| 1 | #!/usr/bin/env python3
|
| 2 | """Differential merge-correctness suite: git merge-ort (oracle) vs Oak's merge engine.
|
| 3 |
|
| 4 | stdout: diagnostic JSONL (one row per directional case), then nothing else.
|
| 5 | --summary FILE: JSON summary (counts per class, merge-law rates, oak-wrong list).
|
| 6 | Exit: 0 when no case is oak-wrong or harness-error, 1 otherwise, 2 on usage error.
|
| 7 | Rows are diagnostic evidence (evidence_class diagnostic_merge_correctness); never
|
| 8 | ingest them into performance reports. See suites/merge/README.md.
|
| 9 | """
|
| 10 | from __future__ import annotations
|
| 11 |
|
| 12 | import argparse
|
| 13 | import fnmatch
|
| 14 | import json
|
| 15 | from pathlib import Path
|
| 16 | import shutil
|
| 17 | import sys
|
| 18 |
|
| 19 | HERE = Path(__file__).resolve().parent
|
| 20 | sys.path.insert(0, str(HERE))
|
| 21 | sys.path.insert(0, str(HERE.parents[1] / "scripts"))
|
| 22 |
|
| 23 | from mergesuite import sources # noqa: E402
|
| 24 | from mergesuite.runner import SCHEMA_VERSION, Suite, stale_schema_rows, summarize # noqa: E402
|
| 25 |
|
| 26 | GROUPS = ("gix-tree", "gix-cartesian", "gix-text", "t64xx", "native", "generated")
|
| 27 |
|
| 28 |
|
| 29 | def resolve_git(explicit: str | None) -> str:
|
| 30 | if explicit:
|
| 31 | return explicit
|
| 32 | try:
|
| 33 | from oakbench.subjects import which_measured_git
|
| 34 | found = which_measured_git()
|
| 35 | except Exception: # pragma: no cover - oakbench import is best effort
|
| 36 | found = shutil.which("git")
|
| 37 | if not found:
|
| 38 | raise SystemExit("git not found")
|
| 39 | return found
|
| 40 |
|
| 41 |
|
| 42 | def load(groups, workdir: Path, git_bin: str, home: Path, gen_count: int, gen_seed: int):
|
| 43 | scenarios = []
|
| 44 | if "gix-tree" in groups:
|
| 45 | scenarios += sources.gitoxide_tree(workdir, git_bin, home)
|
| 46 | if "gix-cartesian" in groups:
|
| 47 | scenarios += sources.gitoxide_cartesian(workdir, git_bin, home)
|
| 48 | if "gix-text" in groups:
|
| 49 | scenarios += sources.gitoxide_text(workdir, git_bin, home)
|
| 50 | if "t64xx" in groups:
|
| 51 | scenarios += sources.git_t64xx(workdir, git_bin, home)
|
| 52 | if "native" in groups:
|
| 53 | scenarios += sources.native_incidents(workdir, git_bin, home)
|
| 54 | if "generated" in groups:
|
| 55 | scenarios += sources.generated(workdir, git_bin, home, gen_count, gen_seed)
|
| 56 | return scenarios
|
| 57 |
|
| 58 |
|
| 59 | def main(argv=None) -> int:
|
| 60 | ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
|
| 61 | ap.add_argument("--workdir", type=Path, required=True, help="scratch dir (fixtures, oak serve data, cases)")
|
| 62 | ap.add_argument("--oak-bin", default="oak")
|
| 63 | ap.add_argument("--git-bin", default=None, help="oracle git (default: real git, xcrun shim resolved)")
|
| 64 | ap.add_argument("--groups", default=",".join(GROUPS), help="comma list of " + ",".join(GROUPS))
|
| 65 | ap.add_argument("--only", default=None, help="fnmatch pattern on scenario id")
|
| 66 | ap.add_argument("--limit", type=int, default=None)
|
| 67 | ap.add_argument("--jobs", type=int, default=4)
|
| 68 | ap.add_argument("--generated-count", type=int, default=40)
|
| 69 | ap.add_argument("--generated-seed", type=int, default=20260928)
|
| 70 | ap.add_argument("--keep-equal", action="store_true", help="keep case dirs of equal-to-git cases")
|
| 71 | ap.add_argument("--summary", type=Path, default=None)
|
| 72 | ap.add_argument("--list", action="store_true", help="list scenario ids and exit")
|
| 73 | ap.add_argument("--resummarize", type=Path, default=None, metavar="JSONL",
|
| 74 | help="recompute the summary (known-bug mapping included) from an existing JSONL run; no Oak")
|
| 75 | ap.add_argument("--allow-stale-schema", action="store_true",
|
| 76 | help="with --resummarize: summarize rows written under an older row/reason schema anyway "
|
| 77 | "(known-bug statuses become 'indeterminate'); default refuses")
|
| 78 | args = ap.parse_args(argv)
|
| 79 | groups = [g for g in args.groups.split(",") if g]
|
| 80 | if not set(groups) <= set(GROUPS):
|
| 81 | ap.error("unknown group")
|
| 82 | if args.resummarize:
|
| 83 | rows = [json.loads(line) for line in args.resummarize.read_text().splitlines() if line.strip()]
|
| 84 | if not rows:
|
| 85 | print(f"resummarize: {args.resummarize} holds no rows; nothing to summarize", file=sys.stderr)
|
| 86 | return 2
|
| 87 | stale = stale_schema_rows(rows)
|
| 88 | if stale:
|
| 89 | newer = [v for v in stale if v != "missing" and v.isdigit() and int(v) > SCHEMA_VERSION]
|
| 90 | what = ("were written by a NEWER merge suite than this one, not on" if newer and len(newer) == len(stale)
|
| 91 | else "predate" if not newer else "do not match")
|
| 92 | print(f"resummarize: {sum(stale.values())}/{len(rows)} row(s) {what} merge-suite row/reason schema "
|
| 93 | f"v{SCHEMA_VERSION} (found {stale}); their reasons cannot be mapped to the current known-bug "
|
| 94 | "patterns, so a missing bug would read as 'fixed'. Re-run the suite"
|
| 95 | + (" (--allow-stale-schema given: statuses reported as indeterminate)."
|
| 96 | if args.allow_stale_schema else ", or pass --allow-stale-schema."), file=sys.stderr)
|
| 97 | if not args.allow_stale_schema:
|
| 98 | return 2
|
| 99 | summary = finish_summary(summarize(rows), rows)
|
| 100 | if args.summary:
|
| 101 | args.summary.write_text(json.dumps(summary, indent=2, sort_keys=True) + "\n")
|
| 102 | print(format_table(summary), file=sys.stderr)
|
| 103 | return 1 if summary["counts"]["oak-wrong"] or summary["counts"]["harness-error"] else 0
|
| 104 | git_bin = resolve_git(args.git_bin)
|
| 105 | workdir = args.workdir.resolve()
|
| 106 | workdir.mkdir(parents=True, exist_ok=True)
|
| 107 | suite = Suite(workdir, args.oak_bin, git_bin, jobs=args.jobs, keep_equal=args.keep_equal)
|
| 108 | scenarios = load(groups, workdir / "fixtures-src", git_bin, suite.git_home,
|
| 109 | args.generated_count, args.generated_seed)
|
| 110 | if args.only:
|
| 111 | scenarios = [s for s in scenarios if fnmatch.fnmatch(s["id"], args.only)]
|
| 112 | if args.limit:
|
| 113 | scenarios = scenarios[:args.limit]
|
| 114 | if args.list:
|
| 115 | for s in scenarios:
|
| 116 | print(s["id"] + (" [unsupported: %s]" % s["unsupported"] if s["unsupported"] else ""))
|
| 117 | return 0
|
| 118 |
|
| 119 | def emit(row):
|
| 120 | print(json.dumps(row, sort_keys=True), flush=True)
|
| 121 |
|
| 122 | rows = suite.run(scenarios, emit)
|
| 123 | summary = finish_summary(summarize(rows), rows)
|
| 124 | summary["meta"] = suite.meta
|
| 125 | summary["run_id"] = suite.run_id
|
| 126 | if args.summary:
|
| 127 | args.summary.write_text(json.dumps(summary, indent=2, sort_keys=True) + "\n")
|
| 128 | print(format_table(summary), file=sys.stderr)
|
| 129 | return 1 if summary["counts"]["oak-wrong"] or summary["counts"]["harness-error"] else 0
|
| 130 |
|
| 131 |
|
| 132 | def finish_summary(summary: dict, rows) -> dict:
|
| 133 | summary["phase2_specs_not_executed"] = [s["id"] for s in sources.PHASE2_SPECS]
|
| 134 | summary["oak_wrong"] = [{"scenario": r["scenario"], "direction": r["direction"], "reasons": r["reasons"],
|
| 135 | "known_bug": r.get("known_bug"),
|
| 136 | "case_dir": r.get("case_dir"), "repro": r["repro"]}
|
| 137 | for r in rows if r["class"] == "oak-wrong"]
|
| 138 | summary["harness_errors"] = [{"scenario": r["scenario"], "direction": r["direction"], "reasons": r["reasons"]}
|
| 139 | for r in rows if r["class"] == "harness-error"]
|
| 140 | return summary
|
| 141 |
|
| 142 |
|
| 143 | def format_table(summary: dict) -> str:
|
| 144 | classes = list(summary["counts"])
|
| 145 | lines = ["| origin | " + " | ".join(classes) + " |", "|---|" + "---|" * len(classes)]
|
| 146 | for origin, counts in sorted(summary["by_origin"].items()):
|
| 147 | lines.append(f"| {origin} | " + " | ".join(str(counts[c]) for c in classes) + " |")
|
| 148 | lines.append("| **total** | " + " | ".join(str(summary["counts"][c]) for c in classes) + " |")
|
| 149 | lines.append("")
|
| 150 | for law, v in summary["laws"].items():
|
| 151 | lines.append(f"{law}: {v['true']}/{v['measured']}")
|
| 152 | lines.append("")
|
| 153 | for bug, info in (summary.get("known_bugs") or {}).items():
|
| 154 | lines.append(f"{bug} ({info['family']}): {info['rows']} row(s), pinned {info['expected_rows']} - "
|
| 155 | f"{info['status']}; new={len(info['new_rows'])} vanished={len(info['vanished_rows'])}")
|
| 156 | if "oak_wrong_unmapped" in summary:
|
| 157 | lines.append(f"oak-wrong rows not mapped to a known bug: {summary['oak_wrong_unmapped']}")
|
| 158 | lines.append(f"failed pulls violating 'failed sync => worktree matches HEAD' (HEAD moved or worktree "
|
| 159 | f"left dirty): {summary['post_failure_hazard_rows']}")
|
| 160 | lines.append(f"missed conflicts by class: {summary['missed_conflict_by_class']}")
|
| 161 | lines.append(f"missed conflicts by git kind: {summary['missed_conflict_by_git_kind']}")
|
| 162 | return "\n".join(lines)
|
| 163 |
|
| 164 |
|
| 165 | if __name__ == "__main__":
|
| 166 | raise SystemExit(main())
|