PYTHON-6188 Fix fuzz Regex divergence and replay the seed corpus in PR CI

.evergreen/scripts/run_tests.py

@@ -139,15 +139,26 @@ def handle_fuzz() -> None:
     targets = ["fuzz_decode", "fuzz_decode_all", "fuzz_is_valid"]
 
     if SUB_TEST_NAME == "regression":
+        # Run the committed reproducers plus the seed corpus.
         # Skip dotfiles (pathlib globbing matches them).
         inputs = sorted(
             str(p) for p in (ROOT / "fuzz" / "regression").glob("*") if not p.name.startswith(".")
         )
         if not inputs:
             raise RuntimeError("No fuzz regression inputs found in fuzz/regression/")
-        # Each reproducer is passed as its own file argument, so libFuzzer
-        # runs each exactly once; the -atheris_runs budget equals the file
-        # count so atheris exits cleanly after the last one.
+        corpus_dir = os.environ.get("FUZZ_CORPUS_DIR", str(ROOT / "fuzz-scratch" / "corpus"))
+        # Fail fast on a missing or empty corpus; libFuzzer would silently
+        # skip it.
+        corpus_files = sorted(
+            str(p) for p in Path(corpus_dir).glob("*") if not p.name.startswith(".")
+        )
+        if not corpus_files:
+            raise RuntimeError(
+                f"No fuzz seed corpus found in {corpus_dir}; was setup_tests.py run?"
+            )
+        inputs.extend(corpus_files)
+        # Each input runs exactly once without random fuzzing; the -atheris_runs
+        # budget equals the file count so atheris exits cleanly after the last.
         runs = len(inputs)
         for target in targets:
             run_command(

.evergreen/scripts/setup_tests.py

@@ -530,8 +530,9 @@ def handle_test_env() -> None:
         else:
             TEST_ARGS = f"test/performance/async_perf_test.py {TEST_ARGS}"
 
-    if test_name == "fuzz" and sub_test_name != "regression":
-        # Build the seed corpus from the vendored spec data.
+    if test_name == "fuzz":
+        # Build the seed corpus from the vendored spec data. Regression runs
+        # run it too.
         scratch_dir = ROOT / "fuzz-scratch"
         shutil.rmtree(scratch_dir, ignore_errors=True)
         corpus_dir = scratch_dir / "corpus"

fuzz/README.md

@@ -47,6 +47,9 @@ just setup-tests fuzz regression
 just run-tests
 ```
 
+On every PR, the `fuzz-regression` task runs the committed reproducers and the full generated seed
+corpus.
+
 Or without the justfile:
 
 ```bash
@@ -114,5 +117,5 @@ fuzz_decode.py      Fuzz target: bson.decode()
 fuzz_decode_all.py  Fuzz target: bson.decode_all()
 fuzz_is_valid.py    Fuzz target: bson.is_valid()
 fuzz_shared.py      Code shared by the fuzz targets
-regression/         Known bug reproducers (no corpus setup required)
+regression/         Known bug reproducers; run with the seed corpus on PRs
 ```

fuzz/fuzz_shared.py

@@ -57,6 +57,10 @@ def normalize(obj: object) -> object:
         return [normalize(v) for v in obj]
     if isinstance(obj, float) and math.isnan(obj):
         return "nan"
+    if type(obj).__name__ == "Regex":
+        # int() canonicalizes the C extension's plain int flags and pure
+        # Python's re.RegexFlag.
+        return f"Regex:{obj.pattern!r}:{int(obj.flags)}"
     return f"{type(obj).__name__}:{obj!r}"