Add the OSS-Fuzz reproducers for 474400817 and 474480402 as unit tests

Following the convention added to tests/fuzzing.md, the reproducers of
the two BJData fuzzer asserts tracked since January are now unit tests:

- 474400817 (assert(false)): an empty object _ArraySize_ was written as
  the ND-array header length, which from_bjdata() could not read back.
  Fixed by #5455.

- 474480402 (to_bjdata(j2, false, false) == vec2): a one-byte Draft 3
  binary array is written in Draft 2 mode as a uint8 array and then
  re-serialized with the int8 marker. This is the documented exception to
  byte stability, not a library bug; OSS-Fuzz closed it after #5494
  relaxed the harness to value stability. The test pins the exact bytes
  so the exception stays deliberate.

The 563659413 reproducer is already a unit test (#5542). A comment also
ties the existing UBJSON excessive-count test to the timeout OSS-Fuzz
reported for that shape (testcase 6347769435193344).

OSS-Fuzz: 474400817
OSS-Fuzz: 474480402

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

.github/PULL_REQUEST_TEMPLATE.md

@@ -2,6 +2,7 @@
 
 - [ ] The changes are described in detail, both the what and why.
 - [ ] If applicable, an [existing issue](https://github.com/nlohmann/json/issues) is referenced.
+- [ ] If applicable, a fixed [OSS-Fuzz](https://issues.oss-fuzz.com) issue is referenced as `OSS-Fuzz: <id>` (see [fuzz testing](https://github.com/nlohmann/json/blob/develop/tests/fuzzing.md#handling-oss-fuzz-reports)).
 - [ ] The [Code coverage](https://coveralls.io/github/nlohmann/json) remained at 100%. A test case for every new line of code.
 - [ ] If applicable, the [documentation](https://json.nlohmann.me) is updated.
 - [ ] The source code is amalgamated by running `make amalgamate`.

docs/mkdocs/docs/community/quality_assurance.md

@@ -164,6 +164,9 @@ Note: Some modern features (like C++20 ranges or filesystem support) may be disa
 - [x] The parser is tested against extensive correctness suites for JSON compliance.
 - [x] In addition, the library is continuously fuzz-tested at [OSS-Fuzz](https://google.github.io/oss-fuzz/) where the
   library is checked against billions of inputs.
+- [x] Every crash reported by OSS-Fuzz is fixed together with a unit test that reproduces it, and the fix references
+  the OSS-Fuzz issue. The round-trip checks of the fuzzer drivers are also part of the unit tests. See the
+  [fuzz testing documentation](https://github.com/nlohmann/json/blob/develop/tests/fuzzing.md#handling-oss-fuzz-reports).
 
 ## Static analysis
 

docs/mkdocs/docs/features/binary_formats/bjdata.md

@@ -208,6 +208,16 @@ The library maps BJData types to JSON value types as follows:
 
     The mapping is **complete** in the sense that any BJData value can be converted to a JSON value.
 
+!!! info "Round trips"
+
+    A value returned by [`from_bjdata`](../../api/basic_json/from_bjdata.md) can be serialized with
+    [`to_bjdata`](../../api/basic_json/to_bjdata.md) using any combination of options and parsed back into an equal
+    value, and serializing that value again with the same options produces the same bytes. The exception is binary
+    values: they are only written as an optimized binary array (`[$B`) if Draft 3 is enabled and both `use_size` and
+    `use_type` are set. Otherwise, they are written as arrays of integers and parsed back as such (see the notes on
+    binary values above), and serializing such an array again may choose different, but equally valid, type markers.
+    The bytes can then differ, but parsing them again yields the same value.
+
 ??? example
 
     ```cpp

tests/fuzzing.md

@@ -79,3 +79,26 @@ the same `fuzzers` target as above and also relies on the `FUZZER_ENGINE` variab
 [build script](https://github.com/google/oss-fuzz/blob/master/projects/json/build.sh) for more information.
 
 In case the build at OSS-Fuzz fails, an issue will be created automatically.
+
+### Handling OSS-Fuzz reports
+
+OSS-Fuzz files the crashes it finds in its own [issue tracker](https://issues.oss-fuzz.com), not on GitHub. So that
+each report can be traced to the change that fixed it, and each fix to the report it answers, fixes follow these
+conventions:
+
+- **Reference the OSS-Fuzz issue in the pull request**, next to any GitHub issue it closes, as `OSS-Fuzz: <id>` (for
+  example, `OSS-Fuzz: 563659413`), and in the commit message. The ID alone does not disclose the crash. If the report
+  was triaged into a GitHub issue, link the OSS-Fuzz issue there too.
+- **Turn the reproducer into a unit test.** Download the testcase from the OSS-Fuzz report, reduce it if possible, and
+  add it as a regression test to the unit test of the affected format (e.g., `tests/src/unit-bjdata.cpp`), with a
+  comment naming the OSS-Fuzz issue. This way the input is checked by every CI run rather than only by OSS-Fuzz, and
+  it stays covered even if OSS-Fuzz later closes the report as not reproducible.
+- **Keep the fuzzer drivers and the unit tests in sync.** The round-trip checks of the UBJSON and BJData drivers are
+  also run on a fixed corpus in the unit tests (see `tests/src/round_trip_corpus.hpp` and the "round-trip invariants"
+  test cases), so a regression shows up in CI first. When a driver's checks change, change the unit tests with them.
+- **Record in the report whether the bug shipped.** OSS-Fuzz asks whether a crash was a short-lived regression or
+  affects a released version; answer it when the fix is merged, as it decides whether the fix needs a release note or
+  a security advisory (see the [security policy](../.github/SECURITY.md)).
+
+After the fix is merged, OSS-Fuzz re-runs the reproducer on its next build and marks the report as verified and
+closed. If it does not, the fix is incomplete.

tests/src/fuzzer-parse_bjdata.cpp

@@ -42,6 +42,9 @@ dump() serializes any non-finite double the same deterministic way (as JSON
 `null`, since JSON itself cannot represent NaN/Infinity), so comparing
 dumps is stable under exactly the same values that break operator==.
 
+The unit tests run the same checks on a fixed corpus (see the "BJData round-trip
+invariants" test case), so keep both in sync.
+
 The provided function `LLVMFuzzerTestOneInput` can be used in different fuzzer
 drivers.
 */

tests/src/fuzzer-parse_ubjson.cpp

@@ -21,6 +21,9 @@ array data, it performs the following steps:
 - j4 = from_ubjson(vec3)
 - assert(j1 == j4)
 
+The unit tests run the same checks on a fixed corpus (see the "UBJSON round-trip
+invariants" test case), so keep both in sync.
+
 The provided function `LLVMFuzzerTestOneInput` can be used in different fuzzer
 drivers.
 */

tests/src/round_trip_corpus.hpp

@@ -0,0 +1,214 @@
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++ (supporting code)
+// |  |  |__   |  |  | | | |  version 3.12.0
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2026 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include <cmath> // nan
+#include <cstddef> // size_t
+#include <cstdint> // int32_t, int64_t, uint32_t, uint64_t
+#include <limits> // numeric_limits
+#include <random> // mt19937
+#include <string> // string, to_string
+#include <utility> // move
+#include <vector> // vector
+
+#include <nlohmann/json.hpp>
+
+// Values for the round-trip property tests of the UBJSON and BJData writers.
+//
+// The fuzzer drivers (tests/src/fuzzer-parse_ubjson.cpp and
+// fuzzer-parse_bjdata.cpp) check that anything the library parses can be
+// serialized, parsed back, and serialized again without loss. Those checks
+// only run at OSS-Fuzz, so a regression used to surface days later as an
+// external report. The unit tests run the same checks on this corpus in CI.
+//
+// The corpus is deterministic: std::mt19937's output sequence is fixed by
+// the standard, and it is used directly rather than through a distribution
+// (whose results are implementation-defined).
+namespace utils
+{
+
+class round_trip_corpus
+{
+  public:
+    using json = nlohmann::json;
+
+    static std::vector<json> values()
+    {
+        round_trip_corpus corpus;
+        return corpus.build();
+    }
+
+    // whether a value contains a binary value, which a BJData or UBJSON round
+    // trip may turn into an array of integers
+    static bool contains_binary(const json& j)
+    {
+        if (j.is_binary())
+        {
+            return true;
+        }
+        if (j.is_structured())
+        {
+            for (const auto& element : j)
+            {
+                if (contains_binary(element))
+                {
+                    return true;
+                }
+            }
+        }
+        return false;
+    }
+
+  private:
+    std::vector<json> atoms;
+    // a fixed seed is the point: the corpus must be the same in every run
+    std::mt19937 generator{42}; // NOLINT(cert-msc32-c,cert-msc51-cpp,bugprone-random-generator-seed)
+
+    round_trip_corpus()
+    {
+        atoms =
+        {
+            nullptr, true, false,
+            // integers at the boundaries of every UBJSON/BJData integer type
+            0, 1, -1, 127, 128, 255, 256, -128, -129,
+            32767, 32768, 65535, 65536, -32768, -32769,
+            (std::numeric_limits<std::int32_t>::min)(), (std::numeric_limits<std::int32_t>::max)(),
+            (std::numeric_limits<std::uint32_t>::max)(),
+            (std::numeric_limits<std::int64_t>::min)(), (std::numeric_limits<std::int64_t>::max)(),
+            static_cast<std::uint64_t>((std::numeric_limits<std::int64_t>::max)()) + 1u,
+            (std::numeric_limits<std::uint64_t>::max)(),
+            // floating-point numbers, including non-finite ones
+            0.0, -0.0, 1.5, -2.25, 3.4e38, (std::numeric_limits<double>::max)(),
+            std::nan(""), std::numeric_limits<double>::infinity(), -std::numeric_limits<double>::infinity(),
+            // strings, including a non-ASCII one and one longer than 255 bytes
+            "", "a", "\xC3\xA4", std::string(300, 'x'),
+            // binary values with and without subtype
+            json::binary({}), json::binary({1, 2, 255}), json::binary({0x80, 0x7F}, 42), json::binary({1}, 0)
+        };
+    }
+
+    std::vector<json> build()
+    {
+        std::vector<json> result = atoms;
+
+        // each atom inside containers, including homogeneous ones that the
+        // writers encode as optimized (typed) containers
+        result.emplace_back(json::array());
+        result.emplace_back(json::object());
+        for (const auto& atom : atoms)
+        {
+            result.push_back(json::array({atom}));
+            result.push_back(json::array({atom, atom, atom}));
+            result.push_back(json::array({json::array({atom})}));
+            result.push_back(json::object({{"key", atom}}));
+        }
+        result.push_back(json::array({1, 1.5}));
+        result.push_back(json::array({-1, 255}));
+        result.push_back(json::array({"a", "b"}));
+
+        // deep, but well below any recursion or depth limit
+        json nested_array = 1;
+        json nested_object = 1;
+        for (int i = 0; i < 300; ++i)
+        {
+            nested_array = json::array({nested_array});
+            nested_object = json::object({{"key", nested_object}});
+        }
+        result.push_back(nested_array);
+        result.push_back(nested_object);
+
+        add_annotated_arrays(result);
+        add_random_values(result);
+        return result;
+    }
+
+    // objects in the JData annotated array format, which the BJData writer
+    // encodes as ND-arrays when the annotation describes a packed array, and
+    // as plain objects otherwise (see #5398, #5399, #5403, #5404, and #5542)
+    static void add_annotated_arrays(std::vector<json>& result)
+    {
+        const std::vector<json> types =
+        {
+            "uint8", "int8", "uint16", "int16", "uint32", "int32", "uint64", "int64",
+            "single", "double", "char", "byte", "bool", "unknown", 5, nullptr
+        };
+        const std::vector<json> sizes =
+        {
+            json::array(), {3}, {1, 3}, {3, 1}, {2, 3}, {2, 0}, {0, 2}, {2, 2, 2}, {-1, 2}, {2, 1.5},
+            "3", 3, nullptr, json::binary({})
+        };
+        const std::vector<json> data =
+        {
+            nullptr, 5, "s", json::object({{"a", 1}}), json::array(),
+            {1, 2, 3}, {1, 2, 3, 4, 5, 6}, {1, 2, 3, 4, 5, 6, 7, 8},
+            {1.5, 2.5, 3.5, 4.5, 5.5, 6.5}, {300, -300, 70000, -70000, 1, 2},
+            {"a", "b", "c", "d", "e", "f"}, {json::array({1, 2, 3}), json::array({4, 5, 6})}
+        };
+
+        for (const auto& type : types)
+        {
+            for (const auto& size : sizes)
+            {
+                for (const auto& d : data)
+                {
+                    result.push_back({{"_ArrayType_", type}, {"_ArraySize_", size}, {"_ArrayData_", d}});
+                }
+            }
+        }
+
+        // incomplete annotations and annotations with an extra key
+        result.push_back({{"_ArraySize_", {2, 3}}, {"_ArrayData_", {1, 2, 3, 4, 5, 6}}});
+        result.push_back({{"_ArrayType_", "uint8"}, {"_ArrayData_", {1, 2, 3, 4, 5, 6}}});
+        result.push_back({{"_ArrayType_", "uint8"}, {"_ArraySize_", {2, 3}}});
+        result.push_back({{"_ArrayType_", "uint8"}, {"_ArraySize_", {2, 3}}, {"_ArrayData_", {1, 2, 3, 4, 5, 6}}, {"extra", 1}});
+    }
+
+    // random containers of atoms, both homogeneous and mixed
+    void add_random_values(std::vector<json>& result)
+    {
+        for (int i = 0; i < 1000; ++i)
+        {
+            result.push_back(random_value(0));
+        }
+    }
+
+    std::size_t random_below(std::size_t bound)
+    {
+        return static_cast<std::size_t>(generator()) % bound;
+    }
+
+    json random_value(int depth)
+    {
+        const auto kind = random_below(10);
+        if (depth > 3 || kind < 5)
+        {
+            return atoms[random_below(atoms.size())];
+        }
+
+        json result = kind < 8 ? json::array() : json::object();
+        const auto count = random_below(5);
+        const bool homogeneous = random_below(2) == 0;
+        const json fixed = atoms[random_below(atoms.size())];
+        for (std::size_t i = 0; i < count; ++i)
+        {
+            json element = homogeneous ? fixed : random_value(depth + 1);
+            if (result.is_array())
+            {
+                result.push_back(std::move(element));
+            }
+            else
+            {
+                result[std::to_string(i)] = std::move(element);
+            }
+        }
+        return result;
+    }
+};
+
+} // namespace utils

tests/src/unit-bjdata.cpp

@@ -19,6 +19,7 @@ using nlohmann::json;
 #include <fstream>
 #include <set>
 #include "make_test_data_available.hpp"
+#include "round_trip_corpus.hpp"
 #include "test_utils.hpp"
 
 namespace
@@ -2867,6 +2868,21 @@ TEST_CASE("BJData")
                 const auto out_num = json::to_bjdata(j_num);
                 CHECK(out_num.at(0) == '{');
                 CHECK(json::from_bjdata(out_num) == j_num);
+
+                // OSS-Fuzz issue 474400817: an empty object _ArraySize_ was
+                // written as the ND-array header length, which from_bjdata()
+                // could not read back
+                const std::vector<uint8_t> input =
+                {
+                    '[', '{', 'U', 11, '_', 'A', 'r', 'r', 'a', 'y', 'D', 'a', 't', 'a', '_', 'Z',
+                    'U', 11, '_', 'A', 'r', 'r', 'a', 'y', 'T', 'y', 'p', 'e', '_', 'S', 'i', 5, 'i', 'n', 't', '1', '6',
+                    'U', 11, '_', 'A', 'r', 'r', 'a', 'y', 'S', 'i', 'z', 'e', '_', '{', '}', '}', ']'
+                };
+                const json j1 = json::from_bjdata(input);
+                CHECK(j1 == json::parse(R"([{"_ArrayType_":"int16","_ArraySize_":{},"_ArrayData_":null}])"));
+                json j2;
+                CHECK_NOTHROW(j2 = json::from_bjdata(json::to_bjdata(j1, false, false)));
+                CHECK(j2 == j1);
             }
 
             SECTION("ndarray with out-of-range _ArrayData_ elements stays as object")
@@ -4273,6 +4289,93 @@ TEST_CASE("BJData use_type requires use_size")
     }
 }
 
+TEST_CASE("BJData round-trip invariants")
+{
+    // This checks what the parse_bjdata_fuzzer driver checks (see
+    // tests/src/fuzzer-parse_bjdata.cpp), so that a regression shows up in CI
+    // rather than as an OSS-Fuzz report: every value from_bjdata() returns
+    // (j1) can be serialized with any combination of options, the result can
+    // be parsed back (j2), and serializing j2 again with the same options
+    // yields a value-equal result.
+    //
+    // Beyond the driver, this also checks that j2 equals j1 and that
+    // serializing j2 reproduces the exact bytes, both except for values that
+    // contain a binary value: a binary value is only written as a binary
+    // value with Draft 3's optimized binary array, and otherwise read back as
+    // an array of integers, for which the writer may choose different (but
+    // equally valid) type markers when it is serialized again (see #5494).
+    //
+    // Values are compared with dump() rather than operator==, because a NaN
+    // never compares equal to itself.
+    struct options
+    {
+        bool use_size;
+        bool use_type;
+        json::bjdata_version_t version;
+    };
+    const std::vector<options> all_options =
+    {
+        {false, false, json::bjdata_version_t::draft2},
+        {true, false, json::bjdata_version_t::draft2},
+        {true, true, json::bjdata_version_t::draft2},
+        {false, false, json::bjdata_version_t::draft3},
+        {true, false, json::bjdata_version_t::draft3},
+        {true, true, json::bjdata_version_t::draft3},
+    };
+
+    for (const auto& j0 : utils::round_trip_corpus::values())
+    {
+        // turn the corpus value into a value as from_bjdata() returns it
+        for (const auto& initial : all_options)
+        {
+            const json j1 = json::from_bjdata(json::to_bjdata(j0, initial.use_size, initial.use_type, initial.version));
+            const bool has_binary = utils::round_trip_corpus::contains_binary(j1);
+
+            for (const auto& o : all_options)
+            {
+                INFO("j1 = " << j1.dump() << ", use_size = " << o.use_size << ", use_type = " << o.use_type
+                     << ", draft3 = " << (o.version == json::bjdata_version_t::draft3));
+
+                const std::vector<std::uint8_t> vec = json::to_bjdata(j1, o.use_size, o.use_type, o.version);
+                json j2;
+                // anything the library writes must be parsable by the library
+                REQUIRE_NOTHROW(j2 = json::from_bjdata(vec));
+                const std::vector<std::uint8_t> vec2 = json::to_bjdata(j2, o.use_size, o.use_type, o.version);
+                CHECK(json::from_bjdata(vec2).dump() == j2.dump());
+
+                if (!has_binary)
+                {
+                    CHECK(j2.dump() == j1.dump());
+                    CHECK(vec2 == vec);
+                }
+            }
+        }
+    }
+}
+
+TEST_CASE("BJData round trip of a binary value is value-stable, not byte-stable")
+{
+    // OSS-Fuzz issue 474480402: a Draft 3 optimized binary array is read as a
+    // binary value, which to_bjdata() writes in the default Draft 2 mode as a
+    // plain array of uint8 numbers. That is read back as an array of numbers,
+    // for which the writer then picks the smallest type marker, int8 ('i'),
+    // so re-serializing changes the bytes, but not the value. This is the
+    // exception described in the "Round trips" note of the BJData
+    // documentation, and why the fuzzer checks value stability (see #5494).
+    const std::vector<uint8_t> input = {'[', '$', 'B', '#', 'U', 1, 0x20};
+    const json j1 = json::from_bjdata(input);
+    CHECK(j1 == json::binary({0x20}));
+
+    const std::vector<uint8_t> vec = json::to_bjdata(j1, false, false);
+    CHECK(vec == std::vector<uint8_t>({'[', 'U', 0x20, ']'}));
+    const json j2 = json::from_bjdata(vec);
+    CHECK(j2 == json::array({0x20}));
+
+    const std::vector<uint8_t> vec2 = json::to_bjdata(j2, false, false);
+    CHECK(vec2 == std::vector<uint8_t>({'[', 'i', 0x20, ']'}));
+    CHECK(json::from_bjdata(vec2) == j2);
+}
+
 TEST_CASE("BJData roundtrips" * doctest::skip())
 {
     SECTION("input from self-generated BJData files")

tests/src/unit-ubjson.cpp

@@ -15,6 +15,7 @@ using nlohmann::json;
 #include <fstream>
 #include <set>
 #include "make_test_data_available.hpp"
+#include "round_trip_corpus.hpp"
 #include "test_utils.hpp"
 
 namespace
@@ -2265,7 +2266,9 @@ TEST_CASE("UBJSON optimized arrays of a valueless type are bounded")
 
     SECTION("an excessive count is rejected")
     {
-        // 'l' is a big-endian int32: 0x7FFFFFFF elements, about 34 GB of value
+        // 'l' is a big-endian int32: 0x7FFFFFFF elements, about 34 GB of value;
+        // OSS-Fuzz reported this shape as a parse_ubjson_fuzzer timeout
+        // (testcase 6347769435193344, no issue filed)
         for (const auto marker :
                 {'Z', 'T', 'F'
                 })
@@ -2817,6 +2820,51 @@ TEST_CASE("UBJSON use_type requires use_size")
     }
 }
 
+TEST_CASE("UBJSON round-trip invariants")
+{
+    // This checks what the parse_ubjson_fuzzer driver checks (see
+    // tests/src/fuzzer-parse_ubjson.cpp), so that a regression shows up in CI
+    // rather than as an OSS-Fuzz report: every value from_ubjson() returns
+    // (j1) can be serialized with any combination of options, the result can
+    // be parsed back (j2), and serializing j2 again with the same options
+    // reproduces the exact bytes. Beyond the driver, this also checks that j2
+    // equals j1. Values are compared with dump() rather than operator==,
+    // because a NaN never compares equal to itself.
+    struct options
+    {
+        bool use_size;
+        bool use_type;
+    };
+    const std::vector<options> all_options =
+    {
+        {false, false},
+        {true, false},
+        {true, true},
+    };
+
+    for (const auto& j0 : utils::round_trip_corpus::values())
+    {
+        // turn the corpus value into a value as from_ubjson() returns it; this
+        // has no binary values, as UBJSON writes them as arrays of integers
+        for (const auto& initial : all_options)
+        {
+            const json j1 = json::from_ubjson(json::to_ubjson(j0, initial.use_size, initial.use_type));
+
+            for (const auto& o : all_options)
+            {
+                INFO("j1 = " << j1.dump() << ", use_size = " << o.use_size << ", use_type = " << o.use_type);
+
+                const std::vector<std::uint8_t> vec = json::to_ubjson(j1, o.use_size, o.use_type);
+                json j2;
+                // anything the library writes must be parsable by the library
+                REQUIRE_NOTHROW(j2 = json::from_ubjson(vec));
+                CHECK(j2.dump() == j1.dump());
+                CHECK(json::to_ubjson(j2, o.use_size, o.use_type) == vec);
+            }
+        }
+    }
+}
+
 TEST_CASE("UBJSON roundtrips" * doctest::skip())
 {
     SECTION("input from self-generated UBJSON files")