CI: select members by package identity, not by substring

.github/workflows/validate.yml

@@ -704,6 +704,16 @@ jobs:
           fetch-depth: 0
       - name: Install lua
         run: sudo apt-get install -y --no-install-recommends lua5.4
+      # The dependencies stated by the manifests inside packages' archives,
+      # read by tests/select_members.py. An archive is pinned by its sha256,
+      # so an entry never goes stale; a new key per descriptor set keeps the
+      # entries of the previous one and adds the archives it adds.
+      - name: Restore the archive manifests read before
+        uses: actions/cache@v4
+        with:
+          path: ${{ runner.temp }}/select-manifests.json
+          key: select-manifests-${{ hashFiles('pkgs/**/*.lua') }}
+          restore-keys: select-manifests-
       # ── Selective member testing ──────────────────────────────────────
       # `mcpp test --workspace` builds every member (opencv, ffmpeg, …) and
       # dominates CI wall-clock, while a PR almost always touches one
@@ -830,7 +840,7 @@ jobs:
             changed=""
           fi
           printf 'changed files vs %s:\n%s\n' "${base:-<none>}" "$changed"
-          sel=""; pkgsel=""
+          sel=""; pkgsel=""; descriptors=""
           add() { case " $sel " in *" $1 "*) ;; *) sel="$sel $1" ;; esac; }
           while IFS= read -r f; do
             [ -n "$f" ] || continue
@@ -856,15 +866,11 @@ jobs:
                 if [ -d "tests/examples/$m" ] || [ -d "tests/mwrap/$m" ]; then add "$m"; fi ;;
               pkgs/*.lua|pkgs/*/*.lua)
                 # A DELETED descriptor is in the diff but not on disk, and its
-                # CN urls are gone with it — nothing left to reach.
+                # CN urls are gone with it — nothing left to reach. The members
+                # it reaches are selected after the loop, all descriptors at
+                # once (tests/select_members.py).
                 if [ -f "$f" ]; then pkgsel="$pkgsel $f"; fi
-                lib=$(basename "$f" .lua); lib=${lib##*.}
-                hit=0
-                for mt in tests/examples/*/mcpp.toml tests/mwrap/*/mcpp.toml; do
-                  [ -f "$mt" ] || continue
-                  if grep -Fq "$lib" "$mt"; then add "$(basename "$(dirname "$mt")")"; hit=1; fi
-                done
-                [ "$hit" = 1 ] || echo "note: no workspace member exercises $f" ;;
+                descriptors="$descriptors $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
@@ -890,6 +896,18 @@ jobs:
           done <<EOF
           $changed
           EOF
+          # The members a changed descriptor reaches: those whose declared
+          # dependency graph holds its identity, directly or through other
+          # packages of this index (tests/select_members.py says how the
+          # graph is read). It replaced a substring search for the short name,
+          # under which pkgs/o/os.lua selected 113 members and two consumed it.
+          python3 tests/test_select_members.py
+          if [ -n "$descriptors" ]; then
+            for m in $(python3 tests/select_members.py ${base:+--base "$base"} \
+                         --cache "$RUNNER_TEMP/select-manifests.json" $descriptors); do
+              add "$m"
+            done
+          fi
           sel=${sel# }
           echo "MEMBERS=$sel" >> "$GITHUB_ENV"
           echo "selected members: ${sel:-<none>}"

CHANGELOG.md

@@ -16,6 +16,13 @@
 
 ### Changed
 
+- **CI 按包身份选择受影响的成员。** `validate.yml` 的 select 曾在成员清单中查找描述符短名的
+  子串:`pkgs/o/os.lua` 选中 205 个成员中的 113 个(`boost-*`、`macos`、`cfg(os = ...)`、
+  注释里的 `exists`),而经 asio、curl 依赖 `compat.openssl` 的成员一个也未选中。
+  `tests/select_members.py` 以 mcpp 的规则读取成员声明的 `(namespace, name)`,沿描述符的
+  `deps` 与包归档内清单(按 tag 读 raw.githubusercontent.com;release 资产则读归档本身,
+  并经 `path` 依赖到达同一归档的其他清单)陈述的依赖求传递闭包,选出依赖图含被修改身份的成员。
+  归档以 sha256 固定,已读结果由 `actions/cache` 保存。读不到的清单记为警告,由每周全量运行覆盖。
 - **`compat.sycl-runtime` 在 Linux 上要求 `mcpp:c++-abi=libstdc++`。** `libsycl.so` 在 libstdc++
   上抛出 `sycl::exception`,而一个进程中的两套 C++ 运行时之间带类型的 `catch` 不会匹配
   (mcpp-community/mcpp#838)。gcc 工具链,或 llvm 加 `gcc.libstdcxx`,满足该要求;其他构建

tests/descriptor_identities.lua

@@ -0,0 +1,76 @@
+-- Print what each descriptor declares about itself and about its neighbours:
+--
+--     <file> TAB <namespace> TAB <name> TAB <dependency>,... TAB <archive>#<manifest> ...
+--
+-- A dependency is a key of any `deps` table inside `package`, spelled as the
+-- descriptor spells it (`compat.openssl`, `compat:openssl`, `zlib`); the
+-- reader of this output (tests/select_members.py) resolves the spelling
+-- against the identities printed here.
+--
+-- The last column is where the rest of the package's dependencies are stated
+-- when the descriptor does not state them: a package whose `mcpp` is not an
+-- inline table builds from the manifest inside its archive (`mcpp =
+-- "*/ld/mcpp.toml"`, or `mcpp.toml` at the archive's root when the field is
+-- absent). One entry per version: the archive's GLOBAL url, `#`, and the
+-- manifest's path inside it, prefixed with `?` when the field is absent --
+-- the archive may then hold no manifest at all (a package built by xpm, or
+-- one whose install() writes it). A descriptor that does not load prints no line.
+--
+-- Usage: lua5.4 tests/descriptor_identities.lua pkgs/*/*.lua
+
+function import(...)
+    return setmetatable({}, {__index = function() return function() end end})
+end
+
+local function collect(t, out, seen)
+    if seen[t] then return end
+    seen[t] = true
+    for k, v in pairs(t) do
+        if k == "deps" and type(v) == "table" then
+            for dk, dv in pairs(v) do
+                if type(dk) == "string" then out[dk] = true
+                elseif type(dv) == "string" then out[dv] = true end
+            end
+        end
+        if type(v) == "table" then collect(v, out, seen) end
+    end
+end
+
+for i = 1, #arg do
+    package = nil
+    local chunk = loadfile(arg[i], "t")
+    if chunk then
+        -- `package = { ... }` is the first statement, so it is populated even
+        -- if a later top-level line fails under the stubbed `import`.
+        pcall(chunk)
+        local p = package
+        if type(p) == "table" and type(p.name) == "string" then
+            local deps = {}
+            collect(p, deps, {})
+            local list = {}
+            for d in pairs(deps) do list[#list + 1] = d end
+            table.sort(list)
+            local manifests = {}
+            if type(p.mcpp) ~= "table" then
+                local inside = "?mcpp.toml"
+                if type(p.mcpp) == "string" then inside = (p.mcpp:gsub("^%*/", "")) end
+                local seen = {}
+                for _, versions in pairs(type(p.xpm) == "table" and p.xpm or {}) do
+                    for _, entry in pairs(type(versions) == "table" and versions or {}) do
+                        local url = type(entry) == "table" and entry.url
+                        if type(url) == "table" then url = url.GLOBAL end
+                        if type(url) == "string" and not seen[url] then
+                            seen[url] = true
+                            manifests[#manifests + 1] = url .. "#" .. inside
+                        end
+                    end
+                end
+                table.sort(manifests)
+            end
+            print(table.concat({ arg[i],
+                                 type(p.namespace) == "string" and p.namespace or "",
+                                 p.name, table.concat(list, ","),
+                                 table.concat(manifests, " ") }, "\t"))
+        end
+    end
+end