.github/workflows/wai-pack.yml

@@ -47,8 +47,41 @@ on:
       - '.github/workflows/wai-pack.yml'
 
 jobs:
+  # Forgejo matches the `paths:` filters above against the **push delta**, not
+  # the pull request's diff. A rebase moves a branch across every commit main
+  # has gained, so its delta touches nearly every path and this workflow fires
+  # on changes that cannot affect it. Measured on this repo: a jcp-only change
+  # pushed to a fresh branch ran 2 workflows; the same change force-pushed after
+  # a rebase ran 15, this one among them, putting three platforms of wai-pack and
+  # the JPEG-AI whole decode in a queue they had no business being in.
+  #
+  # So the filters above still gate the *trigger*, and this job gates the
+  # *work*, from the pull request's own diff. It reads the path list out of this
+  # file, so there is one list and not a third copy to drift. On a push to main
+  # there is no pull-request context, the shas are empty, and the gate answers
+  # `true` — a push to main runs everything, as before.
+  relevant:
+    name: wai-pack · does this change touch it
+    runs-on: linux
+    outputs:
+      run: ${{ steps.gate.outputs.run }}
+    steps:
+      - uses: actions/checkout@v5
+        with:
+          fetch-depth: 0
+      - id: gate
+        run: |
+          set -euo pipefail
+          r=$(python3 scripts/pr-touches-paths.py .github/workflows/wai-pack.yml \
+                "${{ github.event.pull_request.base.sha }}" \
+                "${{ github.event.pull_request.head.sha }}")
+          echo "this change touches wai-pack: $r"
+          echo "run=$r" >> "$GITHUB_OUTPUT"
+
   legs:
     name: wai-pack · ${{ matrix.os }}
+    needs: relevant
+    if: needs.relevant.outputs.run == 'true'
     runs-on: ${{ matrix.os }}
     strategy:
       fail-fast: false
@@ -112,6 +145,8 @@ jobs:
 
   independent:
     name: wai-pack · independent verifier (Python)
+    needs: relevant
+    if: needs.relevant.outputs.run == 'true'
     runs-on: linux
     steps:
       - uses: actions/checkout@v5
@@ -167,6 +202,8 @@ jobs:
 
   media:
     name: wai-pack · the floor through a media prober
+    needs: relevant
+    if: needs.relevant.outputs.run == 'true'
     runs-on: linux
     steps:
       - uses: actions/checkout@v5
@@ -198,6 +235,8 @@ jobs:
 
   licences:
     name: wai-pack · dependency licences
+    needs: relevant
+    if: needs.relevant.outputs.run == 'true'
     runs-on: linux
     steps:
       - uses: actions/checkout@v5

.github/workflows/workflow-hygiene.yml

@@ -2,8 +2,8 @@ name: workflow hygiene
 
 # Properties of the workflows themselves, checked rather than remembered.
 #
-# Today that is one property: every workflow declares a concurrency group that
-# varies by ref. None of them did until 2026-10-05, and the consequence was not
+Two properties today. The first: every workflow declares a concurrency group
+# that varies by ref. None of them did until 2026-10-05, and the consequence was not
 # a slow build — it was that CI stopped being usable for anyone. A session
 # pushing four versions of a branch queued four complete sets of runs, all of
 # which had to drain; one PR's 21 contexts settled none in 70 minutes behind a
@@ -12,6 +12,15 @@ name: workflow hygiene
 # whole repo, invisibly: the workflow still passes, it just queues forever
 # under load.
 #
+# **The path gate.** `scripts/pr-touches-paths.py` decides whether a
+# path-filtered workflow's jobs do any work, from the pull request's own diff —
+# because Forgejo matches a `paths:` filter against the push delta, so a rebase
+# fires every path-filtered workflow in the repo (measured: a jcp-only change
+# ran 2 workflows pushed fresh and 15 force-pushed after a rebase). A gate that
+# decides whether tests run is load-bearing in the dangerous direction: a wrong
+# `false` silently skips a workflow. Its self-test pins both directions and that
+# every way of failing fails *open*.
+#
 # This job runs on `linux`, which is jetson-hub — deliberately not the dev Mac,
 # which already serves 13 of this repo's 21 contexts through two runners. It
 # needs no toolchain and finishes in seconds.
@@ -23,6 +32,9 @@ on:
     paths:
       - '.github/workflows/**'
       - 'scripts/check-workflow-concurrency.sh'
+      - 'scripts/check-workflow-concurrency-selftest.sh'
+      - 'scripts/pr-touches-paths.py'
+      - 'scripts/pr-touches-paths-selftest.sh'
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}
@@ -45,3 +57,8 @@ jobs:
 
       - name: Every workflow is grouped per ref
         run: ./scripts/check-workflow-concurrency.sh
+
+      # The path gate fails open in every failure mode, and still answers
+      # `false` for a diff that cannot affect the workflow.
+      - name: The path gate fails open (self-test)
+        run: ./scripts/pr-touches-paths-selftest.sh

scripts/pr-touches-paths-selftest.sh

@@ -0,0 +1,70 @@
+#!/usr/bin/env bash
+# Self-test for pr-touches-paths.py, run by the workflow-hygiene job.
+#
+# A gate that decides whether tests run is itself load-bearing: a wrong `false`
+# silently skips a whole workflow. These cases pin both directions and, most
+# importantly, that every way of failing fails OPEN.
+set -euo pipefail
+cd "$(dirname "$0")/.."
+G=scripts/pr-touches-paths.py
+tmp=$(mktemp -d); trap 'rm -rf "$tmp"' EXIT
+fail=0
+check() { # name expected actual
+  if [ "$2" = "$3" ]; then printf '  ok    %s\n' "$1"
+  else printf '  FAIL  %s: expected %s, got %s\n' "$1" "$2" "$3"; fail=1; fi
+}
+
+# A throwaway repo, so the diffs are real git diffs and not a mock.
+repo="$tmp/r"; mkdir -p "$repo/wai/wai-pack" "$repo/jcp"
+git -C "$repo" init -q
+git -C "$repo" config user.email t@e; git -C "$repo" config user.name t
+echo x > "$repo/jcp/SPEC.md"; echo y > "$repo/wai/wai-pack/lib.rs"
+git -C "$repo" add -A; git -C "$repo" commit -qm base
+BASE=$(git -C "$repo" rev-parse HEAD)
+echo changed > "$repo/jcp/SPEC.md"
+git -C "$repo" add -A; git -C "$repo" commit -qm jcp-only
+JCP=$(git -C "$repo" rev-parse HEAD)
+echo changed > "$repo/wai/wai-pack/lib.rs"
+git -C "$repo" add -A; git -C "$repo" commit -qm wai-too
+WAI=$(git -C "$repo" rev-parse HEAD)
+
+wf="$tmp/w.yml"
+cat > "$wf" <<'Y'
+on:
+  pull_request:
+    paths:
+      - 'wai/wai-pack/**'
+Y
+
+run() { ( cd "$repo" && python3 "$OLDPWD/$G" "$1" "$2" "$3" ); }
+check "a jcp-only diff does not touch a wai filter" false "$(run "$wf" "$BASE" "$JCP")"
+check "a diff that touches the filter runs"          true  "$(run "$wf" "$BASE" "$WAI")"
+
+# Every failure mode must fail OPEN.
+check "a missing workflow file"      true "$(run "$tmp/nope.yml" "$BASE" "$JCP")"
+check "a workflow with no filter"    true "$(printf 'on:\n  pull_request:\n' > "$tmp/nofilter.yml"; run "$tmp/nofilter.yml" "$BASE" "$JCP")"
+check "an empty base sha"            true "$(run "$wf" "" "$JCP")"
+check "a sha git cannot resolve"     true "$(run "$wf" "$BASE" deadbeefdeadbeefdeadbeefdeadbeefdeadbeef)"
+check "too few arguments"            true "$( cd "$repo" && python3 "$OLDPWD/$G" "$wf" )"
+check "an empty diff"                true "$(run "$wf" "$JCP" "$JCP")"
+
+# `**` spans separators; `*` does not.
+cat > "$tmp/star.yml" <<'Y'
+on:
+  pull_request:
+    paths:
+      - 'jcp/*.md'
+Y
+# Only the NESTED file changes, so `jcp/*.md` must not match it. (A top-level
+# .md in the same commit would match legitimately and prove nothing — which is
+# how this case failed the first time it was written.)
+mkdir -p "$repo/jcp/sub"; echo z > "$repo/jcp/sub/x.md"
+git -C "$repo" add -A >/dev/null; git -C "$repo" commit -qm sub >/dev/null
+SUB=$(git -C "$repo" rev-parse HEAD)
+check "a single star does not span a separator" false "$(run "$tmp/star.yml" "$WAI" "$SUB")"
+# And it does match a file at that level, so the glob is not simply never matching.
+echo z > "$repo/jcp/top.md"
+git -C "$repo" add -A >/dev/null; git -C "$repo" commit -qm top >/dev/null
+check "a single star matches at its own level" true "$(run "$tmp/star.yml" "$SUB" "$(git -C "$repo" rev-parse HEAD)")"
+
+[ "$fail" = 0 ] && echo "pr-touches-paths self-test: all cases pass" || { echo "pr-touches-paths self-test FAILED"; exit 1; }

scripts/pr-touches-paths.py

@@ -0,0 +1,104 @@
+#!/usr/bin/env python3
+"""Does this pull request's own diff touch any path its workflow filters on?
+
+Forgejo matches a workflow's `paths:` filter against the **push delta**, not the
+pull request's diff. A rebase moves a branch across every commit main has
+gained, so its push delta touches nearly every path and every path-filtered
+workflow fires. Measured on this repo: a jcp-only change pushed to a fresh
+branch ran 2 workflows; the same change force-pushed after a rebase ran 15,
+including JPEG-AI whole decode and wai-pack on three platforms. That is what
+fills the queue, and it delays everyone's real jobs behind irrelevant ones.
+
+This reads the path list out of the workflow file itself — one source of truth,
+rather than a third copy of the list to drift from the two `on:` blocks — and
+compares it against `git diff --name-only BASE...HEAD`.
+
+    pr-touches-paths.py <workflow.yml> <base-sha> <head-sha>
+
+Prints `true` or `false`. **Fails open**: anything it cannot determine prints
+`true`, because a wrong `false` silently skips a test and a wrong `true` only
+costs runner time. A gate that guesses must guess toward running.
+"""
+import re
+import subprocess
+import sys
+from pathlib import Path
+
+
+def paths_of(workflow: Path) -> list[str]:
+    """The union of every `paths:` list under `on:` in this workflow.
+
+    Parsed rather than imported from a library: the runner image has no PyYAML
+    guaranteed, and the shape here is a flat `paths:` followed by `- 'glob'`
+    lines. A list this parser fails to read yields none, which fails open.
+    """
+    globs: list[str] = []
+    in_paths = False
+    for line in workflow.read_text().splitlines():
+        if re.match(r"\s*paths(-ignore)?:\s*$", line):
+            in_paths = True
+            continue
+        if in_paths:
+            m = re.match(r"\s*-\s*'([^']+)'\s*$", line) or re.match(r'\s*-\s*"([^"]+)"\s*$', line)
+            if m:
+                globs.append(m.group(1))
+                continue
+            if line.strip().startswith("- "):
+                globs.append(line.strip()[2:].strip())
+                continue
+            in_paths = False
+    return globs
+
+
+def matches(path: str, glob: str) -> bool:
+    """`glob` in the Actions sense: `**` spans separators, `*` does not."""
+    rx = []
+    i = 0
+    while i < len(glob):
+        if glob.startswith("**", i):
+            rx.append(".*")
+            i += 2
+        elif glob[i] == "*":
+            rx.append("[^/]*")
+            i += 1
+        elif glob[i] == "?":
+            rx.append("[^/]")
+            i += 1
+        else:
+            rx.append(re.escape(glob[i]))
+            i += 1
+    return re.fullmatch("".join(rx), path) is not None
+
+
+def main() -> None:
+    if len(sys.argv) != 4:
+        print("true")  # cannot tell: run it
+        return
+    workflow, base, head = Path(sys.argv[1]), sys.argv[2], sys.argv[3]
+    if not workflow.is_file() or not base or not head:
+        print("true")
+        return
+    globs = paths_of(workflow)
+    if not globs:
+        # No filter, or a shape this parser does not read: the workflow wanted
+        # to run on everything, or we cannot say it did not.
+        print("true")
+        return
+    try:
+        out = subprocess.run(
+            ["git", "diff", "--name-only", f"{base}...{head}"],
+            capture_output=True, text=True, check=True,
+        ).stdout
+    except (subprocess.CalledProcessError, OSError):
+        print("true")
+        return
+    changed = [l for l in out.splitlines() if l.strip()]
+    if not changed:
+        print("true")
+        return
+    hit = any(matches(p, g) for p in changed for g in globs)
+    print("true" if hit else "false")
+
+
+if __name__ == "__main__":
+    main()