "upstream change — selective CI cannot select a member no diff names." \
            "" \
            "Check the \`workspace\` job's failed shards for which members regressed.")
          if [ -n "$NUM" ]; then
            gh issue reopen "$NUM" 2>/dev/null || true
            gh issue comment "$NUM" --body "$BODY"
            echo "commented on #$NUM"
          else
            gh issue create --title "$TITLE" --body "$BODY"
          fi