| 1 | """Git/Oak committed-evidence parity matrix on real binaries (QA P1 A1/A4).
|
| 2 |
|
| 3 | Every row states the same repository situation for each arm; the arms must
|
| 4 | agree on measured pass/fail. "Unmeasured" (None) is allowed only for a
|
| 5 | missing instrument, never for subject-controlled state or a failing subject
|
| 6 | binary. Ported from the reviewer's stageA/parity.py.
|
| 7 | """
|
| 8 | from __future__ import annotations
|
| 9 |
|
| 10 | import importlib.util
|
| 11 | import os
|
| 12 | import shutil
|
| 13 | import subprocess
|
| 14 | import sys
|
| 15 | import tempfile
|
| 16 | import unittest
|
| 17 | from pathlib import Path
|
| 18 | from unittest import mock
|
| 19 |
|
| 20 | ROOT = Path(__file__).resolve().parents[1]
|
| 21 | sys.path.insert(0, str(ROOT / "scripts"))
|
| 22 |
|
| 23 | from oakbench import workflow_oracles as O # noqa: E402
|
| 24 | from oakbench.subjects import which_measured_git # noqa: E402
|
| 25 |
|
| 26 | GIT = which_measured_git()
|
| 27 | OAK = shutil.which("oak")
|
| 28 |
|
| 29 |
|
| 30 | def reader_python() -> str | None:
|
| 31 | configured = os.environ.get("OAK_BENCH_OAK_READER_PYTHON")
|
| 32 | if configured:
|
| 33 | return configured
|
| 34 | if importlib.util.find_spec("blake3") and importlib.util.find_spec("zstandard"):
|
| 35 | return sys.executable
|
| 36 | return None
|
| 37 |
|
| 38 |
|
| 39 | def run(*command, cwd):
|
| 40 | result = subprocess.run(list(command), cwd=cwd, capture_output=True, text=True,
|
| 41 | env={**os.environ, "OAK_NO_UPDATE_CHECK": "1"})
|
| 42 | if result.returncode:
|
| 43 | raise RuntimeError(f"{command}: {result.stderr}")
|
| 44 |
|
| 45 |
|
| 46 | def commit(kind, repo):
|
| 47 | if kind == "git":
|
| 48 | run(GIT, "add", "-A", cwd=repo)
|
| 49 | run(GIT, "-c", "user.email=a@b", "-c", "user.name=a", "commit", "-qm", "x", "--allow-empty", cwd=repo)
|
| 50 | else:
|
| 51 | run(OAK, "commit", "--no-verify", cwd=repo)
|
| 52 |
|
| 53 |
|
| 54 | def init(kind, repo):
|
| 55 | (repo / "seed.txt").write_text("seed\n")
|
| 56 | run(*([GIT, "init", "-q"] if kind == "git" else [OAK, "init"]), cwd=repo)
|
| 57 | commit(kind, repo)
|
| 58 |
|
| 59 |
|
| 60 | def write_payload(repo, scenario):
|
| 61 | expected = dict(O.COMMITTED_PAYLOADS[scenario])
|
| 62 | if scenario == "sync_pull_dirty":
|
| 63 | expected["work.txt"] = "dirty local edit"
|
| 64 | for relative, marker in expected.items():
|
| 65 | path = repo / relative
|
| 66 | path.parent.mkdir(parents=True, exist_ok=True)
|
| 67 | if path.is_symlink() or path.exists():
|
| 68 | path.unlink()
|
| 69 | path.write_text(marker + "\n")
|
| 70 | os.chmod(path, 0o644)
|
| 71 |
|
| 72 |
|
| 73 | READER = reader_python() # captured before setUp isolates the environment
|
| 74 |
|
| 75 |
|
| 76 | @unittest.skipUnless(GIT and OAK, "real Git and Oak required")
|
| 77 | class OracleParityMatrixTest(unittest.TestCase):
|
| 78 | def setUp(self):
|
| 79 | patcher = mock.patch.dict(os.environ)
|
| 80 | patcher.start()
|
| 81 | self.addCleanup(patcher.stop)
|
| 82 | os.environ.pop("OAK_BENCH_OAK_READER_PYTHON", None)
|
| 83 | os.environ.pop("OAK_BENCH_GIT_INSTRUMENT", None)
|
| 84 | self.tmp = Path(tempfile.mkdtemp(prefix="oakbench-parity-")).resolve()
|
| 85 | self.addCleanup(shutil.rmtree, self.tmp, True)
|
| 86 | self.stubs = self.tmp / "stubs"
|
| 87 | self.stubs.mkdir()
|
| 88 |
|
| 89 | def stub(self, name, command, code):
|
| 90 | path = self.stubs / name
|
| 91 | path.write_text(f'#!/bin/sh\n[ "$1" = {command} ] && {{ echo "simulated oak {command} bug" >&2; exit {code}; }}\n'
|
| 92 | f'exec {OAK} "$@"\n')
|
| 93 | path.chmod(0o755)
|
| 94 | return str(path)
|
| 95 |
|
| 96 | def arms(self):
|
| 97 | yield "git", GIT, None
|
| 98 | yield "oak-export", OAK, None
|
| 99 | if READER:
|
| 100 | yield "oak-reader", OAK, READER
|
| 101 |
|
| 102 | def evaluate(self, scenario, situation, *, oak_binary=None):
|
| 103 | results = {}
|
| 104 | for label, binary, reader in self.arms():
|
| 105 | kind = "git" if label == "git" else "oak"
|
| 106 | repo = Path(tempfile.mkdtemp(dir=self.tmp)).resolve()
|
| 107 | init(kind, repo)
|
| 108 | situation(kind, repo)
|
| 109 | evidence = O.task_evidence(scenario, repo, kind=kind,
|
| 110 | binary=(oak_binary or binary) if kind == "oak" else binary,
|
| 111 | oak_reader_python=reader)
|
| 112 | results[label] = (evidence["workflow_integrity_passed"],
|
| 113 | evidence["workflow_integrity_failure_reason"] or evidence["workflow_integrity_unmeasured_reason"])
|
| 114 | return results
|
| 115 |
|
| 116 | def assert_all(self, results, passed):
|
| 117 | for label, (value, reason) in results.items():
|
| 118 | self.assertIs(value, passed, f"{label}: {reason} (all arms: {results})")
|
| 119 |
|
| 120 | def test_payload_situations_agree_across_arms(self):
|
| 121 | scenario = "sync_push_divergence"
|
| 122 |
|
| 123 | def correct(kind, repo):
|
| 124 | write_payload(repo, scenario)
|
| 125 | commit(kind, repo)
|
| 126 |
|
| 127 | def wrong_committed(kind, repo):
|
| 128 | write_payload(repo, scenario)
|
| 129 | (repo / "notes.txt").write_text("wrong\n")
|
| 130 | commit(kind, repo)
|
| 131 | write_payload(repo, scenario)
|
| 132 |
|
| 133 | def exec_committed(kind, repo):
|
| 134 | write_payload(repo, scenario)
|
| 135 | os.chmod(repo / "notes.txt", 0o755)
|
| 136 | commit(kind, repo)
|
| 137 | os.chmod(repo / "notes.txt", 0o644)
|
| 138 |
|
| 139 | def oversized_committed(kind, repo): # A4: task limit (1 MiB) fails first, as in Git
|
| 140 | write_payload(repo, scenario)
|
| 141 | (repo / "notes.txt").write_bytes(b"upstream change\n" + b"a" * (2 * 1024 * 1024))
|
| 142 | commit(kind, repo)
|
| 143 | write_payload(repo, scenario)
|
| 144 |
|
| 145 | def never_committed(kind, repo):
|
| 146 | write_payload(repo, scenario)
|
| 147 |
|
| 148 | def metadata_destroyed(kind, repo): # A1 case 1
|
| 149 | write_payload(repo, scenario)
|
| 150 | commit(kind, repo)
|
| 151 | shutil.rmtree(repo / (".git" if kind == "git" else ".oak"))
|
| 152 |
|
| 153 | self.assert_all(self.evaluate(scenario, correct), True)
|
| 154 | for situation in (wrong_committed, exec_committed, oversized_committed, never_committed, metadata_destroyed):
|
| 155 | with self.subTest(situation=situation.__name__):
|
| 156 | self.assert_all(self.evaluate(scenario, situation), False)
|
| 157 |
|
| 158 | def test_failing_subject_oak_is_a_measured_failure_in_every_payload_workflow(self):
|
| 159 | # A1 case 2: a released Oak whose export or status exits nonzero must
|
| 160 | # fail the Oak arm, never turn it into an unmeasured skip.
|
| 161 | failing = {"export": self.stub("oak-export-fails", "export", 1),
|
| 162 | "status": self.stub("oak-status-fails", "status", 1)}
|
| 163 | for scenario in sorted(O.COMMITTED_PAYLOADS):
|
| 164 | def correct(kind, repo, scenario=scenario):
|
| 165 | write_payload(repo, scenario)
|
| 166 | commit(kind, repo)
|
| 167 | for command, binary in failing.items():
|
| 168 | with self.subTest(scenario=scenario, command=command):
|
| 169 | repo = Path(tempfile.mkdtemp(dir=self.tmp)).resolve()
|
| 170 | init("oak", repo)
|
| 171 | correct("oak", repo)
|
| 172 | evidence = O.task_evidence(scenario, repo, kind="oak", binary=binary)
|
| 173 | self.assertIs(evidence["workflow_integrity_passed"], False, evidence)
|
| 174 | self.assertIsNone(evidence["workflow_integrity_unmeasured_reason"])
|
| 175 |
|
| 176 | def test_oracle_cli_exit_is_failure_not_sentinel_for_failing_export(self):
|
| 177 | scenario = "sync_pull_upstream"
|
| 178 | repo = Path(tempfile.mkdtemp(dir=self.tmp)).resolve()
|
| 179 | init("oak", repo)
|
| 180 | write_payload(repo, scenario)
|
| 181 | commit("oak", repo)
|
| 182 | binary = self.stub("oak-export-fails", "export", 1)
|
| 183 | result = subprocess.run([sys.executable, "-I", str(ROOT / "scripts/oakbench/workflow_oracles.py"), scenario,
|
| 184 | str(repo), "--oak-head", "--binary", binary], capture_output=True, text=True,
|
| 185 | env={k: v for k, v in os.environ.items() if k != "OAK_BENCH_OAK_READER_PYTHON"})
|
| 186 | self.assertEqual(result.returncode, 1, result.stdout + result.stderr)
|
| 187 | self.assertIn("oak_head_export_failed", result.stdout)
|
| 188 |
|
| 189 | def test_reader_arm_is_exercised_when_configured(self):
|
| 190 | if not READER:
|
| 191 | self.skipTest("optional reader dependencies not configured")
|
| 192 | self.assertIn("oak-reader", [label for label, _, _ in self.arms()])
|
| 193 |
|
| 194 | def test_missing_instrument_is_still_unmeasured(self):
|
| 195 | scenario = "sync_pull_upstream"
|
| 196 | repo = Path(tempfile.mkdtemp(dir=self.tmp)).resolve()
|
| 197 | init("oak", repo)
|
| 198 | write_payload(repo, scenario)
|
| 199 | commit("oak", repo)
|
| 200 | evidence = O.task_evidence(scenario, repo, kind="oak", binary=str(self.tmp / "no-such-oak"))
|
| 201 | self.assertIsNone(evidence["workflow_integrity_passed"])
|
| 202 | evidence = O.task_evidence(scenario, repo, kind="oak", oak_reader_python=str(self.tmp / "no-python"))
|
| 203 | self.assertIsNone(evidence["workflow_integrity_passed"])
|
| 204 |
|
| 205 |
|
| 206 | if __name__ == "__main__":
|
| 207 | unittest.main()
|