Fix macOS CI job processing in scripts and tests

Update ci-e2e-bundle.py to include e2e-macos in MATRIX_JOBS so macOS
test binaries are bundled. Update ci-vlt-proof-suites.py and all test
scripts (test_ci_vlt_rows.py, test_ci_e2e_tiers.py, test_vlt_coverage.py)
to read from the new macOS-specific jobs (e2e-macos, cargo-vex-matrix-macos,
install-proof-macos) alongside the original jobs.

Author: cursoragent

scripts/ci-e2e-bundle.py

@@ -29,7 +29,7 @@
 ROOT = Path(__file__).resolve().parents[1]
 CI = ROOT / ".github" / "workflows" / "ci.yml"
 TESTS = ROOT / "crates" / "socket-patch-cli" / "tests"
-MATRIX_JOBS = ("e2e", "e2e-full")
+MATRIX_JOBS = ("e2e", "e2e-full", "e2e-macos")
 # The binaries cargo-vex-matrix and cargo-vex-matrix-full run on every OS.
 CARGO_VEX_SUITES = (
     "e2e_redirect_cargo_build",

scripts/ci-vlt-proof-suites.py

@@ -39,7 +39,8 @@ def proof_upgrade(vlt, node):
 
 def ci_cells(text=None):
     reader = load_reader()
-    rows = reader.matrix_include(reader.jobs(text if text is not None else CI.read_text(encoding="utf-8"))["e2e"])
+    jobs = reader.jobs(text if text is not None else CI.read_text(encoding="utf-8"))
+    rows = reader.matrix_include(jobs["e2e"]) + reader.matrix_include(jobs["e2e-macos"])
     return {(r["suite"], r["os"], r["vlt"], r.get("vlt_store_linker", ""), r.get("vlt_upgrade", ""))
             for r in rows if r.get("vlt") and r.get("test_filter") == "--include-ignored vlt_pinned_matrix"}
 

scripts/tests/test_ci_e2e_tiers.py

@@ -42,7 +42,7 @@ class Tiers(unittest.TestCase):
     def test_every_row_is_a_test_target(self):
         # `suite` may list several binaries; an `allow_empty` row may name
         # suites that have not landed yet (test_ci_gradle_prefixes.py).
-        for job in ("e2e", "e2e-full"):
+        for job in ("e2e", "e2e-full", "e2e-macos"):
             for row in rows(job):
                 for suite in bundle.row_suites(row):
                     with self.subTest(job=job, row=row, suite=suite):
@@ -78,15 +78,16 @@ def test_nightly_schedule_runs_the_full_tier(self):
 
     def test_cargo_cross_is_split_exactly(self):
         pr = [(r["os"], r["toolchain"], r.get("lock", "")) for r in rows("cargo-vex-matrix")]
+        macos = [(r["os"], r["toolchain"], r.get("lock", "")) for r in rows("cargo-vex-matrix-macos")]
         full = [(r["os"], r["toolchain"], r.get("lock", "")) for r in rows("cargo-vex-matrix-full")]
         want = {("ubuntu-latest", t, l) for t, l in itertools.product(("1.82.0", "1.93.1", "stable"),
                                                                     ("", "1", "2", "3", "4"))}
         want |= {("macos-latest", "stable", "1"), ("windows-latest", "stable", "1"),
                  ("macos-latest", "1.93.1", ""), ("windows-latest", "1.93.1", "")}
-        self.assertEqual(len(pr + full), len(want))
-        self.assertEqual(set(pr) | set(full), want)
+        self.assertEqual(len(pr + macos + full), len(want))
+        self.assertEqual(set(pr) | set(macos) | set(full), want)
         for os_name in ("ubuntu-latest", "macos-latest", "windows-latest"):
-            self.assertIn((os_name, "1.93.1", ""), pr, "the pinned toolchain's own lock on every OS")
+            self.assertIn((os_name, "1.93.1", ""), pr + macos, "the pinned toolchain's own lock on every OS")
         ubuntu = [c for c in pr if c[0] == "ubuntu-latest"]
         self.assertEqual({c[1] for c in ubuntu}, {"1.82.0", "1.93.1", "stable"}, "every toolchain on PRs")
         self.assertEqual({c[2] for c in ubuntu}, {"", "1", "2", "3", "4"}, "every lock on PRs")
@@ -100,7 +101,7 @@ def test_bundle_covers_every_os(self):
         for os_name in ("ubuntu-latest", "macos-latest", "windows-latest"):
             with self.subTest(os=os_name):
                 suites = bundle.suites_for(os_name, TEXT)
-                for job in ("e2e", "e2e-full"):
+                for job in ("e2e", "e2e-full", "e2e-macos"):
                     for row in rows(job):
                         if row["os"] == os_name:
                             for suite in bundle.row_suites(row):
@@ -268,11 +269,13 @@ class PdmCapstone(unittest.TestCase):
     def test_excludes_exactly_the_cells_ci_runs_on_every_pr(self):
         excluded = rows_mod.matrix_include([l.replace("exclude:", "include:") for l in self.job])
         ci = {(r["os"], r["pdm"]) for r in rows("e2e") if "pdm" in r}
-        self.assertEqual({(r["os"], r["pdm"]) for r in excluded}, ci)
-        self.assertEqual(len(excluded), len(ci))
-        for row in rows("e2e"):
-            if "pdm" in row:
-                self.assertEqual(row.get("test_filter"), "pdm:: --ignored")
+        ci |= {(r["os"], r["pdm"]) for r in rows("e2e-macos") if "pdm" in r}
+        self.assertEqual({(r["os"], r["pdm"]) for r in excluded if "pdm" in r}, ci)
+        self.assertEqual(len([r for r in excluded if "pdm" in r]), len(ci))
+        for job in ("e2e", "e2e-macos"):
+            for row in rows(job):
+                if "pdm" in row:
+                    self.assertEqual(row.get("test_filter"), "pdm:: --ignored")
         self.assertFalse(any("pdm" in r for r in rows("e2e-full")),
                          "a pdm row off the PR tier would leave its cell unrun on PRs")
         versions = re.search(r"pdm: \[([^\]]*)\]", "\n".join(self.job)).group(1)

scripts/tests/test_ci_vlt_rows.py

@@ -148,7 +148,7 @@ def step(job_lines, name):
 
 class CiE2eVltRows(unittest.TestCase):
     ci = jobs(CI.read_text(encoding="utf-8"))
-    rows = matrix_include(ci["e2e"])
+    rows = matrix_include(ci["e2e"]) + matrix_include(ci["e2e-macos"])
     vlt_rows = [r for r in rows if "vlt" in r or "vlt" in r.get("suite", "")]
 
     def test_every_vlt_row_pins_a_release_and_includes_the_ignored_legs(self):
@@ -241,7 +241,7 @@ def test_hosted_e2e_proves_vlt_against_production(self):
 
 class CompatibilityWorkflow(unittest.TestCase):
     compat = jobs(COMPAT.read_text(encoding="utf-8"))
-    rows = matrix_include(compat["install-proof"])
+    rows = matrix_include(compat["install-proof"]) + matrix_include(compat["install-proof-macos"])
 
     def covered(self):
         cells = set()
@@ -287,7 +287,7 @@ def test_node_floors_and_store_linkers(self):
                       step(self.compat["install-proof"], "Collation golden under this Node"))
 
     def test_jobs_and_triggers(self):
-        for job in ("build", "install-proof", "native", "lock-diff", "canary", "downgrade"):
+        for job in ("build", "install-proof", "install-proof-macos", "native", "lock-diff", "canary", "downgrade"):
             self.assertIn(job, self.compat)
         text = COMPAT.read_text(encoding="utf-8")
         self.assertIn("schedule:", text)

scripts/tests/test_vlt_coverage.py

@@ -119,8 +119,12 @@ def ci_ids(self):
             (ROOT / ".github/workflows/vlt-compatibility.yml").read_text(encoding="utf-8"))
         ids = {f"ci:{r['suite']}:{r['os']}:{r['vlt']}"
                for r in self.rows.matrix_include(ci["e2e"]) if "vlt" in r}
+        ids |= {f"ci:{r['suite']}:{r['os']}:{r['vlt']}"
+                for r in self.rows.matrix_include(ci["e2e-macos"]) if "vlt" in r}
         ids |= {f"compat:{r['os']}:{r['vlt']}"
                 for r in self.rows.matrix_include(compat["install-proof"])}
+        ids |= {f"compat:{r['os']}:{r['vlt']}"
+                for r in self.rows.matrix_include(compat["install-proof-macos"])}
         return ids
 
     def test_every_code_is_mapped_and_documented(self):