diff --git a/python/sglang/srt/debug_utils/comparator/bundle_comparator.py b/python/sglang/srt/debug_utils/comparator/bundle_comparator.py
index 683d28f0d..021990b51 100644
--- a/python/sglang/srt/debug_utils/comparator/bundle_comparator.py
+++ b/python/sglang/srt/debug_utils/comparator/bundle_comparator.py
@@ -315,8 +315,16 @@ def _apply_dim_names_from_meta(
 
 
 def _load_all_values(filenames: list[str], base_path: Path) -> list[ValueWithMeta]:
-    return [
-        item
-        for f in filenames
-        if (item := ValueWithMeta.load(base_path / f)).value is not LOAD_FAILED
-    ]
+    result: list[ValueWithMeta] = []
+    for f in filenames:
+        item: ValueWithMeta = ValueWithMeta.load(base_path / f)
+        if item.value is LOAD_FAILED:
+            warning_sink.add(
+                GeneralWarning(
+                    category="load_failed",
+                    message=f"Failed to load tensor file: {f}",
+                )
+            )
+            continue
+        result.append(item)
+    return result
diff --git a/python/sglang/srt/debug_utils/comparator/display.py b/python/sglang/srt/debug_utils/comparator/display.py
index 56a9f46b7..0ef7e5ab6 100644
--- a/python/sglang/srt/debug_utils/comparator/display.py
+++ b/python/sglang/srt/debug_utils/comparator/display.py
@@ -10,7 +10,7 @@ import polars as pl
 from sglang.srt.debug_utils.comparator.output_types import (
     InputIdsRecord,
     RankInfoRecord,
-    print_record,
+    report_sink,
 )
 from sglang.srt.debug_utils.dump_loader import LOAD_FAILED, ValueWithMeta
 
@@ -23,25 +23,18 @@ def emit_display_records(
     dump_dir: Path,
     label: str,
     tokenizer: Any,
-    output_format: str,
 ) -> None:
     rank_rows: Optional[list[dict[str, Any]]] = _collect_rank_info(
         df, dump_dir=dump_dir
     )
     if rank_rows is not None:
-        print_record(
-            RankInfoRecord(label=label, rows=rank_rows),
-            output_format=output_format,
-        )
+        report_sink.add(RankInfoRecord(label=label, rows=rank_rows))
 
     input_ids_rows: Optional[list[dict[str, Any]]] = _collect_input_ids_and_positions(
         df, dump_dir=dump_dir, tokenizer=tokenizer
     )
     if input_ids_rows is not None:
-        print_record(
-            InputIdsRecord(label=label, rows=input_ids_rows),
-            output_format=output_format,
-        )
+        report_sink.add(InputIdsRecord(label=label, rows=input_ids_rows))
 
 
 def _render_polars_as_text(df: pl.DataFrame, *, title: Optional[str] = None) -> str:
diff --git a/python/sglang/srt/debug_utils/comparator/entrypoint.py b/python/sglang/srt/debug_utils/comparator/entrypoint.py
index ced54e0bc..ce52aa57a 100644
--- a/python/sglang/srt/debug_utils/comparator/entrypoint.py
+++ b/python/sglang/srt/debug_utils/comparator/entrypoint.py
@@ -1,6 +1,8 @@
 from __future__ import annotations
 
 import argparse
+import re
+import sys
 from pathlib import Path
 from typing import Any, Iterator, Optional, Union
 
@@ -29,88 +31,124 @@ from sglang.srt.debug_utils.comparator.output_types import (
     NonTensorRecord,
     SkipRecord,
     SummaryRecord,
-    print_record,
+    report_sink,
+)
+from sglang.srt.debug_utils.comparator.per_token_visualizer import (
+    generate_per_token_heatmap,
 )
 from sglang.srt.debug_utils.comparator.per_token_visualizer import (
     generate_per_token_heatmap,
 )
 from sglang.srt.debug_utils.comparator.utils import Pair
-from sglang.srt.debug_utils.comparator.warning_sink import warning_sink
 from sglang.srt.debug_utils.dump_loader import read_meta, read_tokenizer_path
 
 
 def main() -> None:
     args = _parse_args()
-    run(args)
+    sys.exit(run(args))
 
 
-def run(args: argparse.Namespace) -> None:
-    print_record(
-        ConfigRecord.from_args(args),
+def run(args: argparse.Namespace) -> int:
+    report_path: Optional[Path] = _resolve_report_path(args)
+    report_sink.configure(
         output_format=args.output_format,
+        report_path=report_path,
     )
 
-    warning_sink.set_output_format(args.output_format)
+    try:
+        report_sink.add(ConfigRecord.from_args(args))
 
-    dfs: Pair[pl.DataFrame] = _read_df(args)
+        dfs: Pair[pl.DataFrame] = _read_df(args)
 
-    tokenizer: Any = _maybe_load_tokenizer(args)
-    for label, df, dump_dir in [
-        ("baseline", dfs.x, Path(args.baseline_path)),
-        ("target", dfs.y, Path(args.target_path)),
-    ]:
-        emit_display_records(
-            df=df,
-            dump_dir=dump_dir,
-            label=label,
-            tokenizer=tokenizer,
-            output_format=args.output_format,
+        tokenizer: Any = _maybe_load_tokenizer(args)
+        for label, df, dump_dir in [
+            ("baseline", dfs.x, Path(args.baseline_path)),
+            ("target", dfs.y, Path(args.target_path)),
+        ]:
+            emit_display_records(
+                df=df,
+                dump_dir=dump_dir,
+                label=label,
+                tokenizer=tokenizer,
+            )
+
+        ta_result: TokenAlignerResult = compute_maybe_token_aligner_result(args, dfs)
+
+        if ta_result.mode == "smart":
+            dfs = dfs.map(lambda df: df.filter(~pl.col("name").is_in(AUX_NAMES)))
+
+        bundle_info_pairs: list[Pair[TensorBundleInfo]] = match_bundles(
+            dfs=dfs,
+            skip_keys=_compute_skip_keys(
+                args, has_token_aligner=ta_result.mode is not None
+            ),
         )
 
-    ta_result: TokenAlignerResult = compute_maybe_token_aligner_result(args, dfs)
+        viz_output_dir: Optional[Path] = (
+            Path(args.viz_output_dir) if args.viz_bundle_details else None
+        )
 
-    if ta_result.mode == "smart":
-        dfs = dfs.map(lambda df: df.filter(~pl.col("name").is_in(AUX_NAMES)))
+        visualize_per_token: Optional[Path] = (
+            Path(args.visualize_per_token) if args.visualize_per_token else None
+        )
 
-    bundle_info_pairs: list[Pair[TensorBundleInfo]] = match_bundles(
-        dfs=dfs,
-        skip_keys=_compute_skip_keys(
-            args, has_token_aligner=ta_result.mode is not None
-        ),
-    )
+        meta_overrider: MetaOverrider = MetaOverrider.from_args_and_config(
+            override_dims=args.override_dims,
+            override_baseline_dims=args.override_baseline_dims,
+            override_target_dims=args.override_target_dims,
+            override_config=(
+                Path(args.override_config) if args.override_config else None
+            ),
+        )
 
-    viz_output_dir: Optional[Path] = (
-        Path(args.viz_output_dir) if args.viz_bundle_details else None
-    )
+        comparison_records = _compare_bundle_pairs(
+            bundle_info_pairs=bundle_info_pairs,
+            baseline_path=Path(args.baseline_path),
+            target_path=Path(args.target_path),
+            token_aligner_mode=ta_result.mode,
+            token_aligner_plan=ta_result.plan,
+            diff_threshold=args.diff_threshold,
+            thd_seq_lens_by_step_pair=ta_result.thd_seq_lens_by_step_pair,
+            viz_output_dir=viz_output_dir,
+            compute_per_token=visualize_per_token is not None,
+            meta_overrider=meta_overrider,
+        )
+        summary, skipped_names = _consume_comparison_records(
+            comparison_records=comparison_records,
+            visualize_per_token=visualize_per_token,
+        )
+        return _compute_exit_code(
+            summary,
+            allow_skip_pattern=args.allow_skip_pattern,
+            skipped_names=skipped_names,
+        )
+    finally:
+        report_sink.close()
+        if report_path is not None:
+            print(f"Report: {report_path}", file=sys.stderr)
 
-    visualize_per_token: Optional[Path] = (
-        Path(args.visualize_per_token) if args.visualize_per_token else None
-    )
 
-    meta_overrider: MetaOverrider = MetaOverrider.from_args_and_config(
-        override_dims=args.override_dims,
-        override_baseline_dims=args.override_baseline_dims,
-        override_target_dims=args.override_target_dims,
-        override_config=Path(args.override_config) if args.override_config else None,
-    )
+def _compute_exit_code(
+    summary: SummaryRecord,
+    *,
+    allow_skip_pattern: str,
+    skipped_names: list[str],
+) -> int:
+    if summary.failed > 0:
+        return 1
 
-    comparison_records = _compare_bundle_pairs(
-        bundle_info_pairs=bundle_info_pairs,
-        baseline_path=Path(args.baseline_path),
-        target_path=Path(args.target_path),
-        token_aligner_mode=ta_result.mode,
-        token_aligner_plan=ta_result.plan,
-        diff_threshold=args.diff_threshold,
-        thd_seq_lens_by_step_pair=ta_result.thd_seq_lens_by_step_pair,
-        viz_output_dir=viz_output_dir,
-        compute_per_token=visualize_per_token is not None,
-        meta_overrider=meta_overrider,
-    )
-    _consume_comparison_records(
-        comparison_records=comparison_records,
-        output_format=args.output_format,
-        visualize_per_token=visualize_per_token,
-    )
+    pattern: re.Pattern[str] = re.compile(allow_skip_pattern)
+    forbidden: list[str] = [n for n in skipped_names if not pattern.fullmatch(n)]
+    if forbidden:
+        return 1
+
+    return 0
+
+
+def _resolve_report_path(args: argparse.Namespace) -> Optional[Path]:
+    if args.report_path is not None:
+        return Path(args.report_path) if args.report_path else None
+    return Path(args.target_path) / "comparator_report.jsonl"
 
 
 def _maybe_load_tokenizer(args: argparse.Namespace) -> Any:
@@ -195,22 +233,30 @@ def _compare_bundle_pairs(
 def _consume_comparison_records(
     *,
     comparison_records: Iterator[Union[ComparisonRecord, SkipRecord, NonTensorRecord]],
-    output_format: str,
     visualize_per_token: Optional[Path] = None,
-) -> None:
+) -> tuple[SummaryRecord, list[str]]:
     counts: dict[str, int] = {"passed": 0, "failed": 0, "skipped": 0}
     collected_comparisons: list[ComparisonRecord] = []
+    skipped_names: list[str] = []
 
     for record in comparison_records:
         counts[record.category] += 1
-        print_record(record, output_format=output_format)
+        report_sink.add(record)
+        if isinstance(record, SkipRecord) and record.category == "skipped":
+            skipped_names.append(record.name)
         if visualize_per_token is not None and isinstance(record, ComparisonRecord):
             collected_comparisons.append(record)
 
-    print_record(
-        SummaryRecord(total=sum(counts.values()), **counts),
-        output_format=output_format,
-    )
+    summary: SummaryRecord = SummaryRecord(total=sum(counts.values()), **counts)
+    report_sink.add(summary)
+
+    if visualize_per_token is not None and collected_comparisons:
+        generate_per_token_heatmap(
+            records=collected_comparisons,
+            output_path=visualize_per_token,
+        )
+
+    return summary, skipped_names
 
     if visualize_per_token is not None and collected_comparisons:
         generate_per_token_heatmap(
@@ -300,5 +346,21 @@ def _parse_args() -> argparse.Namespace:
         default=None,
         help="Path to YAML override config file (dims overrides, etc.)",
     )
+    parser.add_argument(
+        "--allow-skip-pattern",
+        type=str,
+        default=".*",
+        help="Regex pattern for tensor names allowed to be skipped. "
+        "Default '.*' allows all skips. Use '^$' to forbid all skips.",
+    )
+
+    # Report output
+    parser.add_argument(
+        "--report-path",
+        type=str,
+        default=None,
+        help="Path for JSONL report (default: <target-path>/comparator_report.jsonl). "
+        "Pass empty string '' to disable.",
+    )
 
     return parser.parse_args()
diff --git a/python/sglang/srt/debug_utils/comparator/output_types.py b/python/sglang/srt/debug_utils/comparator/output_types.py
index c83db102e..f75179276 100644
--- a/python/sglang/srt/debug_utils/comparator/output_types.py
+++ b/python/sglang/srt/debug_utils/comparator/output_types.py
@@ -1,7 +1,9 @@
 from __future__ import annotations
 
+import sys
 from abc import abstractmethod
-from typing import TYPE_CHECKING, Annotated, Any, Literal, Optional, Union
+from pathlib import Path
+from typing import IO, TYPE_CHECKING, Annotated, Any, Literal, Optional, Union
 
 import polars as pl
 from pydantic import ConfigDict, Discriminator, Field, TypeAdapter, model_validator
@@ -253,8 +255,62 @@ def parse_record_json(json_str: str | bytes) -> AnyRecord:
     return _get_any_record_adapter().validate_json(json_str)
 
 
-def print_record(record: _OutputRecord, output_format: str) -> None:
+def _print_to_stdout(record: _OutputRecord, *, output_format: str) -> None:
     if output_format == "json":
         print(record.model_dump_json())
     else:
         print(record.to_text())
+
+
+class ReportSink:
+    """Unified entry point for all record output."""
+
+    def __init__(self) -> None:
+        self._output_format: str = "text"
+        self._report_file: Optional[IO[str]] = None
+        self._report_path: Optional[Path] = None
+
+    def configure(
+        self,
+        *,
+        output_format: str = "text",
+        report_path: Optional[Path] = None,
+    ) -> None:
+        self._output_format = output_format
+
+        if report_path is not None:
+            try:
+                report_path.parent.mkdir(parents=True, exist_ok=True)
+                self._report_file = open(report_path, "w", encoding="utf-8")
+                self._report_path = report_path
+            except OSError as exc:
+                print(
+                    f"Warning: cannot open report file {report_path}: {exc}",
+                    file=sys.stderr,
+                )
+
+    def add(self, record: _OutputRecord) -> None:
+        _print_to_stdout(record, output_format=self._output_format)
+
+        if self._report_file is not None:
+            self._report_file.write(record.model_dump_json())
+            self._report_file.write("\n")
+            self._report_file.flush()
+
+    def close(self) -> None:
+        if self._report_file is not None:
+            self._report_file.close()
+            self._report_file = None
+
+    @property
+    def report_path(self) -> Optional[Path]:
+        return self._report_path
+
+    def _reset(self) -> None:
+        """Reset state for test isolation."""
+        self.close()
+        self._output_format = "text"
+        self._report_path = None
+
+
+report_sink = ReportSink()
diff --git a/python/sglang/srt/debug_utils/comparator/warning_sink.py b/python/sglang/srt/debug_utils/comparator/warning_sink.py
index 1c443f8a6..a5e106f42 100644
--- a/python/sglang/srt/debug_utils/comparator/warning_sink.py
+++ b/python/sglang/srt/debug_utils/comparator/warning_sink.py
@@ -9,10 +9,6 @@ from sglang.srt.debug_utils.comparator.output_types import AnyWarning
 class WarningSink:
     def __init__(self) -> None:
         self._stack: list[list[AnyWarning]] = []
-        self._output_format: str = "text"
-
-    def set_output_format(self, output_format: str) -> None:
-        self._output_format = output_format
 
     @contextmanager
     def context(self) -> Generator[list[AnyWarning], None, None]:
@@ -30,13 +26,10 @@ class WarningSink:
         else:
             from sglang.srt.debug_utils.comparator.output_types import (
                 WarningRecord,
-                print_record,
+                report_sink,
             )
 
-            print_record(
-                WarningRecord(warnings=[warning]),
-                output_format=self._output_format,
-            )
+            report_sink.add(WarningRecord(warnings=[warning]))
 
 
 warning_sink = WarningSink()
diff --git a/test/registered/debug_utils/comparator/conftest.py b/test/registered/debug_utils/comparator/conftest.py
index e69de29bb..f356008e3 100644
--- a/test/registered/debug_utils/comparator/conftest.py
+++ b/test/registered/debug_utils/comparator/conftest.py
@@ -0,0 +1,9 @@
+import pytest
+
+from sglang.srt.debug_utils.comparator.output_types import report_sink
+
+
+@pytest.fixture(autouse=True)
+def _reset_report_sink() -> None:
+    yield
+    report_sink._reset()
diff --git a/test/registered/debug_utils/comparator/test_bundle_comparator.py b/test/registered/debug_utils/comparator/test_bundle_comparator.py
new file mode 100644
index 000000000..f9491b9c8
--- /dev/null
+++ b/test/registered/debug_utils/comparator/test_bundle_comparator.py
@@ -0,0 +1,90 @@
+import sys
+from pathlib import Path
+from unittest.mock import patch
+
+import pytest
+import torch
+
+from sglang.srt.debug_utils.comparator.bundle_comparator import _load_all_values
+from sglang.srt.debug_utils.comparator.output_types import GeneralWarning
+from sglang.srt.debug_utils.comparator.warning_sink import WarningSink
+from sglang.test.ci.ci_register import register_cpu_ci
+
+register_cpu_ci(est_time=15, suite="default", nightly=True)
+
+
+def _save_tensor(
+    dump_path: Path,
+    *,
+    name: str,
+    step: int = 0,
+    rank: int = 0,
+) -> str:
+    filename: str = f"step={step}___rank={rank}___dump_index=0___name={name}.pt"
+    tensor: torch.Tensor = torch.randn(4)
+    torch.save({"value": tensor, "meta": {}}, dump_path / filename)
+    return filename
+
+
+class TestLoadAllValues:
+    def test_all_success(self, tmp_path: Path) -> None:
+        """All files load successfully — no warnings emitted."""
+        fn0: str = _save_tensor(tmp_path, name="a", rank=0)
+        fn1: str = _save_tensor(tmp_path, name="a", rank=1)
+
+        sink = WarningSink()
+        with sink.context() as warnings:
+            with patch(
+                "sglang.srt.debug_utils.comparator.bundle_comparator.warning_sink",
+                sink,
+            ):
+                result = _load_all_values(filenames=[fn0, fn1], base_path=tmp_path)
+
+        assert len(result) == 2
+        assert len(warnings) == 0
+
+    def test_one_corrupted_emits_warning(self, tmp_path: Path) -> None:
+        """One corrupted file is filtered out and emits a load_failed warning."""
+        fn_good: str = _save_tensor(tmp_path, name="a", rank=0)
+
+        fn_bad: str = "step=0___rank=1___dump_index=0___name=a.pt"
+        (tmp_path / fn_bad).write_text("not a valid pt file")
+
+        sink = WarningSink()
+        with sink.context() as warnings:
+            with patch(
+                "sglang.srt.debug_utils.comparator.bundle_comparator.warning_sink",
+                sink,
+            ):
+                result = _load_all_values(
+                    filenames=[fn_good, fn_bad], base_path=tmp_path
+                )
+
+        assert len(result) == 1
+        assert len(warnings) == 1
+        assert isinstance(warnings[0], GeneralWarning)
+        assert warnings[0].category == "load_failed"
+        assert fn_bad in warnings[0].message
+
+    def test_all_corrupted_emits_warnings_returns_empty(self, tmp_path: Path) -> None:
+        """All files corrupted — returns empty list and emits one warning per file."""
+        fn0: str = "step=0___rank=0___dump_index=0___name=a.pt"
+        fn1: str = "step=0___rank=1___dump_index=0___name=a.pt"
+        (tmp_path / fn0).write_text("corrupt")
+        (tmp_path / fn1).write_text("corrupt")
+
+        sink = WarningSink()
+        with sink.context() as warnings:
+            with patch(
+                "sglang.srt.debug_utils.comparator.bundle_comparator.warning_sink",
+                sink,
+            ):
+                result = _load_all_values(filenames=[fn0, fn1], base_path=tmp_path)
+
+        assert len(result) == 0
+        assert len(warnings) == 2
+        assert all(w.category == "load_failed" for w in warnings)
+
+
+if __name__ == "__main__":
+    sys.exit(pytest.main([__file__]))
diff --git a/test/registered/debug_utils/comparator/test_entrypoint.py b/test/registered/debug_utils/comparator/test_entrypoint.py
index 7d705816e..add958808 100644
--- a/test/registered/debug_utils/comparator/test_entrypoint.py
+++ b/test/registered/debug_utils/comparator/test_entrypoint.py
@@ -1,3 +1,4 @@
+import subprocess
 import sys
 import textwrap
 from argparse import Namespace
@@ -7,7 +8,7 @@ import pytest
 import torch
 
 import sglang.srt.debug_utils.dumper as _dumper_module
-from sglang.srt.debug_utils.comparator.entrypoint import run
+from sglang.srt.debug_utils.comparator.entrypoint import _compute_exit_code, run
 from sglang.srt.debug_utils.comparator.output_types import (
     AnyRecord,
     ComparisonRecord,
@@ -39,7 +40,7 @@ class TestEntrypointGroupingRaw:
         baseline_path, target_path = _create_dumps(tmp_path, ["tensor_a", "tensor_b"])
         args = _make_args(baseline_path, target_path, grouping="raw")
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         assert isinstance(records[0], ConfigRecord)
 
         assert len(_get_comparisons(records)) == 2
@@ -54,7 +55,7 @@ class TestEntrypointGroupingRaw:
         baseline_path, target_path = _create_dumps(tmp_path, ["tensor_a", "tensor_b"])
         args = _make_args(baseline_path, target_path, filter="tensor_a", grouping="raw")
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         assert len(_get_comparisons(records)) == 1
 
     def test_no_baseline_skip(self, tmp_path, capsys):
@@ -66,7 +67,7 @@ class TestEntrypointGroupingRaw:
         )
         args = _make_args(baseline_path, target_path, grouping="raw")
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         skips = [r for r in records if isinstance(r, SkipRecord)]
         assert len(skips) == 1
         assert skips[0].reason == "baseline_load_failed"
@@ -82,7 +83,7 @@ class TestEntrypointGroupingRaw:
             baseline_path, target_path, start_step=1, end_step=1, grouping="raw"
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         summary = records[-1]
         assert isinstance(summary, SummaryRecord)
         assert summary.total == 1
@@ -92,7 +93,7 @@ class TestEntrypointGroupingRaw:
         baseline_path, target_path = _create_dumps(tmp_path, ["t"], num_steps=2)
         args = _make_args(baseline_path, target_path, grouping="raw")
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         assert all(isinstance(r, _OutputRecord) for r in records)
 
     def test_comparison_failed(self, tmp_path, capsys):
@@ -111,7 +112,7 @@ class TestEntrypointGroupingRaw:
             baseline_path, target_path, grouping="raw", diff_threshold=1e-3
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comparisons = _get_comparisons(records)
         assert len(comparisons) == 1
         assert comparisons[0].diff is not None
@@ -133,7 +134,7 @@ class TestEntrypointGroupingRaw:
         )
         args = _make_args(baseline_path, target_path, grouping="raw")
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comparisons = _get_comparisons(records)
         assert len(comparisons) == 1
         assert comparisons[0].shape_mismatch is True
@@ -159,7 +160,7 @@ class TestEntrypointGroupingRaw:
         )
         args = _make_args(baseline_path, target_path, grouping="raw")
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comparisons = _get_comparisons(records)
         assert len(comparisons) == 1
 
@@ -189,7 +190,7 @@ class TestEntrypointGroupingRaw:
             baseline_path, target_path, grouping="raw", diff_threshold=0.01
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comparisons = _get_comparisons(records)
         assert len(comparisons) == 1
         assert comparisons[0].diff_downcast is not None
@@ -227,7 +228,7 @@ class TestEntrypointGroupingRaw:
             diff_threshold=1e-3,
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         summary = records[-1]
         assert isinstance(summary, SummaryRecord)
         assert summary.passed == 1
@@ -242,7 +243,7 @@ class TestEntrypointGroupingRaw:
             baseline_path, target_path, filter="nonexistent_pattern", grouping="raw"
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         summary = records[-1]
         assert isinstance(summary, SummaryRecord)
         assert summary.total == 0
@@ -271,7 +272,7 @@ class TestEntrypointGroupingRaw:
             diff_threshold=0.01,
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comparisons = _get_comparisons(records)
         assert len(comparisons) == 2
 
@@ -348,7 +349,7 @@ class TestEntrypointGroupingRaw:
             grouping="raw",
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comparisons = _get_comparisons(records)
         assert len(comparisons) == 2
         assert all(c.diff is not None and c.diff.passed for c in comparisons)
@@ -367,7 +368,7 @@ class TestEntrypointGroupingLogical:
         baseline_path, target_path = _create_dumps(tmp_path, ["tensor_a", "tensor_b"])
         args = _make_args(baseline_path, target_path)
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         assert len(_get_comparisons(records)) == 2
         summary = records[-1]
         assert isinstance(summary, SummaryRecord)
@@ -402,7 +403,7 @@ class TestEntrypointGroupingLogical:
 
         args = _make_args(baseline_path, target_path, diff_threshold=0.01)
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comp = _assert_single_comparison_passed(records)
         assert comp.name == "hidden"
 
@@ -439,7 +440,7 @@ class TestEntrypointGroupingLogical:
 
         args = _make_args(baseline_path, target_path, diff_threshold=0.01)
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         _assert_single_comparison_passed(records)
 
     def test_one_side_dims_single_baseline(self, tmp_path, capsys):
@@ -466,7 +467,7 @@ class TestEntrypointGroupingLogical:
 
         args = _make_args(baseline_path, target_path, diff_threshold=0.01)
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         _assert_single_comparison_passed(records)
 
     @pytest.mark.parametrize(
@@ -496,7 +497,7 @@ class TestEntrypointGroupingLogical:
             target_dir / _FIXED_EXP_NAME,
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         skips = [r for r in records if isinstance(r, SkipRecord)]
         assert len(skips) == 1
         assert skips[0].reason == expected_reason
@@ -535,7 +536,7 @@ class TestEntrypointGroupingLogical:
 
         args = _make_args(baseline_path, target_path, diff_threshold=0.01)
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         summary = records[-1]
         assert isinstance(summary, SummaryRecord)
         assert summary.total == 2
@@ -572,7 +573,7 @@ class TestEntrypointGroupingLogical:
 
         args = _make_args(baseline_path, target_path, diff_threshold=0.01)
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comparisons = _get_comparisons(records)
         assert len(comparisons) == 1
         # concat along dim 0 (fallback, no token dim) → 2 steps × [4, 8] = [8, 8]
@@ -613,7 +614,7 @@ class TestEntrypointGroupingLogical:
             diff_threshold=0.01,
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comp = _assert_single_comparison_passed(records)
         assert comp.name == "attn_out"
 
@@ -651,7 +652,7 @@ class TestEntrypointGroupingLogical:
             diff_threshold=0.01,
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comparisons = _get_comparisons(records)
         assert len(comparisons) == 1
         assert comparisons[0].name == "t_a"
@@ -696,7 +697,7 @@ class TestEntrypointGroupingLogical:
             diff_threshold=0.01,
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comparisons = _get_comparisons(records)
         assert len(comparisons) == 2
         assert all(c.diff is not None and c.diff.passed for c in comparisons)
@@ -737,7 +738,7 @@ class TestEntrypointGroupingLogical:
             diff_threshold=0.01,
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comp = _assert_single_comparison_passed(records)
         assert comp.name == "hidden"
 
@@ -776,7 +777,7 @@ class TestEntrypointGroupingLogical:
             diff_threshold=0.01,
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         _assert_single_comparison_passed(records)
 
     def test_ep_cp_tp_three_axis_unshard(self, tmp_path, capsys):
@@ -811,7 +812,7 @@ class TestEntrypointGroupingLogical:
             diff_threshold=0.01,
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comp = _assert_single_comparison_passed(records)
         assert comp.name == "hidden"
 
@@ -845,7 +846,7 @@ class TestEntrypointGroupingLogical:
             diff_threshold=0.01,
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comp = _assert_single_comparison_passed(records)
         assert comp.name == "attn_out"
 
@@ -879,7 +880,7 @@ class TestEntrypointGroupingLogical:
             diff_threshold=0.01,
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comp = _assert_single_comparison_passed(records)
         assert comp.name == "hidden"
 
@@ -906,7 +907,7 @@ class TestEntrypointGroupingLogical:
             diff_threshold=0.01,
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comp = _assert_single_comparison_passed(records)
         assert comp.name == "hidden"
 
@@ -934,7 +935,7 @@ class TestEntrypointGroupingLogical:
             diff_threshold=0.01,
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comparisons = _get_comparisons(records)
         assert len(comparisons) == 1
 
@@ -970,7 +971,7 @@ class TestEntrypointGroupingLogical:
 
         args = _make_args(baseline_path, target_path, diff_threshold=0.01)
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comp = _assert_single_comparison_passed(records)
         assert comp.name == "attn_out"
 
@@ -1001,7 +1002,7 @@ class TestEntrypointGroupingLogical:
 
         args = _make_args(baseline_path, target_path, diff_threshold=0.01)
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comp = _assert_single_comparison_passed(records)
         assert comp.name == "attn_out"
 
@@ -1041,7 +1042,7 @@ class TestEntrypointGroupingLogical:
             diff_threshold=0.01,
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comp = _assert_single_comparison_passed(records)
         assert comp.name == "hidden"
 
@@ -1105,7 +1106,8 @@ class TestEntrypointConcatMode:
         args: Namespace = _make_args(
             baseline_path, target_path, diff_threshold=diff_threshold
         )
-        return _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
+        return records
 
     def test_concat_multi_step_different_data(self, tmp_path, capsys):
         """Multi-step concat with different data per step + truncation."""
@@ -1168,7 +1170,7 @@ class TestEntrypointConcatMode:
             diff_threshold=0.01,
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comparisons = _get_comparisons(records)
         assert len(comparisons) == 1
         # 2 steps × [4, 8] concat along dim 0 (fallback) → [8, 8]
@@ -1330,7 +1332,7 @@ class TestEntrypointConcatMode:
             diff_threshold=0.01,
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comparisons = _get_comparisons(records)
         # all 3 tensors should be compared (not filtered out)
         names = {c.name for c in comparisons}
@@ -1425,7 +1427,7 @@ class TestEntrypointConcatMode:
             diff_threshold=0.01,
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comparisons = _get_comparisons(records)
         assert len(comparisons) == 1
         # CP unshard: [4,4,6] × 2 ranks → [4,8,6] per step
@@ -1477,7 +1479,7 @@ class TestEntrypointConcatMode:
             token_aligner="concat_steps",
             diff_threshold=1e-3,
         )
-        records: list[AnyRecord] = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
 
         comparisons: list[ComparisonRecord] = _get_comparisons(records)
         hidden_comparisons: list[ComparisonRecord] = [
@@ -1519,7 +1521,7 @@ class TestEntrypointAxisAligner:
             diff_threshold=1e-3,
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comp = _assert_single_comparison_passed(records)
         assert comp.name == "hidden"
         assert comp.baseline.shape == [4, 16, 8]
@@ -1556,7 +1558,7 @@ class TestEntrypointAxisAligner:
             diff_threshold=1e-3,
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comp = _assert_single_comparison_passed(records)
         assert comp.name == "hidden"
 
@@ -1589,7 +1591,7 @@ class TestEntrypointAxisAligner:
             diff_threshold=1e-3,
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comp = _assert_single_comparison_passed(records)
         assert comp.name == "hidden"
         assert comp.baseline.shape == [4, 8]
@@ -1702,7 +1704,7 @@ class TestEntrypointReplicatedAxis:
             diff_threshold=0.01,
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comp = _assert_single_comparison_passed(records)
         assert comp.warnings == []
         assert all(c.passed for c in comp.replicated_checks)
@@ -1741,7 +1743,7 @@ class TestEntrypointReplicatedAxis:
             diff_threshold=0.01,
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comparisons = _get_comparisons(records)
         assert len(comparisons) == 1
         assert comparisons[0].category == "failed"
@@ -1787,7 +1789,7 @@ class TestEntrypointReplicatedAxis:
             diff_threshold=0.5,
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         comparisons = _get_comparisons(records)
         assert len(comparisons) == 1
 
@@ -1850,7 +1852,7 @@ class TestEntrypointAlignment:
         args = _make_args(
             exp_paths[0], exp_paths[1], grouping="logical", token_aligner="smart"
         )
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
 
         comparisons = _get_comparisons(records)
         # AUX_NAMES are filtered out after plan computation → only hidden_states remains
@@ -1964,7 +1966,7 @@ class TestEntrypointAlignment:
             token_aligner="smart",
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
 
         warning_records = [r for r in records if isinstance(r, WarningRecord)]
         layout_warnings = [
@@ -2031,7 +2033,7 @@ class TestEntrypointNonTensorValues:
             tmp_path, name="sm_scale", baseline_value=0.125, target_value=0.125
         )
         args = _make_args(baseline_path, target_path, grouping="raw")
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
 
         non_tensors = _get_non_tensors(records)
         assert len(non_tensors) == 1
@@ -2050,7 +2052,7 @@ class TestEntrypointNonTensorValues:
             tmp_path, name="sm_scale", baseline_value=0.125, target_value=0.25
         )
         args = _make_args(baseline_path, target_path, grouping="raw")
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
 
         non_tensors = _get_non_tensors(records)
         assert len(non_tensors) == 1
@@ -2070,7 +2072,7 @@ class TestEntrypointNonTensorValues:
             target_value="flash_attn",
         )
         args = _make_args(baseline_path, target_path, grouping="raw")
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
 
         non_tensors = _get_non_tensors(records)
         assert len(non_tensors) == 1
@@ -2100,7 +2102,7 @@ class TestEntrypointNonTensorValues:
             target_dir / _FIXED_EXP_NAME,
             grouping="raw",
         )
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
 
         comparisons = _get_comparisons(records)
         non_tensors = _get_non_tensors(records)
@@ -2121,7 +2123,7 @@ class TestEntrypointNonTensorValues:
             tmp_path, name="debug_info", baseline_value=value, target_value=value
         )
         args = _make_args(baseline_path, target_path, grouping="raw")
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
 
         non_tensors = _get_non_tensors(records)
         assert len(non_tensors) == 1
@@ -2135,7 +2137,7 @@ class TestEntrypointNonTensorValues:
             tmp_path, name="optional_param", baseline_value=None, target_value=None
         )
         args = _make_args(baseline_path, target_path, grouping="raw")
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
 
         non_tensors = _get_non_tensors(records)
         assert len(non_tensors) == 1
@@ -2151,7 +2153,7 @@ class TestEntrypointNonTensorValues:
             tmp_path, name="sm_scale", baseline_value=0.125, target_value=0.125
         )
         args = _make_args(baseline_path, target_path, grouping="raw")
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
 
         non_tensors = _get_non_tensors(records)
         assert len(non_tensors) == 1
@@ -2186,7 +2188,7 @@ class TestEntrypointVisualize:
             viz_output_dir=str(viz_dir),
         )
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
         assert len(_get_comparisons(records)) == 1
 
         png_files = list(viz_dir.glob("*.png"))
@@ -2341,15 +2343,19 @@ def _make_args(baseline_path: Path, target_path: Path, **overrides) -> Namespace
         override_baseline_dims=[],
         override_target_dims=[],
         override_config=None,
+        allow_skip_pattern=".*",
+        report_path="",
     )
     defaults.update(overrides)
     return Namespace(**defaults)
 
 
-def _run_and_parse(args: Namespace, capsys: pytest.CaptureFixture) -> list[AnyRecord]:
+def _run_and_parse(
+    args: Namespace, capsys: pytest.CaptureFixture
+) -> tuple[list[AnyRecord], int]:
     capsys.readouterr()
-    run(args)
-    return _parse_jsonl(capsys.readouterr().out)
+    exit_code: int = run(args)
+    return _parse_jsonl(capsys.readouterr().out), exit_code
 
 
 def _parse_jsonl(output: str) -> list[AnyRecord]:
@@ -2875,7 +2881,7 @@ class TestEntrypointPerTokenVisualization:
             grouping="raw",
             visualize_per_token=str(output_png),
         )
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
 
         comparisons = _get_comparisons(records)
         assert len(comparisons) == 2
@@ -2892,7 +2898,7 @@ class TestEntrypointPerTokenVisualization:
         baseline_path, target_path = _create_dumps(tmp_path, ["tensor_a"])
         args = _make_args(baseline_path, target_path, grouping="raw")
 
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
 
         comparisons = _get_comparisons(records)
         assert len(comparisons) == 1
@@ -2992,7 +2998,7 @@ class TestEntrypointThdCpZigzag:
             token_aligner="smart",
             diff_threshold=1e-3,
         )
-        records: list[AnyRecord] = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
 
         comparisons: list[ComparisonRecord] = _get_comparisons(records)
         hidden_comparisons: list[ComparisonRecord] = [
@@ -3043,7 +3049,7 @@ class TestEntrypointThdCpZigzag:
             token_aligner="smart",
             diff_threshold=1e-3,
         )
-        records: list[AnyRecord] = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
 
         # hidden_states should pass comparison (after unshard + reorder)
         comparisons: list[ComparisonRecord] = _get_comparisons(records)
@@ -3113,7 +3119,7 @@ class TestEntrypointDpFilter:
             grouping="logical",
             diff_threshold=1e-3,
         )
-        records: list[AnyRecord] = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
 
         comparison: ComparisonRecord = _assert_single_comparison_passed(records)
         assert comparison.name == "hidden"
@@ -3169,7 +3175,7 @@ class TestEntrypointDpFilter:
             grouping="logical",
             diff_threshold=1e-3,
         )
-        records: list[AnyRecord] = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
 
         comparison: ComparisonRecord = _assert_single_comparison_passed(records)
         assert comparison.name == "hidden"
@@ -3218,7 +3224,7 @@ class TestEntrypointDpFilter:
             grouping="logical",
             diff_threshold=1e-3,
         )
-        records: list[AnyRecord] = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
 
         comparison: ComparisonRecord = _assert_single_comparison_passed(records)
         assert comparison.name == "hidden"
@@ -3344,7 +3350,7 @@ class TestEntrypointMetaOverride:
             grouping="logical",
             override_dims=["hidden:t h(tp)"],
         )
-        self._assert_all_passed(_run_and_parse(args, capsys))
+        self._assert_all_passed(_run_and_parse(args, capsys)[0])
 
     @pytest.mark.parametrize(
         "baseline_dims, target_dims, override_kwarg",
@@ -3371,7 +3377,7 @@ class TestEntrypointMetaOverride:
         )
 
         args = _make_args(baseline_path, target_path, grouping="raw", **override_kwarg)
-        self._assert_all_passed(_run_and_parse(args, capsys))
+        self._assert_all_passed(_run_and_parse(args, capsys)[0])
 
     def test_override_config_yaml(self, tmp_path: Path, capsys) -> None:
         """--override-config YAML overrides dims."""
@@ -3390,7 +3396,7 @@ class TestEntrypointMetaOverride:
             grouping="raw",
             override_config=str(yaml_path),
         )
-        self._assert_all_passed(_run_and_parse(args, capsys))
+        self._assert_all_passed(_run_and_parse(args, capsys)[0])
 
     def test_no_match_uses_original_dims(self, tmp_path: Path, capsys) -> None:
         """When override regex doesn't match, original dims from dump are used."""
@@ -3406,7 +3412,7 @@ class TestEntrypointMetaOverride:
             grouping="raw",
             override_dims=["no_match_pattern:b s d"],
         )
-        self._assert_all_passed(_run_and_parse(args, capsys))
+        self._assert_all_passed(_run_and_parse(args, capsys)[0])
 
     def test_selective_match_multi_tensor(self, tmp_path: Path, capsys) -> None:
         """Override matches only 'logits'; 'hidden' uses original dims."""
@@ -3437,7 +3443,7 @@ class TestEntrypointMetaOverride:
             grouping="raw",
             override_dims=["logits:t v"],
         )
-        self._assert_all_passed(_run_and_parse(args, capsys), expected_count=2)
+        self._assert_all_passed(_run_and_parse(args, capsys)[0], expected_count=2)
 
     def test_multiple_cli_override_dims(self, tmp_path: Path, capsys) -> None:
         """Multiple --override-dims for different tensors."""
@@ -3470,7 +3476,7 @@ class TestEntrypointMetaOverride:
             grouping="raw",
             override_dims=["hidden:t h", "logits:t v"],
         )
-        self._assert_all_passed(_run_and_parse(args, capsys), expected_count=2)
+        self._assert_all_passed(_run_and_parse(args, capsys)[0], expected_count=2)
 
     def test_per_side_dims_different_parallelism(self, tmp_path: Path, capsys) -> None:
         """baseline TP-sharded, target EP-sharded — per-side override fixes both."""
@@ -3512,7 +3518,7 @@ class TestEntrypointMetaOverride:
             override_baseline_dims=["hidden:t h(tp)"],
             override_target_dims=["hidden:t h(ep)"],
         )
-        self._assert_all_passed(_run_and_parse(args, capsys))
+        self._assert_all_passed(_run_and_parse(args, capsys)[0])
 
     def test_yaml_first_match_wins_e2e(self, tmp_path: Path, capsys) -> None:
         """YAML with two matching rules: first rule wins in real pipeline."""
@@ -3533,7 +3539,7 @@ class TestEntrypointMetaOverride:
             grouping="raw",
             override_config=str(yaml_path),
         )
-        self._assert_all_passed(_run_and_parse(args, capsys))
+        self._assert_all_passed(_run_and_parse(args, capsys)[0])
 
     def test_cli_overrides_yaml_e2e(self, tmp_path: Path, capsys) -> None:
         """CLI --override-dims wins over YAML rule for the same tensor."""
@@ -3553,7 +3559,7 @@ class TestEntrypointMetaOverride:
             override_dims=["hidden:t h"],
             override_config=str(yaml_path),
         )
-        self._assert_all_passed(_run_and_parse(args, capsys))
+        self._assert_all_passed(_run_and_parse(args, capsys)[0])
 
     def test_override_injects_dims_when_absent(self, tmp_path: Path, capsys) -> None:
         """Override injects dims into meta even when dump had no dims annotation."""
@@ -3569,7 +3575,7 @@ class TestEntrypointMetaOverride:
             grouping="raw",
             override_dims=["hidden:t h"],
         )
-        self._assert_all_passed(_run_and_parse(args, capsys))
+        self._assert_all_passed(_run_and_parse(args, capsys)[0])
 
     def test_non_tensor_unaffected_by_override(self, tmp_path: Path, capsys) -> None:
         """Non-tensor values pass through without error even with active override."""
@@ -3596,7 +3602,7 @@ class TestEntrypointMetaOverride:
             grouping="raw",
             override_dims=["hidden:x y"],
         )
-        records = _run_and_parse(args, capsys)
+        records, _ = _run_and_parse(args, capsys)
 
         non_tensors: list[NonTensorRecord] = [
             r for r in records if isinstance(r, NonTensorRecord)
@@ -3613,5 +3619,297 @@ class TestEntrypointMetaOverride:
         assert summary.failed == 0
 
 
+class TestExitCode:
+    """Tests for exit code behavior based on comparison results."""
+
+    def test_all_passed(self):
+        """All passed → exit 0."""
+        summary = SummaryRecord(total=3, passed=3, failed=0, skipped=0)
+        assert (
+            _compute_exit_code(summary, allow_skip_pattern=".*", skipped_names=[]) == 0
+        )
+
+    def test_has_failed_and_passed(self):
+        """Has failed and passed → exit 1."""
+        summary = SummaryRecord(total=4, passed=2, failed=2, skipped=0)
+        assert (
+            _compute_exit_code(summary, allow_skip_pattern=".*", skipped_names=[]) == 1
+        )
+
+    def test_all_failed(self):
+        """All failed (0 passed) → exit 1."""
+        summary = SummaryRecord(total=3, passed=0, failed=3, skipped=0)
+        assert (
+            _compute_exit_code(summary, allow_skip_pattern=".*", skipped_names=[]) == 1
+        )
+
+    def test_all_skipped_allow_all(self):
+        """All skipped + allow_skip_pattern='.*' → exit 0."""
+        summary = SummaryRecord(total=2, passed=0, failed=0, skipped=2)
+        assert (
+            _compute_exit_code(
+                summary, allow_skip_pattern=".*", skipped_names=["a", "b"]
+            )
+            == 0
+        )
+
+    def test_all_skipped_forbid_all(self):
+        """All skipped + allow_skip_pattern='^$' → exit 1."""
+        summary = SummaryRecord(total=2, passed=0, failed=0, skipped=2)
+        assert (
+            _compute_exit_code(
+                summary, allow_skip_pattern="^$", skipped_names=["a", "b"]
+            )
+            == 1
+        )
+
+    def test_passed_and_skipped_allow_all(self):
+        """Passed + skipped, allow all → exit 0."""
+        summary = SummaryRecord(total=3, passed=2, failed=0, skipped=1)
+        assert (
+            _compute_exit_code(summary, allow_skip_pattern=".*", skipped_names=["a"])
+            == 0
+        )
+
+    def test_passed_and_skipped_forbid_all(self):
+        """Passed + skipped + forbid all → exit 1."""
+        summary = SummaryRecord(total=3, passed=2, failed=0, skipped=1)
+        assert (
+            _compute_exit_code(summary, allow_skip_pattern="^$", skipped_names=["a"])
+            == 1
+        )
+
+    def test_skip_pattern_matches_specific_name(self):
+        """Pattern matching specific name allows that skip, forbids others."""
+        summary = SummaryRecord(total=4, passed=2, failed=0, skipped=2)
+        assert (
+            _compute_exit_code(
+                summary,
+                allow_skip_pattern="positions|seq_lens",
+                skipped_names=["positions", "seq_lens"],
+            )
+            == 0
+        )
+
+    def test_skip_pattern_partial_match_forbidden(self):
+        """Pattern matches some skips but not all → exit 1."""
+        summary = SummaryRecord(total=4, passed=1, failed=0, skipped=3)
+        assert (
+            _compute_exit_code(
+                summary,
+                allow_skip_pattern="positions|seq_lens",
+                skipped_names=["positions", "seq_lens", "hidden_states"],
+            )
+            == 1
+        )
+
+    def test_e2e_all_passed_exit_zero(self, tmp_path, capsys):
+        """Integration: all comparisons pass → run() returns 0."""
+        baseline_path, target_path = _create_dumps(tmp_path, ["tensor_a", "tensor_b"])
+        args = _make_args(baseline_path, target_path, grouping="raw")
+
+        records, exit_code = _run_and_parse(args, capsys)
+        summary = records[-1]
+        assert isinstance(summary, SummaryRecord)
+        assert summary.passed == 2
+        assert summary.failed == 0
+        assert exit_code == 0
+
+    def test_e2e_has_failed_exit_nonzero(self, tmp_path, capsys):
+        """Integration: a failed comparison → run() returns 1."""
+        torch.manual_seed(42)
+        baseline_path = _create_rank_dump(
+            tmp_path / "baseline", rank=0, name="tensor_a", tensor=torch.randn(10, 10)
+        )
+        target_path = _create_rank_dump(
+            tmp_path / "target",
+            rank=0,
+            name="tensor_a",
+            tensor=torch.randn(10, 10) * 100,
+        )
+        args = _make_args(
+            baseline_path, target_path, grouping="raw", diff_threshold=1e-3
+        )
+
+        records, exit_code = _run_and_parse(args, capsys)
+        summary = records[-1]
+        assert isinstance(summary, SummaryRecord)
+        assert summary.failed == 1
+        assert exit_code == 1
+
+
+class TestExitCodeSubprocess:
+    """E2E subprocess tests: invoke comparator as a child process and verify exit code."""
+
+    @staticmethod
+    def _run_comparator(
+        baseline_path: Path,
+        target_path: Path,
+        *,
+        grouping: str = "raw",
+        allow_skip_pattern: str = ".*",
+    ) -> subprocess.CompletedProcess[str]:
+        cmd: list[str] = [
+            sys.executable,
+            "-m",
+            "sglang.srt.debug_utils.comparator",
+            "--baseline-path",
+            str(baseline_path),
+            "--target-path",
+            str(target_path),
+            "--grouping",
+            grouping,
+            "--output-format",
+            "json",
+            "--allow-skip-pattern",
+            allow_skip_pattern,
+        ]
+        return subprocess.run(cmd, capture_output=True, text=True)
+
+    def test_all_passed_exit_zero(self, tmp_path):
+        """Subprocess: all comparisons pass → exit 0."""
+        baseline_path, target_path = _create_dumps(tmp_path, ["tensor_a"])
+        result = self._run_comparator(baseline_path, target_path)
+        assert result.returncode == 0
+
+    def test_failed_exit_nonzero(self, tmp_path):
+        """Subprocess: failed comparison → exit 1."""
+        torch.manual_seed(42)
+        baseline_path = _create_rank_dump(
+            tmp_path / "baseline", rank=0, name="t", tensor=torch.randn(10, 10)
+        )
+        target_path = _create_rank_dump(
+            tmp_path / "target", rank=0, name="t", tensor=torch.randn(10, 10) * 100
+        )
+        result = self._run_comparator(baseline_path, target_path)
+        assert result.returncode == 1
+
+    def test_skipped_allow_all_exit_zero(self, tmp_path):
+        """Subprocess: skipped comparison with allow_skip_pattern='.*' → exit 0."""
+        baseline_path, target_path = _create_dumps(
+            tmp_path,
+            tensor_names=["tensor_a", "tensor_extra"],
+            baseline_names=["tensor_a"],
+        )
+        result = self._run_comparator(
+            baseline_path, target_path, allow_skip_pattern=".*"
+        )
+        assert result.returncode == 0
+
+    def test_skipped_forbid_all_exit_nonzero(self, tmp_path):
+        """Subprocess: skipped comparison with allow_skip_pattern='^$' → exit 1."""
+        baseline_path, target_path = _create_dumps(
+            tmp_path,
+            tensor_names=["tensor_a", "tensor_extra"],
+            baseline_names=["tensor_a"],
+        )
+        result = self._run_comparator(
+            baseline_path, target_path, allow_skip_pattern="^$"
+        )
+        assert result.returncode == 1
+
+
+class TestReportOutput:
+    """Test JSONL report file output via ReportSink."""
+
+    def test_default_report_path(self, tmp_path, capsys):
+        """Default writes to <target>/comparator_report.jsonl with ConfigRecord + SummaryRecord."""
+        baseline_path, target_path = _create_dumps(tmp_path, ["tensor_a"])
+        args = _make_args(baseline_path, target_path, grouping="raw", report_path=None)
+
+        exit_code: int = run(args)
+
+        report_file: Path = target_path / "comparator_report.jsonl"
+        assert report_file.exists()
+
+        report_records: list[AnyRecord] = _parse_jsonl(report_file.read_text())
+        assert isinstance(report_records[0], ConfigRecord)
+        assert isinstance(report_records[-1], SummaryRecord)
+        assert exit_code == 0
+
+    def test_custom_report_path(self, tmp_path, capsys):
+        """--report-path writes to the specified location."""
+        baseline_path, target_path = _create_dumps(tmp_path, ["tensor_a"])
+        custom_path: Path = tmp_path / "custom" / "report.jsonl"
+        args = _make_args(
+            baseline_path,
+            target_path,
+            grouping="raw",
+            report_path=str(custom_path),
+        )
+
+        run(args)
+
+        assert custom_path.exists()
+        report_records: list[AnyRecord] = _parse_jsonl(custom_path.read_text())
+        assert isinstance(report_records[0], ConfigRecord)
+        assert isinstance(report_records[-1], SummaryRecord)
+
+    def test_disabled_report(self, tmp_path, capsys):
+        """--report-path '' disables file generation."""
+        baseline_path, target_path = _create_dumps(tmp_path, ["tensor_a"])
+        args = _make_args(baseline_path, target_path, grouping="raw", report_path="")
+
+        run(args)
+
+        report_file: Path = target_path / "comparator_report.jsonl"
+        assert not report_file.exists()
+
+    def test_report_matches_stdout_json(self, tmp_path, capsys):
+        """In json mode, report content matches stdout output."""
+        baseline_path, target_path = _create_dumps(tmp_path, ["tensor_a"])
+        report_file: Path = tmp_path / "report.jsonl"
+        args = _make_args(
+            baseline_path,
+            target_path,
+            grouping="raw",
+            output_format="json",
+            report_path=str(report_file),
+        )
+
+        capsys.readouterr()
+        run(args)
+
+        stdout_lines: list[str] = capsys.readouterr().out.strip().splitlines()
+        report_lines: list[str] = report_file.read_text().strip().splitlines()
+        assert stdout_lines == report_lines
+
+    def test_text_mode_also_writes_report(self, tmp_path, capsys):
+        """Text stdout mode still writes JSONL report."""
+        baseline_path, target_path = _create_dumps(tmp_path, ["tensor_a"])
+        report_file: Path = tmp_path / "report.jsonl"
+        args = _make_args(
+            baseline_path,
+            target_path,
+            grouping="raw",
+            output_format="text",
+            report_path=str(report_file),
+        )
+
+        run(args)
+
+        assert report_file.exists()
+        report_records: list[AnyRecord] = _parse_jsonl(report_file.read_text())
+        assert isinstance(report_records[0], ConfigRecord)
+        assert isinstance(report_records[-1], SummaryRecord)
+
+    def test_streaming_flush(self, tmp_path, capsys):
+        """Report file is flushed after each record (readable before close)."""
+        from sglang.srt.debug_utils.comparator.output_types import report_sink
+
+        report_file: Path = tmp_path / "stream_report.jsonl"
+        report_sink.configure(
+            output_format="json",
+            report_path=report_file,
+        )
+
+        report_sink.add(ConfigRecord(config={"test": True}))
+
+        content: str = report_file.read_text()
+        assert len(content.strip().splitlines()) == 1
+        parsed: AnyRecord = parse_record_json(content.strip())
+        assert isinstance(parsed, ConfigRecord)
+
+
 if __name__ == "__main__":
     sys.exit(pytest.main([__file__]))
diff --git a/test/registered/debug_utils/comparator/test_warning_sink.py b/test/registered/debug_utils/comparator/test_warning_sink.py
index 7a41dcf4c..efc66cc32 100644
--- a/test/registered/debug_utils/comparator/test_warning_sink.py
+++ b/test/registered/debug_utils/comparator/test_warning_sink.py
@@ -3,7 +3,10 @@ import sys
 
 import pytest
 
-from sglang.srt.debug_utils.comparator.output_types import GeneralWarning
+from sglang.srt.debug_utils.comparator.output_types import (
+    GeneralWarning,
+    report_sink,
+)
 from sglang.srt.debug_utils.comparator.warning_sink import WarningSink
 from sglang.test.ci.ci_register import register_cpu_ci
 
@@ -53,7 +56,7 @@ class TestWarningSink:
 
     def test_add_outside_context_prints(self, capsys) -> None:
         sink = WarningSink()
-        sink.set_output_format("text")
+        report_sink.configure(output_format="text")
 
         sink.add(_make_warning())
 
@@ -62,7 +65,7 @@ class TestWarningSink:
 
     def test_context_captures_instead_of_printing(self, capsys) -> None:
         sink = WarningSink()
-        sink.set_output_format("text")
+        report_sink.configure(output_format="text")
 
         with sink.context() as collected:
             sink.add(_make_warning())
@@ -73,7 +76,7 @@ class TestWarningSink:
 
     def test_json_output_outside_context(self, capsys) -> None:
         sink = WarningSink()
-        sink.set_output_format("json")
+        report_sink.configure(output_format="json")
 
         sink.add(_make_warning())
 
@@ -84,7 +87,7 @@ class TestWarningSink:
 
     def test_exception_in_context_cleans_stack(self, capsys) -> None:
         sink = WarningSink()
-        sink.set_output_format("text")
+        report_sink.configure(output_format="text")
 
         with pytest.raises(RuntimeError):
             with sink.context() as collected:
diff --git a/test/registered/debug_utils/test_engine_dumper_comparator_e2e.py b/test/registered/debug_utils/test_engine_dumper_comparator_e2e.py
index 59477f910..05d46761d 100644
--- a/test/registered/debug_utils/test_engine_dumper_comparator_e2e.py
+++ b/test/registered/debug_utils/test_engine_dumper_comparator_e2e.py
@@ -34,7 +34,7 @@ register_cuda_ci(est_time=120, suite="nightly-2-gpu", nightly=True)
 
 MODEL = "Qwen/Qwen3-0.6B"
 EXP_NAME = "e2e_source_patcher"
-DUMPER_FILTER = r"layer_id=[012]"
+DUMPER_FILTER = "layer_id in [0, 1, 2]"
 
 PATCH_CONFIG_YAML: str = """\
 patches: