Write doubles with the shortest digits (Zmij)

dump() writes doubles with the conversion of Zmij by Victor Zverovich (https://github.com/vitaut/zmij, MIT), ported to C++11 in detail/conversions/zmij.hpp: the shortest decimal in the rounding interval, the closest one if there are several. Grisu2 does not always find the shortest digits; about 0.14% of random doubles are now written differently (0.08% with fewer digits, 0.06% with the closest last digit); short decimals such as 0.1 or 2555.56 are not affected. float keeps Grisu2. The layout of doubles is unchanged, but written differently: the digits are converted eight at a time (the BCD conversion of Xiang JunBo, as in Zmij) and stored with one byte swap per eight digits; leading and trailing zeros are counted from those bytes; and the layouts of format_buffer() are written with fixed-size moves instead of per-digit loops and moves of the buffer (to_chars() uses a local buffer if the caller's is shorter than the 41 bytes this may write). The powers of ten come from the table for number parsing, adjusted where it holds them rounded up, and from the compressed tables of Zmij beyond 10^308. json::dump() gets faster on floats: canada -53%, numbers -46%, mesh -37%, marine_ik -30%. Tests: the powers of ten recomputed with a small big-integer; for random doubles, all powers of two and of ten and their neighbors, and boundary values: the output reads back as the same value, no decimal with one digit fewer does, the layout equals that of format_buffer() for the same digits, and (C++17) the digits equal those of std::to_chars. The size ratios of canada.json in unit-binary_formats.cpp and one expectation in unit-to_chars.cpp change with the shorter output. Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Write compact dumps of json_view without a library call per token
2026-09-29 19:20:30 +00:00 · 2026-09-29 14:03:25 +02:00 · 2026-09-29 14:03:24 +02:00 · 2026-09-29 14:03:23 +02:00 · 2026-09-29 14:03:22 +02:00 · 2026-09-29 14:03:21 +02:00
38 changed files with 5126 additions and 325 deletions
@@ -51,7 +51,7 @@ labels:
    - "include/nlohmann/detail/view/.*"
    - "single_include/nlohmann/json_view\\.hpp"
    - "tests/src/unit-json_view.*"
-    - "tests/src/fuzzer-parse_json_view\\.cpp"
+    - "tests/src/fuzzer-(parse_json_view|json_view_image)\\.cpp"
    - "tests/benchmarks/json_view/.*"
    - "tools/amalgamate/config_json_view\\.json"
    - "docs/mkdocs/docs/features/json_view\\.md"
@@ -26,6 +26,7 @@ cc_library(
        "include/nlohmann/detail/conversions/from_json.hpp",
        "include/nlohmann/detail/conversions/to_chars.hpp",
        "include/nlohmann/detail/conversions/to_json.hpp",
+        "include/nlohmann/detail/conversions/zmij.hpp",
        "include/nlohmann/detail/exceptions.hpp",
        "include/nlohmann/detail/hash.hpp",
        "include/nlohmann/detail/input/binary_reader.hpp",
@@ -72,6 +73,7 @@ cc_library(
        "include/nlohmann/detail/view/edit.hpp",
        "include/nlohmann/detail/view/edit_storage.hpp",
        "include/nlohmann/detail/view/errors.hpp",
+        "include/nlohmann/detail/view/image.hpp",
        "include/nlohmann/detail/view/input.hpp",
        "include/nlohmann/detail/view/iterator.hpp",
        "include/nlohmann/detail/view/lookup.hpp",
@@ -1401,6 +1401,7 @@ THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR I

 - The class contains the UTF-8 Decoder from Bjoern Hoehrmann which is licensed under the [MIT License](https://opensource.org/licenses/MIT) (see above). Copyright &copy; 2008-2009 [Björn Hoehrmann](https://bjoern.hoehrmann.de/) <bjoern@hoehrmann.de>
 - The class contains a slightly modified version of the Grisu2 algorithm from Florian Loitsch which is licensed under the [MIT License](https://opensource.org/licenses/MIT) (see above). Copyright &copy; 2009 [Florian Loitsch](https://florian.loitsch.com/)
+- The class contains a port of the shortest double-to-decimal conversion of [Żmij](https://github.com/vitaut/zmij) by Victor Zverovich, which is licensed under the [MIT License](https://opensource.org/licenses/MIT) (see above). Copyright &copy; 2025 [Victor Zverovich](https://github.com/vitaut)
 - The class contains a copy of [Hedley](https://nemequ.github.io/hedley/) from Evan Nemerson which is licensed as [CC0-1.0](https://creativecommons.org/publicdomain/zero/1.0/).
 - The class contains parts of [Google Abseil](https://github.com/abseil/abseil-cpp) which is licensed under the [Apache 2.0 License](https://opensource.org/licenses/Apache-2.0).
 - The class contains an adapted version of the Eisel-Lemire algorithm and its table of powers of five from [fast_float](https://github.com/fastfloat/fast_float) by Daniel Lemire and contributors, which is available under the [MIT License](https://opensource.org/licenses/MIT) (used here), the Apache 2.0 License, and the Boost Software License. Copyright &copy; 2021 The fast_float authors
@@ -134,6 +134,7 @@ INSERT INTO searchIndex(name, type, path) VALUES ('basic_json_document::accept',
 INSERT INTO searchIndex(name, type, path) VALUES ('basic_json_document::erase', 'Method', 'api/basic_json_document/erase/index.html');
 INSERT INTO searchIndex(name, type, path) VALUES ('basic_json_document::insert', 'Method', 'api/basic_json_document/insert/index.html');
 INSERT INTO searchIndex(name, type, path) VALUES ('basic_json_document::is_discarded', 'Method', 'api/basic_json_document/is_discarded/index.html');
+INSERT INTO searchIndex(name, type, path) VALUES ('basic_json_document::load', 'Function', 'api/basic_json_document/load/index.html');
 INSERT INTO searchIndex(name, type, path) VALUES ('basic_json_document::memory_usage', 'Method', 'api/basic_json_document/memory_usage/index.html');
 INSERT INTO searchIndex(name, type, path) VALUES ('basic_json_document::node_count', 'Method', 'api/basic_json_document/node_count/index.html');
 INSERT INTO searchIndex(name, type, path) VALUES ('basic_json_document::owns_source', 'Method', 'api/basic_json_document/owns_source/index.html');
@@ -142,6 +143,7 @@ INSERT INTO searchIndex(name, type, path) VALUES ('basic_json_document::parse_co
 INSERT INTO searchIndex(name, type, path) VALUES ('basic_json_document::push_back', 'Method', 'api/basic_json_document/push_back/index.html');
 INSERT INTO searchIndex(name, type, path) VALUES ('basic_json_document::read', 'Method', 'api/basic_json_document/read/index.html');
 INSERT INTO searchIndex(name, type, path) VALUES ('basic_json_document::root', 'Method', 'api/basic_json_document/root/index.html');
+INSERT INTO searchIndex(name, type, path) VALUES ('basic_json_document::save', 'Method', 'api/basic_json_document/save/index.html');
 INSERT INTO searchIndex(name, type, path) VALUES ('basic_json_document::set', 'Method', 'api/basic_json_document/set/index.html');
 INSERT INTO searchIndex(name, type, path) VALUES ('basic_json_document::shrink_to_fit', 'Method', 'api/basic_json_document/shrink_to_fit/index.html');
 INSERT INTO searchIndex(name, type, path) VALUES ('basic_json_document::source', 'Method', 'api/basic_json_document/source/index.html');
@@ -60,6 +60,9 @@ Linear.

 ## Notes

+Floating-point numbers are written with the fewest digits that read back as the same value (for `#!cpp double`; see
+[number handling](../../features/types/number_handling.md#number-serialization)).
+
 Binary values are serialized as an object containing two keys:

 - "bytes": an array of bytes as integers
@@ -96,3 +99,5 @@ Binary values are serialized as an object containing two keys:
 - Indentation character `indent_char`, option `ensure_ascii` and exceptions added in version 3.0.0.
 - Error handlers added in version 3.4.0.
 - Serialization of binary values added in version 3.8.0.
+- Doubles are written with the shortest digits (Żmij instead of Grisu2) since version 3.13.0; about 0.1% of doubles are
+  written differently, most of them with fewer digits.
@@ -52,10 +52,16 @@ bookkeeping edits need, and calling any of them on one fails to compile (`#!cpp
 ## Member functions

 - [(constructor)](basic_json_document.md)
+
+### Parsing
+
 - [**parse**](parse.md) (_static_) - deserialize from a compatible input, borrowing or owning it as appropriate
 - [**parse_copy**](parse_copy.md) (_static_) - deserialize a copy of a compatible input
 - [**accept**](accept.md) (_static_) - check whether the input is valid JSON
 - [**read**](read.md) - (re-)parse into this document, reusing its memory
+
+### Access
+
 - [**root**](root.md) - the view of the root value
 - [**is_discarded**](is_discarded.md) - return whether the last parse failed
 - [**source**](source.md) - the parsed text
@@ -63,6 +69,14 @@ bookkeeping edits need, and calling any of them on one fails to compile (`#!cpp
 - [**node_count**](node_count.md) - the number of index entries (values plus object keys)
 - [**memory_usage**](memory_usage.md) - the number of bytes held by the document
 - [**shrink_to_fit**](shrink_to_fit.md) - release unused index capacity
+
+### Images
+
+- [**save**](save.md) - the document as an image that `load()` reads without parsing
+- [**load**](load.md) (_static_) - read an image written by `save()`
+
+### Edits
+
 - [**set**](set.md) - replace a value, or set an object member, an array element, or the value a JSON pointer refers
  to (`#!cpp Editable` documents only)
 - [**push_back**](push_back.md) - append to an array (`#!cpp Editable` documents only)
@@ -0,0 +1,169 @@
+# <small>nlohmann::basic_json_document::</small>load
+
+```cpp
+// (1)
+static basic_json_document load(const std::uint8_t* image, std::size_t size,
+                                 const image_check check = image_check::full);
+
+// (2)
+static basic_json_document load(const std::vector<std::uint8_t>& image,
+                                 const image_check check = image_check::full);
+
+// (3)
+static basic_json_document load(std::vector<std::uint8_t>&& image,
+                                 const image_check check = image_check::full);
+```
+
+1. Reads an image [`save()`](save.md) wrote, from a pointer and a byte count. The image is **borrowed**: `image`
+   must stay alive and unchanged for as long as the returned document, and any view taken from it, is used.
+2. Reads an image from a `#!cpp std::vector`. Also **borrowed** -- equivalent to overload 1 called with
+   `#!cpp image.data()` and `#!cpp image.size()`.
+3. Reads an image, keeping the vector instead of copying it: `image` is moved into the document (no copy), which
+   then owns it for as long as it needs the text and the decoded strings. [`owns_source()`](owns_source.md) is
+   `#!cpp true` afterward.
+
+In every overload, the node index is copied into storage the document itself owns -- so that it is properly aligned,
+and, for an [editable](index.md#edits) document, can be edited -- while the text and the decoded strings stay in
+`image`. The hash indexes [large objects](../../features/json_view.md) use for lookup are rebuilt, exactly as after
+parsing.
+
+## Parameters
+
+`image` (in)
+:   the image [`save()`](save.md) wrote (overloads 1 and 2), or one to take ownership of (overload 3)
+
+`size` (in)
+:   the number of bytes at `image` (overload 1)
+
+`check` (in)
+:   how thoroughly to validate `image` before trusting it; see [`image_check`](#image_check) below (optional,
+    `#!cpp image_check::full` by default)
+
+## Return value
+
+The document read from the image.
+
+## Exception safety
+
+Overloads 1 and 2 give the strong guarantee: `image` is only read, never written, so a thrown exception leaves the
+caller's buffer untouched.
+
+Overload 3 moves `image` into the document *before* validating it, so that a good image is kept without a copy. If
+loading then fails, the partially built document -- and the vector now inside it -- is discarded along with the
+exception, and `image` itself is left **empty**, not restored to what was passed in. Move a copy in instead, or
+validate with overload 2 first, if the original vector must survive a failed load.
+
+## Exceptions
+
+On a big-endian target, throws [`type_error.320`](../../home/exceptions.md#jsonexceptiontype_error320) -- the same
+exception [`save()`](save.md#exceptions) throws there, since the image format is little-endian only.
+
+Otherwise throws [`parse_error.116`](../../home/exceptions.md#jsonexceptionparse_error116) if `image` is not one
+`save()` could have written, or fails the requested `check`:
+
+| message               | when                                                                                                |
+|------------------------|------------------------------------------------------------------------------------------------------|
+| `too short`            | `image` is `#!cpp nullptr`, or `size` is smaller than the 64-byte header                             |
+| `unknown format`       | the header's magic bytes or version do not match, or a reserved header field is not zero             |
+| `sizes out of range`   | the node count, text size, or decoded-string size the header describes does not fit `size`, or the `#!cpp '\0'` after the text or after the decoded strings is missing |
+| `the check failed`     | `check` is not `#!cpp image_check::none`, and the image fails it -- see [`image_check`](#image_check) |
+
+!!! failure "Example messages"
+
+    ```
+    [json.exception.parse_error.116] parse error: invalid json_document image: too short
+    ```
+    ```
+    [json.exception.parse_error.116] parse error: invalid json_document image: unknown format
+    ```
+    ```
+    [json.exception.parse_error.116] parse error: invalid json_document image: sizes out of range
+    ```
+    ```
+    [json.exception.parse_error.116] parse error: invalid json_document image: the check failed
+    ```
+
+## Complexity
+
+Linear in the number of nodes, which are always copied into the document. With `#!cpp check == image_check::full`,
+additionally linear in the combined length of the text and the decoded strings; `#!cpp image_check::bounds` and
+`#!cpp image_check::none` do not read them.
+
+## `image_check`
+
+```cpp
+using image_check = detail::view::image_check;
+
+enum class image_check
+{
+    full,
+    bounds,
+    none
+};
+```
+
+How thoroughly `load()` validates `image` before trusting it.
+
+| value    | checks                                                                                                    | guarantees |
+|----------|--------------------------------------------------------------------------------------------------------------|------------|
+| `full`   | everything the parser itself guarantees: structure and bounds; that every string is valid UTF-8 (and, for a string still in the source text, that it contains no quote, backslash, or control character); and that every number token is well-formed and matches the value stored for it | reading and serializing a checked image is safe and always produces valid JSON, exactly as for a parsed document |
+| `bounds` | structure and bounds only -- that every offset and count in the node index stays inside the image            | reading and serializing stay memory-safe, but a crafted image can hold strings that are not valid UTF-8 or that serialize to invalid JSON ([`dump()`](../basic_json_view/dump.md) writes them unchanged or throws [`type_error.316`](../../home/exceptions.md#jsonexceptiontype_error316)), and numbers whose values differ from their text |
+| `none`   | nothing                                                                                                       | images from a trusted source only -- reading a damaged image is undefined behavior |
+
+`full` is the default and the right choice for an image from anything you do not fully control -- a file, a cache
+shared with other processes, a peer on the network. `bounds` skips scanning the text and the decoded strings, so it
+fits a cache your own process just wrote and reads straight back, where damage would mean a bug or a hardware fault
+rather than adversarial input; it still cannot crash or read out of bounds. `none` skips validation entirely and
+should only be used for an image you trust as much as your own memory.
+
+## Notes
+
+**Lifetime.** Overloads 1 and 2 borrow `image`: it must stay alive and byte-for-byte unchanged for as long as the
+returned document, and any [view](../basic_json_view/index.md) taken from it, is used -- exactly like a document
+[`parse()`](parse.md) borrowed its input for. Overload 3 avoids this by keeping the vector itself; see
+[`owns_source`](owns_source.md).
+
+!!! warning "Experimental"
+
+    The image format is versioned but not yet stable, and may change in an incompatible way before it is declared
+    stable; `load()` already rejects an image written by a different format version with `parse_error.116`
+    ("unknown format"). Use images to cache a document within one build of the library, or to hand one to another
+    process running the *same* build on the *same* (little-endian) machine -- not as a long-term storage format.
+
+**What `image_check::bounds` does not guarantee.** A bounds-checked image can never make `load()`,
+[`root()`](root.md), element access, or [`materialize()`](../basic_json_view/materialize.md) read outside the image,
+so those stay safe on a damaged one. It does *not* guarantee that the image describes valid JSON: a string
+that a `full` check would have rejected can make [`dump()`](../basic_json_view/dump.md) write invalid UTF-8 or invalid
+JSON, or throw `type_error.316`, and a number can read back with a value that does not match how it is spelled. Reserve `bounds` for images you already trust to be well-formed, and use it
+only to skip the extra scan.
+
+## Examples
+
+??? example "Caching a document, ownership, and a rejected image"
+
+    The example below saves a parsed document as an image, checks that `load()` reproduces the original
+    [`dump()`](../basic_json_view/dump.md) without parsing, and shows the difference between
+    `load(std::move(image))` (owned) and `load(image)` (borrowed). It then damages one byte of the image and shows
+    `image_check::full` rejecting it with `parse_error.116`, while `image_check::bounds` -- meant for a cache the
+    process already trusts -- still reads it without going out of bounds.
+
+    ```cpp
+    --8<-- "examples/basic_json_document__load.cpp"
+    ```
+
+    Output:
+
+    ```json
+    --8<-- "examples/basic_json_document__load.output"
+    ```
+
+## See also
+
+- [save](save.md) - write the document as an image
+- [owns_source](owns_source.md) - return whether the document holds its own copy of the text
+- [parse](parse.md) - deserialize from JSON text instead of an image
+- [Images](../../features/json_view.md#images) - why and when to use images
+
+## Version history
+
+- Added in version 3.13.0.
@@ -132,6 +132,7 @@ integer type becomes a floating-point value.
 - [accept](accept.md) - check whether the input is valid JSON
 - [read](read.md) - (re-)parse into this document, reusing its memory
 - [owns_source](owns_source.md) - return whether the document holds its own copy of the text
+- [load](load.md) - read a document from an image instead of parsing JSON text
 - [`BasicJsonType::parse`](../basic_json/parse.md) - the corresponding function of `basic_json`

 ## Version history
@@ -70,6 +70,7 @@ own on the next, since ownership is decided freshly each time.

 - [parse](parse.md) - deserialize from a compatible input
 - [root](root.md) - the view of the root value
+- [load](load.md) - read a document from an image instead of parsing JSON text

 ## Version history

@@ -0,0 +1,104 @@
+# <small>nlohmann::basic_json_document::</small>save
+
+```cpp
+std::vector<std::uint8_t> save() const;
+```
+
+Writes the document as an *image*: a byte buffer that [`load`](load.md) reads back without parsing. The image holds
+the node index, the source text (plus, for an edited document, the number tokens edits wrote), and the decoded
+strings (plus the strings edits wrote) -- everything [`root()`](root.md) needs, with nothing left to parse.
+
+An edited document is written in its *current* state, with its values in document order, the way the library's own
+parser would have produced them for that JSON text: a member [`set`](set.md) added goes at the end, an
+[`erase`](erase.md)d member leaves no trace, and a float that is not finite (NaN or positive/negative infinity)
+becomes null, the same substitution [`dump()`](../basic_json_view/dump.md) makes. The same document always saves to
+the same bytes -- also across `BasicJsonType` and `#!cpp Editable`, since the image reflects document order and
+values only, not which specialization produced them.
+
+## Return value
+
+The image, as a `#!cpp std::vector<std::uint8_t>`. Pass it, or a pointer to its data together with its size, to
+[`load`](load.md) to read the document back.
+
+## Exception safety
+
+Strong guarantee: `save()` does not modify `#!cpp *this` (it is `#!cpp const`), so if it throws, the document is left
+exactly as it was, and the partially built image is discarded with the exception.
+
+## Exceptions
+
+Throws [`type_error.320`](../../home/exceptions.md#jsonexceptiontype_error320) if the document is
+[discarded](is_discarded.md) -- a default-constructed document, or one a failed [`parse()`](parse.md)/
+[`read()`](read.md) with `allow_exceptions == false` left discarded.
+
+On a big-endian target, throws `type_error.320` with a different message instead: the image format is little-endian
+only (see [Notes](#notes)).
+
+Throws [`out_of_range.416`](../../home/exceptions.md#jsonexceptionout_of_range416) if the node count, the text, or
+the decoded strings of the image would individually reach 4 GiB -- the same 32-bit offsets
+[`parse()`](parse.md#exceptions) and, for edits, [`set`](set.md)/[`push_back`](push_back.md) are already limited to.
+
+!!! failure "Example messages"
+
+    ```
+    [json.exception.type_error.320] cannot save a discarded json_document
+    ```
+    ```
+    [json.exception.type_error.320] json_document images need a little-endian target
+    ```
+    ```
+    [json.exception.out_of_range.416] images of 4 GiB or more are not supported by json_document
+    ```
+
+## Complexity
+
+Linear in the size of the document: the number of nodes, plus the length of the text and the decoded strings that end
+up in the image.
+
+## Notes
+
+**Format.** The image begins with a 64-byte header (the magic bytes `#!cpp "NJVI"`, a version number, the node count,
+and the sizes of the text and the decoded strings, all little-endian), followed by the nodes
+([16 bytes each](../../home/architecture.md#node-index-of-json-views)), the text and a `#!cpp '\0'`, and the decoded
+strings and a `#!cpp '\0'`. [`load`](load.md) checks the header, and the sizes it describes, before reading anything
+else -- see [`load`'s Exceptions](load.md#exceptions).
+
+!!! warning "Experimental"
+
+    The image format is versioned but not yet stable: it may change in an incompatible way before it is declared
+    stable. Use images to cache a document within one build of the library, or to hand one to another process running
+    the *same* build on the *same* (little-endian) machine -- not as a long-term storage format. Keep the original
+    JSON text if you need to read a saved document back with a future library version.
+
+**Little-endian only.** The image is written as raw little-endian bytes, with no byte-swapping. `save()` (and
+[`load`](load.md)) throw `type_error.320` on a big-endian target rather than silently produce bytes a big-endian
+reader could not interpret correctly.
+
+## Examples
+
+??? example "Caching a parsed document as an image"
+
+    The example below saves a parsed configuration as an image -- the way a service might cache one to answer later
+    requests without parsing the text again -- and confirms that loading it back gives exactly the same result as
+    parsing did, and that saving is deterministic.
+
+    ```cpp
+    --8<-- "examples/basic_json_document__save.cpp"
+    ```
+
+    Output:
+
+    ```json
+    --8<-- "examples/basic_json_document__save.output"
+    ```
+
+## See also
+
+- [load](load.md) - read an image written by `save()`
+- [owns_source](owns_source.md) - return whether the document holds its own copy of the text
+- [`basic_json_view::dump`](../basic_json_view/dump.md) - serialize the document to JSON text instead of an image
+- [Images](../../features/json_view.md#images) - why and when to use images
+
+## Version history
+
+- Added in version 3.13.0.
@@ -0,0 +1,52 @@
+#include <cstdint>
+#include <iostream>
+#include <string>
+#include <vector>
+#include <nlohmann/json_view.hpp>
+
+using json = nlohmann::json;
+using json_document = nlohmann::json_document;
+using image_check = json_document::image_check;
+
+int main()
+{
+    std::cout << std::boolalpha;
+
+    // the image of a parsed document -- as if read back from a cache file or
+    // received from another process running the same build of the library
+    const std::string text = R"({"name": "cache", "note": "caf\u00e9", "replicas": ["db2", "db3"]})";
+    const json_document parsed = json_document::parse(text);
+    const std::vector<std::uint8_t> image = parsed.save();
+
+    // (1)/(2) load() needs no parsing, yet dumps exactly what parsing did
+    const json_document borrowed = json_document::load(image);
+    std::cout << (borrowed.root().dump() == parsed.root().dump()) << '\n';
+    std::cout << borrowed.owns_source() << '\n'; // borrowed: still points into `image`
+
+    // (3) load(std::move(image)) keeps the vector instead of copying it
+    std::vector<std::uint8_t> to_move = image;
+    const json_document owned = json_document::load(std::move(to_move));
+    std::cout << owned.owns_source() << '\n';
+
+    // a damaged image -- the last byte of the decoded string "note" holds
+    // (an escape sequence, so it was unescaped into the document's own
+    // buffer), flipped, as storage or transport corruption might do
+    std::vector<std::uint8_t> damaged = image;
+    damaged[damaged.size() - 2] = 0xFF;
+
+    // image_check::full inspects strings and numbers, so it catches the damage
+    try
+    {
+        static_cast<void>(json_document::load(damaged, image_check::full));
+    }
+    catch (const json::parse_error& e)
+    {
+        std::cout << e.id << '\n';
+    }
+
+    // image_check::bounds only checks structure and bounds, so a cache the
+    // process already trusts loads without the extra scan -- reading a value
+    // the damage did not touch is still safe
+    const json_document trusted = json_document::load(damaged, image_check::bounds);
+    std::cout << trusted.root()["name"].get<std::string>() << '\n';
+}
@@ -0,0 +1,5 @@
+true
+false
+true
+116
+cache
@@ -0,0 +1,30 @@
+#include <cstdint>
+#include <iostream>
+#include <string>
+#include <vector>
+#include <nlohmann/json_view.hpp>
+
+using json_document = nlohmann::json_document;
+
+int main()
+{
+    std::cout << std::boolalpha;
+
+    // a configuration a service parses once and then caches as an image, so
+    // that later requests can load() it instead of parsing the text again
+    const std::string text = R"({"name": "cache", "host": "db1", "port": 6379, "replicas": ["db2", "db3"]})";
+    const json_document config = json_document::parse(text);
+
+    // save() turns the parsed document into a byte buffer: a 64-byte header,
+    // the node index, the source text, and the decoded strings
+    const std::vector<std::uint8_t> image = config.save();
+    std::cout << image.size() << '\n';
+
+    // the same document always saves to the same bytes
+    std::cout << (image == json_document::parse(text).save()) << '\n';
+
+    // loading the image back needs no parsing, yet dumps exactly what
+    // parsing the text produced
+    const json_document reloaded = json_document::load(image);
+    std::cout << (reloaded.root().dump() == config.root().dump()) << '\n';
+}
@@ -0,0 +1,3 @@
+316
+true
+true
@@ -13,7 +13,8 @@ C++ types, and finally serialize it again.
  [SAX interface](parsing/sax_interface.md), and [error handling](parsing/parse_exceptions.md).
 - [Zero-copy JSON views](json_view.md) — read a JSON text through a flat index instead of building a `json` tree;
  strings and numbers stay in the input and are only decoded when needed.
-  [Editable documents](json_view.md#editing-a-document) can also be modified.
+  [Editable documents](json_view.md#editing-a-document) can also be modified, and [images](json_view.md#images) load a
+  parsed document again without parsing it.
 - [Comments](comments.md) and [trailing commas](trailing_commas.md) — opt-in relaxations of the JSON grammar.

 ## Accessing and modifying values
@@ -250,6 +250,57 @@ edits. See [`basic_json_document`'s Edits](../api/basic_json_document/index.md#e
 throws (the *basic* guarantee, not the strong one `dump()` and the read-only functions provide). How edits are kept in
 the index is described in the [architecture overview](../home/architecture.md#node-index-of-json-views).

+## Images
+
+[`save()`](../api/basic_json_document/save.md) writes a document as an *image*: a byte buffer that
+[`load()`](../api/basic_json_document/load.md) reads back into a document without parsing -- no lexing, no building
+the node index, nothing but copying the nodes and pointing the text and the decoded strings at the image. Where
+[`parse_copy()`](../api/basic_json_document/parse_copy.md) still has to scan the whole input,
+[`load()`](../api/basic_json_document/load.md) turns that scan into a copy of the node index alone.
+
+**Why.** A document that is parsed once and then read many times -- a configuration loaded at startup, a template
+rendered on every request, a large reference dataset a worker process needs in memory -- pays for parsing once but
+can amortize [`save()`](../api/basic_json_document/save.md)'s cost across every later load. That makes images useful
+for a cache: save a document the first time it is parsed (to a file, a shared-memory segment, an in-process cache),
+and [`load()`](../api/basic_json_document/load.md) it on every later use instead of parsing the source text again.
+They are just as useful for handing a parsed document to another process (or a forked worker) running the same build
+of the library, since [`load()`](../api/basic_json_document/load.md) turns the transfer into a copy of the node index
+plus pointers into the received bytes, not a re-parse.
+
+**Choosing a check.** [`load()`](../api/basic_json_document/load.md) takes an
+[`image_check`](../api/basic_json_document/load.md#image_check) that trades validation against speed:
+`image_check::full` (the default) checks everything the parser itself guarantees, so a checked image is exactly as
+safe to read and serialize as a freshly parsed document -- the right choice whenever the image did not come straight
+from this process's own [`save()`](../api/basic_json_document/save.md), such as a file or a network peer.
+`image_check::bounds` only checks structure and bounds -- cheaper, since it skips scanning the text and the decoded
+strings -- and fits a cache the process trusts, one it wrote and reads back itself. `image_check::none` skips
+validation entirely, for an image trusted as much as the process's own memory. See
+[`load()`'s Notes](../api/basic_json_document/load.md#notes) for exactly what each level does and does not guarantee.
+
+??? example "Example: cache a parsed configuration as an image"
+
+    ```cpp
+    --8<-- "examples/basic_json_document__save.cpp"
+    ```
+
+    Output:
+
+    ```json
+    --8<-- "examples/basic_json_document__save.output"
+    ```
+
+!!! warning "Experimental"
+
+    The image format is versioned but not yet stable, and may change in an incompatible way before it is declared
+    stable. It is little-endian only, and tied to the library build that wrote it -- use it to cache a document or to
+    hand one to another process running the *same* build, not as a long-term storage format; keep the original JSON
+    text if a saved document needs to be readable by a future library version.
+
+The idea of a document you can read without parsing comes from zero-copy formats such as
+[FlatBuffers](https://github.com/google/flatbuffers) and [YaFF](https://github.com/yandex/yaff); the check
+[`load()`](../api/basic_json_document/load.md) runs follows the idea of FlatBuffers' Verifier. No code is taken from
+either.
+
 ## Choosing between `json`, `ordered_json`, the SAX interface, and `json_view`

 | | [`json`](../api/json.md) / [`ordered_json`](../api/ordered_json.md) | [SAX interface](parsing/sax_interface.md) | [`json_document`](../api/json_document.md) / [`json_view`](../api/json_view.md) | [`json_editable_document`](../api/json_editable_document.md) / [`json_editable_view`](../api/json_editable_view.md) |
@@ -258,6 +309,7 @@ the index is described in the [architecture overview](../home/architecture.md#no
 | **Mutability** | freely mutable | not applicable (a one-shot event stream) | read-only | [`set`](../api/basic_json_document/set.md)/[`push_back`](../api/basic_json_document/push_back.md)/[`insert`](../api/basic_json_document/insert.md)/[`erase`](../api/basic_json_document/erase.md) edit in place; the source text is never rewritten |
 | **What you get** | a full tree you can read, write, and keep as long as you like | a sequence of callbacks; whatever your handler builds from them | a flat index plus, on demand, [`materialize()`](../api/basic_json_view/materialize.md)d `json`/`ordered_json` values for the parts you actually use | the same, plus [`dump()`](../api/basic_json_view/dump.md) of an edited document that keeps the member order and, with [`number_format::source`](../api/basic_json_view/number_format.md), the spelling of every untouched number |
 | **Typical use** | general-purpose JSON handling: config, request/response bodies you build or modify, anything you hold onto | validating or projecting a text into your own data structure without ever holding the whole thing as JSON | large or high-volume input where you only need part of it, or need it repeatedly, and can keep the source text (or a copy) alive for as long as the document lives | a document you read, patch a few fields of, and write back -- a configuration file, for instance -- where the rest of it should come back exactly as it was |
+| **Caching/reload** | not applicable -- re-parse, or roll your own serialization | not applicable | [`save()`](../api/basic_json_document/save.md)/[`load()`](../api/basic_json_document/load.md): cache the parsed index as an image and reload it without parsing | same, saving the document's current -- possibly edited -- state |

 ## Version history

@@ -118,9 +118,10 @@ That is, `-0` is stored as a signed integer, but the serialization does not repr
 ### Number serialization

 - Integer numbers are serialized as is; that is, no scientific notation is used.
- Floating-point numbers are serialized as specified by the `#!c %g` printf modifier with 
-  [`std::numeric_limits<double>::max_digits10`](https://en.cppreference.com/w/cpp/types/numeric_limits/max_digits10)
-  significant digits. The rationale is to use the shortest representation while still allowing round-tripping.
+- Floating-point numbers are serialized with the fewest digits that read back as the same value (the closest such
+  digits if there are several), in the layout of the `#!c %g` printf modifier: `#!c 1.5`, `#!c 100.0`, `#!c 1e+100`.
+  Doubles are converted with the algorithm of [Żmij](https://github.com/vitaut/zmij), floats with Grisu2, which
+  can write more digits than necessary.

 !!! hint "Notes regarding precision of floating-point numbers"

@@ -540,9 +540,10 @@ therefore silently changes parse results rather than raising an error. See
  specifiers, for which the library likewise provides only `#!cpp double` and `#!cpp long double` overloads
  (`#!cpp float` is promoted to `#!cpp double`).

-If `#!cpp std::numeric_limits<NumberFloatType>` describes an IEEE 754 binary32 or binary64 number, `dump` uses the
-Grisu2 algorithm, which produces the shortest representation that round-trips. Otherwise the `snprintf` fallback with
-`max_digits10` digits is used.
+If `#!cpp std::numeric_limits<NumberFloatType>` describes an IEEE 754 binary64 number, `dump` uses the algorithm of
+Żmij, which produces the shortest representation that round-trips. For IEEE 754 binary32 numbers, it uses Grisu2,
+which produces a short representation that round-trips. Otherwise the `snprintf` fallback with `max_digits10` digits is
+used.

 ### Required for the binary formats

@@ -554,7 +555,7 @@ binary32 or binary64 field and have no encoding for `#!cpp long double`.

 | Type                     | Support                                                                                                               |
 |--------------------------|-----------------------------------------------------------------------------------------------------------------------|
-| `#!cpp double` (default) | full; short round-trip output through Grisu2                                                                          |
+| `#!cpp double` (default) | full; shortest round-trip output through Żmij                                                                         |
 | `#!cpp float`            | full; short round-trip output through Grisu2                                                                          |
 | `#!cpp long double`      | `dump` and `parse` only; the binary format writers do not compile, as they only handle IEEE 754 binary32 and binary64 |
 | any other type           | not usable                                                                                                            |
@@ -259,6 +259,14 @@ edited:
 - Views of read-only documents compile without any of this: how views walk the index is a template parameter
  (`navigation<Editable>`).

+Images ([`save`](../api/basic_json_document/save.md) and [`load`](../api/basic_json_document/load.md),
+[`detail/view/image.hpp`](https://github.com/nlohmann/json/blob/develop/include/nlohmann/detail/view/image.hpp)) store
+the nodes as they are: a 64-byte header (the magic bytes `NJVI`, a format version, the sizes, and reserved bytes that
+must be zero), the nodes, the text, and the decoded strings. An edited document is first written in document order, as
+the parser would have written it (without links), and the numbers of the hash indexes are cleared, since `load`
+rebuilds the indexes. So a change of the node layout is a change of the image format: it must raise `image_version`,
+and `load` then rejects images of other versions (`parse_error.116`) instead of misreading them.
+
 ## Input adapters

 Input is read via **input adapters** that abstract a source. Every input adapter provides this interface:
@@ -388,6 +388,23 @@ A UBJSON high-precision number could not be parsed.
    [json.exception.parse_error.115] parse error at byte 5: syntax error while parsing UBJSON high-precision number: invalid number text: 1A
    ```

+### json.exception.parse_error.116
+
+[`basic_json_document::load()`](../api/basic_json_document/load.md) rejected an
+[image](../features/json_view.md#images): either the bytes are not one [`save()`](../api/basic_json_document/save.md)
+could have written (too short, an unknown magic number or format version, or sizes that do not fit the buffer), or
+they are, but fail the requested [`image_check`](../api/basic_json_document/load.md#image_check).
+
+!!! failure "Example message"
+
+    ```
+    [json.exception.parse_error.116] parse error: invalid json_document image: the check failed
+    ```
+
+!!! note
+
+    This exception was added in version 3.13.0, together with [images](../features/json_view.md#images).
+
 ## Iterator errors

 This exception is thrown if iterators passed to a library function do not match
@@ -800,6 +817,26 @@ from JSON text.

    This exception was added in version 3.13.0, together with editable [`json_document`s](../features/json_view.md).

+### json.exception.type_error.320
+
+[`basic_json_document::save()`](../api/basic_json_document/save.md) cannot write an
+[image](../features/json_view.md#images) of a [discarded](../api/basic_json_document/is_discarded.md) document.
+[`save()`](../api/basic_json_document/save.md) and [`load()`](../api/basic_json_document/load.md) also throw this
+exception on a big-endian target, since the image format is little-endian only.
+
+!!! failure "Example messages"
+
+    ```
+    [json.exception.type_error.320] cannot save a discarded json_document
+    ```
+    ```
+    [json.exception.type_error.320] json_document images need a little-endian target
+    ```
+
+!!! note
+
+    This exception was added in version 3.13.0, together with [images](../features/json_view.md#images).
+
 ## Out of range

 This exception is thrown in case a library function is called on an input parameter that exceeds the expected range, for instance, in the case of array indices or nonexisting object keys.
@@ -1027,7 +1064,9 @@ MessagePack's ext type and BSON's binary subtype are each stored in a single byt
 so they do not support an input of 4 GiB or more. The same 32-bit limit applies to an **editable** document's own
 storage: [`set`](../api/basic_json_document/set.md) and [`push_back`](../api/basic_json_document/push_back.md) throw
 this exception once the strings and number tokens written by edits reach 4 GiB in total, or once more than
-4294967295 arrays/objects have had an element set or appended to them.
+4294967295 arrays/objects have had an element set or appended to them. The same limit applies to an
+[image](../features/json_view.md#images): [`save()`](../api/basic_json_document/save.md) throws it if the node
+count, the text, or the decoded strings it would write would individually reach 4 GiB.

 !!! failure "Example messages"

@@ -1037,6 +1076,9 @@ this exception once the strings and number tokens written by edits reach 4 GiB i
    ```
    [json.exception.out_of_range.416] edits of 4 GiB or more are not supported by json_document
    ```
+    ```
+    [json.exception.out_of_range.416] images of 4 GiB or more are not supported by json_document
+    ```

 !!! note

@@ -18,6 +18,8 @@ The class contains the UTF-8 Decoder from Bjoern Hoehrmann which is licensed und

 The class contains a slightly modified version of the Grisu2 algorithm from Florian Loitsch which is licensed under the [MIT License](https://opensource.org/licenses/MIT) (see above). Copyright &copy; 2009 [Florian Loitsch](https://florian.loitsch.com/)

+The class contains a port of the shortest double-to-decimal conversion of [Żmij](https://github.com/vitaut/zmij) by Victor Zverovich, which is licensed under the [MIT License](https://opensource.org/licenses/MIT) (see above). Copyright &copy; 2025 [Victor Zverovich](https://github.com/vitaut)
+
 The class contains a copy of [Hedley](https://nemequ.github.io/hedley/) from Evan Nemerson which is licensed as [CC0-1.0](https://creativecommons.org/publicdomain/zero/1.0/).

 The class contains an adapted version of the Eisel-Lemire algorithm and its table of powers of five from [fast_float](https://github.com/fastfloat/fast_float) by Daniel Lemire and contributors, which is available under the [MIT License](https://opensource.org/licenses/MIT) (used here), the Apache 2.0 License, and the Boost Software License. Copyright &copy; 2021 The fast_float authors
@@ -236,6 +236,7 @@ nav:
          - 'erase': api/basic_json_document/erase.md
          - 'insert': api/basic_json_document/insert.md
          - 'is_discarded': api/basic_json_document/is_discarded.md
+          - 'load': api/basic_json_document/load.md
          - 'memory_usage': api/basic_json_document/memory_usage.md
          - 'node_count': api/basic_json_document/node_count.md
          - 'owns_source': api/basic_json_document/owns_source.md
@@ -244,6 +245,7 @@ nav:
          - 'push_back': api/basic_json_document/push_back.md
          - 'read': api/basic_json_document/read.md
          - 'root': api/basic_json_document/root.md
+          - 'save': api/basic_json_document/save.md
          - 'set': api/basic_json_document/set.md
          - 'shrink_to_fit': api/basic_json_document/shrink_to_fit.md
          - 'source': api/basic_json_document/source.md
@@ -11,11 +11,17 @@

 #include <array> // array
 #include <cmath>   // signbit, isfinite
+#include <cstddef> // size_t
 #include <cstdint> // intN_t, uintN_t
 #include <cstring> // memcpy, memmove
 #include <limits> // numeric_limits
 #include <type_traits> // conditional

+#ifdef _MSC_VER
+    #include <cstdlib> // _byteswap_uint64
+#endif
+
+#include <nlohmann/detail/conversions/zmij.hpp>
 #include <nlohmann/detail/macro_scope.hpp>

 NLOHMANN_JSON_NAMESPACE_BEGIN
@@ -918,6 +924,87 @@ void grisu2(char* buf, int& len, int& decimal_exponent, FloatType value)
    grisu2(buf, len, decimal_exponent, w.minus, w.w, w.plus);
 }

+/*!
+@brief the shortest digits of a positive finite float (other than double): Grisu2
+*/
+template<typename FloatType>
+JSON_HEDLEY_NON_NULL(1)
+void shortest_digits(char* buf, int& len, int& decimal_exponent, FloatType value)
+{
+    grisu2(buf, len, decimal_exponent, value);
+}
+
+/*!
+@brief the shortest digits of a positive finite double: the conversion of
+Zmij (see zmij.hpp), which always finds the shortest digits that read back as
+the same value (Grisu2 does not for about one double in a thousand), and the
+closest of them if there are several
+
+v = buf * 10^decimal_exponent, as for grisu2()
+*/
+JSON_HEDLEY_NON_NULL(1)
+inline void shortest_digits(char* buf, int& len, int& decimal_exponent, double value)
+{
+    static_assert(std::numeric_limits<double>::is_iec559 && std::numeric_limits<double>::digits == 53,
+                  "internal error: the conversion of Zmij needs IEEE 754 binary64 doubles");
+    JSON_ASSERT(std::isfinite(value));
+    JSON_ASSERT(value > 0);
+
+    std::uint64_t bits = 0;
+    std::memcpy(&bits, &value, sizeof(bits));
+    zmij::decimal d = zmij::to_decimal(bits);
+    // without trailing zeros (up to 16): 8, 4, 2, 1 at a time
+    while (d.significand % 100000000 == 0)
+    {
+        d.significand /= 100000000;
+        d.exponent += 8;
+    }
+    if (d.significand % 10000 == 0)
+    {
+        d.significand /= 10000;
+        d.exponent += 4;
+    }
+    if (d.significand % 100 == 0)
+    {
+        d.significand /= 100;
+        d.exponent += 2;
+    }
+    if (d.significand % 10 == 0)
+    {
+        d.significand /= 10;
+        d.exponent += 1;
+    }
+    // at most 17 digits, written from the back two at a time
+    static constexpr const char* pairs =
+        "00010203040506070809101112131415161718192021222324252627282930313233343536373839"
+        "40414243444546474849505152535455565758596061626364656667686970717273747576777879"
+        "8081828384858687888990919293949596979899";
+    std::array<char, 20> digits{};
+    std::size_t n = digits.size();
+    while (d.significand >= 100)
+    {
+        const auto i = static_cast<std::size_t>(d.significand % 100) * 2;
+        d.significand /= 100;
+        n -= 2;
+        digits[n] = pairs[i];
+        digits[n + 1] = pairs[i + 1];
+    }
+    if (d.significand >= 10)
+    {
+        const auto i = static_cast<std::size_t>(d.significand) * 2;
+        n -= 2;
+        digits[n] = pairs[i];
+        digits[n + 1] = pairs[i + 1];
+    }
+    else
+    {
+        digits[--n] = static_cast<char>('0' + d.significand);
+    }
+    len = static_cast<int>(digits.size() - n);
+    std::memcpy(buf, digits.data() + n, static_cast<std::size_t>(len));
+    decimal_exponent = d.exponent;
+}
+
 /*!
@brief appends a decimal representation of e to buf
@return a pointer to the element following the exponent.
@@ -1047,6 +1134,177 @@ inline char* format_buffer(char* buf, int len, int decimal_exponent,
    return append_exponent(buf, n - 1);
 }

+/// eight decimal digits (a value below 10^8) as bytes 0..9, the first digit
+/// in the most significant byte: three steps that divide all lanes at once
+/// by a multiplication (the conversion of Xiang JunBo, as in Zmij)
+inline std::uint64_t eight_digit_bytes(std::uint64_t abcdefgh) noexcept
+{
+    const std::uint64_t abcd_efgh = abcdefgh + (((std::uint64_t{1} << 32u) - 10000u) * ((abcdefgh * (((std::uint64_t{1} << 40u) / 10000u) + 1u)) >> 40u));
+    const std::uint64_t ab_cd_ef_gh = abcd_efgh + (((std::uint64_t{1} << 16u) - 100u) * (((abcd_efgh * (((std::uint64_t{1} << 19u) / 100u) + 1u)) >> 19u) & 0x7F0000007Fu));
+    return ab_cd_ef_gh + (((std::uint64_t{1} << 8u) - 10u) * (((ab_cd_ef_gh * (((std::uint64_t{1} << 10u) / 10u) + 1u)) >> 10u) & 0x000F000F000F000Fu));
+}
+
+/// store the bytes of v, the most significant one first (one byte swap and
+/// one store where the byte order is known: compilers do not reliably merge
+/// the byte stores once this is inlined)
+inline void store_msb_first(char* p, std::uint64_t v) noexcept
+{
+#if defined(__BYTE_ORDER__) && defined(__ORDER_LITTLE_ENDIAN__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+    v = __builtin_bswap64(v);
+    std::memcpy(p, &v, sizeof(v));
+#elif defined(__BYTE_ORDER__) && defined(__ORDER_BIG_ENDIAN__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
+    std::memcpy(p, &v, sizeof(v));
+#elif defined(_MSC_VER) // (little-endian on all its targets)
+    v = _byteswap_uint64(v);
+    std::memcpy(p, &v, sizeof(v));
+#else
+    for (unsigned i = 0; i < 8; ++i)
+    {
+        p[i] = static_cast<char>(v >> (56u - (8u * i)));
+    }
+#endif
+}
+
+/*!
+@brief digits * 10^exp for a double, in the layout of format_buffer()
+
+The layout is that of format_buffer() with min_exp -4 and max_exp 15 (the
+digits10 of double). The digits are converted eight at a time and placed
+with fixed-size moves instead of per-digit loops and moves of the buffer.
+
+@param[in] digits  the digits (not 0, at most 17 digits; trailing zeros allowed)
+@param[in] exp     the decimal exponent of the last digit
+@return a pointer past the text; up to 41 bytes at @a first are written
+        (some beyond the returned end)
+*/
+JSON_HEDLEY_NON_NULL(1)
+JSON_HEDLEY_RETURNS_NON_NULL
+inline char* write_decimal(char* first, std::uint64_t digits, int exp) noexcept
+{
+    JSON_ASSERT(digits != 0 && digits < 100000000000000000u);
+    const std::uint64_t upper = digits / 100000000u;
+    const std::uint64_t b0 = upper / 100000000u; // (one digit: it is its own byte)
+    const std::uint64_t b1 = eight_digit_bytes(upper % 100000000u);
+    const std::uint64_t b2 = eight_digit_bytes(digits % 100000000u);
+    // leading and trailing zero digits: zero bytes, counted without division
+    int leading = 16;
+    int zeros = 16;
+    if (b0 != 0)
+    {
+        leading = count_leading_zeros(b0) / 8;
+    }
+    else if (b1 != 0)
+    {
+        leading = 8 + (count_leading_zeros(b1) / 8);
+    }
+    else
+    {
+        leading += count_leading_zeros(b2) / 8;
+    }
+    if (b2 != 0)
+    {
+        zeros = count_trailing_zeros(b2) / 8;
+    }
+    else if (b1 != 0)
+    {
+        zeros = 8 + (count_trailing_zeros(b1) / 8);
+    }
+    // (else: 16, b0 is the one digit that is not 0)
+    // the digits as text at text + leading, then '0's, so that fixed-size
+    // moves need not check how many digits there are
+    std::array<char, 64> text; // NOLINT(cppcoreguidelines-pro-type-member-init,hicpp-member-init): written before read
+    store_msb_first(text.data(), b0 + 0x3030303030303030u);
+    store_msb_first(text.data() + 8, b1 + 0x3030303030303030u);
+    store_msb_first(text.data() + 16, b2 + 0x3030303030303030u);
+    std::memset(text.data() + 24, '0', 40);
+    const int k = 24 - leading - zeros; // significant digits
+    const int n = k + exp + zeros;      // position of the decimal point after the first digit
+    const char* const s0 = text.data() + leading;
+
+    if (-4 < n && n <= 15)
+    {
+        // "0.[000]digits" (n <= 0) is the digits after 1 - n leading '0's
+        // with the point after the first; "digits[000].0" (n >= k) and
+        // "dig.its" put the point after n characters
+        const int pad = n <= 0 ? 1 - n : 0;
+        const char* const s = s0 - pad;
+        const int len = k + pad;
+        const int point = n + pad;
+        std::memcpy(first, s, 16);
+        std::memcpy(first + point + 1, s + point, 24);
+        first[point] = '.';
+        return first + (point >= len ? point + 2 : len + 1);
+    }
+
+    // d.igitse+XX, with at least two exponent digits (as append_exponent())
+    std::memcpy(first, s0, 16);
+    std::memcpy(first + 2, s0 + 1, 16);
+    first[1] = '.';
+    char* const end = first + (k == 1 ? 1 : k + 1);
+    const int e = n - 1;
+    const auto ea = static_cast<unsigned>(e < 0 ? -e : e);
+    const bool three = ea >= 100;
+    end[0] = 'e';
+    end[1] = e < 0 ? '-' : '+';
+    end[2] = static_cast<char>('0' + (three ? ea / 100 : (ea / 10) % 10));
+    end[3] = static_cast<char>('0' + (three ? (ea / 10) % 10 : ea % 10));
+    end[4] = static_cast<char>('0' + (ea % 10));
+    return end + (three ? 5 : 4);
+}
+
+/// a positive finite float (other than double): Grisu2 and format_buffer()
+template<typename FloatType>
+JSON_HEDLEY_NON_NULL(1, 2)
+JSON_HEDLEY_RETURNS_NON_NULL
+char* write_positive(char* first, const char* last, FloatType value)
+{
+    JSON_ASSERT(last - first >= std::numeric_limits<FloatType>::max_digits10);
+
+    // Compute v = buffer * 10^decimal_exponent.
+    // The decimal digits are stored in the buffer, which needs to be interpreted
+    // as an unsigned decimal integer.
+    // len is the length of the buffer, i.e., the number of decimal digits.
+    int len = 0;
+    int decimal_exponent = 0;
+    shortest_digits(first, len, decimal_exponent, value);
+
+    JSON_ASSERT(len <= std::numeric_limits<FloatType>::max_digits10);
+
+    // Format the buffer like printf("%.*g", prec, value)
+    constexpr int kMinExp = -4;
+    // Use digits10 here to increase compatibility with version 2.
+    constexpr int kMaxExp = std::numeric_limits<FloatType>::digits10;
+
+    JSON_ASSERT(last - first >= kMaxExp + 2);
+    JSON_ASSERT(last - first >= 2 + (-kMinExp - 1) + std::numeric_limits<FloatType>::max_digits10);
+    JSON_ASSERT(last - first >= std::numeric_limits<FloatType>::max_digits10 + 6);
+
+    return format_buffer(first, len, decimal_exponent, kMinExp, kMaxExp);
+}
+
+/// a positive finite double: the shortest digits (Zmij), laid out by
+/// write_decimal() (through a local buffer if [first, last) is shorter than
+/// the 41 bytes it may write)
+JSON_HEDLEY_NON_NULL(1, 2)
+JSON_HEDLEY_RETURNS_NON_NULL
+inline char* write_positive(char* first, const char* last, double value)
+{
+    static_assert(std::numeric_limits<double>::is_iec559 && std::numeric_limits<double>::digits == 53,
+                  "internal error: the conversion of Zmij needs IEEE 754 binary64 doubles");
+    std::uint64_t bits = 0;
+    std::memcpy(&bits, &value, sizeof(bits));
+    const zmij::decimal d = zmij::to_decimal(bits);
+    if (JSON_HEDLEY_LIKELY(last - first >= 41))
+    {
+        return write_decimal(first, d.significand, d.exponent);
+    }
+    std::array<char, 64> buf; // NOLINT(cppcoreguidelines-pro-type-member-init,hicpp-member-init): written before read
+    const auto len = static_cast<std::size_t>(write_decimal(buf.data(), d.significand, d.exponent) - buf.data());
+    JSON_ASSERT(static_cast<std::size_t>(last - first) >= len);
+    std::memcpy(first, buf.data(), len);
+    return first + len;
+}
+
 }  // namespace dtoa_impl

 /*!
@@ -1064,7 +1322,6 @@ JSON_HEDLEY_NON_NULL(1, 2)
 JSON_HEDLEY_RETURNS_NON_NULL
 char* to_chars(char* first, const char* last, FloatType value)
 {
-    static_cast<void>(last); // maybe unused - fix warning
    JSON_ASSERT(std::isfinite(value));

    // Use signbit(value) instead of (value < 0) since signbit works for -0.
@@ -1090,28 +1347,7 @@ char* to_chars(char* first, const char* last, FloatType value)
    JSON_HEDLEY_DIAGNOSTIC_POP
 #endif

-    JSON_ASSERT(last - first >= std::numeric_limits<FloatType>::max_digits10);
-
-    // Compute v = buffer * 10^decimal_exponent.
-    // The decimal digits are stored in the buffer, which needs to be interpreted
-    // as an unsigned decimal integer.
-    // len is the length of the buffer, i.e., the number of decimal digits.
-    int len = 0;
-    int decimal_exponent = 0;
-    dtoa_impl::grisu2(first, len, decimal_exponent, value);
-
-    JSON_ASSERT(len <= std::numeric_limits<FloatType>::max_digits10);
-
-    // Format the buffer like printf("%.*g", prec, value)
-    constexpr int kMinExp = -4;
-    // Use digits10 here to increase compatibility with version 2.
-    constexpr int kMaxExp = std::numeric_limits<FloatType>::digits10;
-
-    JSON_ASSERT(last - first >= kMaxExp + 2);
-    JSON_ASSERT(last - first >= 2 + (-kMinExp - 1) + std::numeric_limits<FloatType>::max_digits10);
-    JSON_ASSERT(last - first >= std::numeric_limits<FloatType>::max_digits10 + 6);
-
-    return dtoa_impl::format_buffer(first, len, decimal_exponent, kMinExp, kMaxExp);
+    return dtoa_impl::write_positive(first, last, value);
 }

 }  // namespace detail
@@ -0,0 +1,218 @@
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.12.0
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2025 Victor Zverovich <https://github.com/vitaut/zmij>
+// SPDX-FileCopyrightText: 2013-2026 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include <array> // array
+#include <cstddef> // size_t
+#include <cstdint> // uint32_t, uint64_t
+
+#include <nlohmann/detail/abi_macros.hpp>
+#include <nlohmann/detail/bit_ops.hpp>
+#include <nlohmann/detail/input/pow5_table.hpp>
+#include <nlohmann/detail/macro_scope.hpp>
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+/*!
+@brief the shortest decimal representation of a double
+
+A C++11 port of the conversion of Zmij by Victor Zverovich
+(https://github.com/vitaut/zmij, MIT license): the shortest decimal in the
+rounding interval of a double, the closest one if there are several. Zmij
+credits Xiang JunBo (producing the shorter candidate without a division) and
+Dougall Johnson (the compressed powers of ten). The powers of ten are taken
+from the table for number parsing (pow5_table.hpp) where it holds them, and
+computed from the compressed tables of Zmij beyond it.
+*/
+namespace zmij
+{
+
+/// significand * 10^exponent
+struct decimal
+{
+    std::uint64_t significand;
+    int exponent;
+};
+
+/// the compressed powers of ten of Zmij
+inline const std::array<std::uint64_t, 28>& pow10_minor() noexcept
+{
+    static const std::array<std::uint64_t, 28> table =
+    {
+        {
+            0x8000000000000000u, 0xa000000000000000u, 0xc800000000000000u, 0xfa00000000000000u, 0x9c40000000000000u,
+            0xc350000000000000u, 0xf424000000000000u, 0x9896800000000000u, 0xbebc200000000000u, 0xee6b280000000000u,
+            0x9502f90000000000u, 0xba43b74000000000u, 0xe8d4a51000000000u, 0x9184e72a00000000u, 0xb5e620f480000000u,
+            0xe35fa931a0000000u, 0x8e1bc9bf04000000u, 0xb1a2bc2ec5000000u, 0xde0b6b3a76400000u, 0x8ac7230489e80000u,
+            0xad78ebc5ac620000u, 0xd8d726b7177a8000u, 0x878678326eac9000u, 0xa968163f0a57b400u, 0xd3c21bcecceda100u,
+            0x84595161401484a0u, 0xa56fa5b99019a5c8u, 0xcecb8f27f4200f3au
+        }
+    };
+    return table;
+}
+
+/// (high, low) pairs
+inline const std::array<std::uint64_t, 50>& pow10_major() noexcept
+{
+    static const std::array<std::uint64_t, 50> table =
+    {
+        {
+            0xaddcb9e83c6b1793u, 0xdf4abe242a1bbf3eu, 0xaf8e5410288e1b6fu, 0x07ecf0ae5ee44ddau, 0xb1442798f49ffb4au, 0x99cd11cfdf41779du,
+            0xb2fe3f0b8599ef07u, 0x861fa7e6dcb4aa15u, 0xb4bca50b065abe63u, 0x0fed077a756b53aau, 0xb67f6455292cbf08u, 0x1a3bc84c17b1d543u,
+            0xb84687c269ef3bfbu, 0x3d5d514f40eea742u, 0xba121a4650e4ddebu, 0x92f34d62616ce413u, 0xbbe226efb628afeau, 0x890489f70a55368cu,
+            0xbdb6b8e905cb600fu, 0x5400e987bbc1c921u, 0xbf8fdb78849a5f96u, 0xde98520472bdd034u, 0xc16d9a0095928a27u, 0x75b7053c0f178294u,
+            0xc350000000000000u, 0x0000000000000000u, 0xc5371912364ce305u, 0x6c28000000000000u, 0xc722f0ef9d80aad6u, 0x424d3ad2b7b97ef6u,
+            0xc913936dd571c84cu, 0x03bc3a19cd1e38eau, 0xcb090c8001ab551cu, 0x5cadf5bfd3072cc6u, 0xcd036837130890a1u, 0x36dba887c37a8c10u,
+            0xcf02b2c21207ef2eu, 0x94f967e45e03f4bcu, 0xd106f86e69d785c7u, 0xe13336d701beba52u, 0xd31045a8341ca07cu, 0x1ede48111209a051u,
+            0xd51ea6fa85785631u, 0x552a74227f3ea566u, 0xd732290fbacaf133u, 0xa97c177947ad4096u, 0xd94ad8b1c7380874u, 0x18375281ae7822bdu,
+            0xdb68c2ca82ed2a05u, 0xa67398db9f6820e1u
+        }
+    };
+    return table;
+}
+
+/// one bit per power: whether the computed value is one unit too large
+inline const std::array<std::uint32_t, 21>& pow10_fixups() noexcept
+{
+    static const std::array<std::uint32_t, 21> table =
+    {
+        {
+            0x8d8fc810u, 0x06100293u, 0x19000000u, 0x00100000u, 0x00000908u, 0x00000000u, 0x04e00300u, 0x3807e0b2u, 0x3d83d793u, 0x0006f5ccu,
+            0x00000000u, 0xffff0000u, 0x8076337du, 0x4ff45ba0u, 0x09405033u, 0x034376d9u, 0x09000000u, 0x4e100501u, 0x076d14dcu, 0xf964f45eu,
+            0x0000003du
+        }
+    };
+    return table;
+}
+
+/// the 128-bit significand of 10^k, rounded down, for k in [-307, 341]
+/// (compute_pow10 of Zmij)
+inline uint128_parts compute_pow10(int k) noexcept
+{
+    const auto i = static_cast<unsigned>(k + 307);
+    const std::uint64_t m = pow10_minor()[(i + 24) % 28];
+    const std::size_t j = 2 * static_cast<std::size_t>((i + 24) / 28);
+    const std::uint64_t h_hi = pow10_major()[j];
+    const std::uint64_t h_lo = pow10_major()[j + 1];
+    const std::uint64_t h1 = full_multiplication(h_lo, m).high;
+    const std::uint64_t c0 = h_lo * m;
+    const std::uint64_t c1 = h1 + (h_hi * m);
+    const std::uint64_t c2 = (c1 < h1 ? 1u : 0u) + full_multiplication(h_hi, m).high;
+    uint128_parts r{};
+    if ((c2 >> 63u) != 0)
+    {
+        r.high = c2;
+        r.low = c1;
+    }
+    else
+    {
+        r.high = (c2 << 1u) | (c1 >> 63u);
+        r.low = (c1 << 1u) | (c0 >> 63u);
+    }
+    r.low -= (pow10_fixups()[i >> 5u] >> (i & 31u)) & 1u;
+    return r;
+}
+
+/// The 128-bit significand of 10^k, rounded down, for k in [-342, 341].
+/// Up to 10^308, the table for number parsing holds the same significands
+/// (those of 5^k), except for k in [-27, -1], where it holds them one unit
+/// larger (as the Eisel-Lemire algorithm needs them).
+inline uint128_parts pow10(int k) noexcept
+{
+    if (k > pow5_128_largest_power)
+    {
+        return compute_pow10(k); // (only for the smallest doubles)
+    }
+    const auto i = 2 * static_cast<std::size_t>(k - pow5_128_smallest_power);
+    uint128_parts r{pow5_128()[i + 1], pow5_128()[i]};
+    const std::uint64_t adjust = static_cast<unsigned>(k + 27) < 27u ? 1u : 0u;
+    r.high -= r.low < adjust ? 1u : 0u;
+    r.low -= adjust;
+    return r;
+}
+
+/// (x_hi * 2^64 + x_lo) * y >> 64, as 128 bits
+inline uint128_parts umul192_hi128(std::uint64_t x_hi, std::uint64_t x_lo, std::uint64_t y) noexcept
+{
+    const uint128_parts p = full_multiplication(x_hi, y);
+    uint128_parts r{};
+    r.low = p.low + full_multiplication(x_lo, y).high;
+    r.high = p.high + (r.low < p.low ? 1u : 0u);
+    return r;
+}
+
+/// (x * y + c) >> 64
+inline std::uint64_t umul128_add_hi64(std::uint64_t x, std::uint64_t y, std::uint64_t c) noexcept
+{
+    const uint128_parts p = full_multiplication(x, y);
+    return p.high + (p.low + c < p.low ? 1u : 0u);
+}
+
+/// The shortest decimal in the rounding interval of a positive finite double
+/// given by its bits, the closest one if there are several (to_decimal of
+/// Zmij). The significand can end in zeros.
+inline decimal to_decimal(std::uint64_t bits) noexcept
+{
+    constexpr int extra_shift = 9;
+    const auto raw_exp = static_cast<int>((bits >> 52u) & 0x7FFu);
+    std::uint64_t bin_sig = bits & ((std::uint64_t{1} << 52u) - 1);
+    // a power of two has a narrower interval below (except the smallest normal)
+    const bool regular = bin_sig != 0 || raw_exp <= 1;
+    const int bin_exp = (raw_exp == 0 ? 1 : raw_exp) - 1075;
+    if (raw_exp != 0)
+    {
+        bin_sig |= std::uint64_t{1} << 52u;
+    }
+    // floor(log10(2^bin_exp)), or floor(log10(3/4 * 2^bin_exp)) for the irregular case
+    const int dec_exp = ((bin_exp * 315653) - (regular ? 0 : 131072)) >> 20;
+    // scaled by 10^(-dec_exp - 1): the integral part is the shorter candidate
+    const int shift = bin_exp + ((-(dec_exp + 1) * 217707) >> 16) + 1 + extra_shift;
+    const uint128_parts p10 = pow10(-dec_exp - 1);
+    const uint128_parts p = umul192_hi128(p10.high, p10.low, bin_sig << static_cast<unsigned>(shift));
+    std::uint64_t integral = p.high >> static_cast<unsigned>(extra_shift);
+    const std::uint64_t fractional = (p.high << static_cast<unsigned>(64 - extra_shift)) | (p.low >> static_cast<unsigned>(extra_shift));
+    std::uint64_t digit = 0;
+    bool round_up = false;
+    bool round_down = false;
+    if (JSON_HEDLEY_LIKELY(regular))
+    {
+        const std::uint64_t half_ulp = (p10.high >> static_cast<unsigned>(extra_shift + 1 - shift)) + (1 - (bin_sig & 1u));
+        round_up = fractional + half_ulp < fractional;
+        round_down = half_ulp > fractional;
+        // the last digit of the longer candidate, rounded to nearest
+        digit = umul128_add_hi64(fractional, 10, (std::uint64_t{1} << 63u) + 6);
+        if (fractional == (std::uint64_t{1} << 62u))
+        {
+            digit = 2; // 2.5 rounds to 2
+        }
+    }
+    else
+    {
+        const std::uint64_t half_ulp = p10.high >> static_cast<unsigned>(extra_shift + 1 - shift);
+        round_up = half_ulp > ~std::uint64_t{0} - fractional;
+        round_down = (half_ulp >> 1u) > fractional;
+        digit = umul128_add_hi64(fractional, 10, (std::uint64_t{1} << 63u) - 1);
+        const std::uint64_t lowest = umul128_add_hi64(fractional - (half_ulp >> 1u), 10, ~std::uint64_t{0});
+        digit = digit < lowest ? lowest : digit;
+    }
+    integral += round_up ? 1u : 0u;
+    if (!round_up && !round_down)
+    {
+        // the shorter candidate is outside the rounding interval: one digit more
+        return decimal{(integral * 10) + digit, dec_exp};
+    }
+    return decimal{integral, dec_exp + 1};
+}
+
+}  // namespace zmij
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
@@ -10,7 +10,7 @@

 #include <array> // array
 #include <cstddef> // size_t
-#include <cstdint> // uint32_t
+#include <cstdint> // uint8_t, uint32_t
 #include <cstring> // memcpy
 #include <functional> // less
 #include <map> // map
@@ -41,7 +41,9 @@ struct document_data
    node* inline_tape = nullptr; ///< node array allocated together with this header
    std::size_t inline_cap = 0;
    std::string arena{}; ///< decoded strings that contained escapes // NOLINT(readability-redundant-member-init)
+    std::size_t arena_size = 0; ///< bytes of decoded strings at base[1] (the arena, or those of a loaded image)
    std::string owned{}; ///< owned copy of the input, if any // NOLINT(readability-redundant-member-init)
+    std::vector<std::uint8_t> owned_image{}; ///< a loaded image the document owns (the text and the decoded strings point into it) // NOLINT(readability-redundant-member-init)

    // hash indexes of large objects (see object_index.hpp)
    static constexpr std::uint32_t index_min_members = 128;
@@ -0,0 +1,602 @@
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  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 <array> // array
+#include <cstddef> // size_t
+#include <cstdint> // int64_t, uint8_t, uint16_t, uint32_t, uint64_t
+#include <cstring> // memcmp, memcpy
+#include <limits> // numeric_limits
+#include <string> // string
+#include <vector> // vector
+
+#include <nlohmann/json.hpp>
+#include <nlohmann/detail/view/document_data.hpp>
+#include <nlohmann/detail/view/errors.hpp>
+#include <nlohmann/detail/view/macro_scope.hpp>
+#include <nlohmann/detail/view/node.hpp>
+#include <nlohmann/detail/view/number.hpp>
+#include <nlohmann/detail/view/object_index.hpp>
+#include <nlohmann/detail/view/scan.hpp>
+
+// Images: a document stored so that loading it needs no parsing.
+//
+// Layout (little-endian): a 64-byte header, the nodes, the text (the source,
+// followed by the number tokens written by edits), a NUL, the decoded strings
+// (followed by the strings written by edits), a NUL. The idea is that of
+// zero-copy formats such as FlatBuffers (https://github.com/google/flatbuffers)
+// and YaFF (https://github.com/yandex/yaff); no code is taken from them.
+// check_image follows the idea of FlatBuffers' Verifier (bounds and
+// structure) and also checks what the parser guarantees about strings and
+// numbers, so that reading and serializing a checked image is safe and yields
+// valid JSON.
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+namespace view
+{
+
+/// how load() checks an image
+enum class image_check
+{
+    /// everything the parser guarantees: structure and bounds, strings (valid
+    /// UTF-8; source strings without quotes, backslashes, and control
+    /// characters), and numbers (well-formed, matching the stored values)
+    full,
+    /// structure and bounds only: reading and serializing are safe, but a
+    /// crafted image can yield invalid UTF-8, strings that serialize to
+    /// invalid JSON, or numbers that differ from their text
+    bounds,
+    /// none: for images from a trusted source only (a damaged image is
+    /// undefined behavior)
+    none,
+};
+
+struct image_header
+{
+    std::array<char, 4> magic;              ///< "NJVI"
+    std::uint32_t version;                  ///< 1
+    std::uint64_t node_count;
+    std::uint64_t text_size;
+    std::uint64_t arena_size;
+    std::array<std::uint64_t, 4> reserved;  ///< zero (for later versions)
+};
+static_assert(sizeof(image_header) == 64, "the image header must be 64 bytes");
+
+constexpr std::uint32_t image_version = 1;
+
+/// the largest node count and text or string size of an image (as for parsed
+/// documents, offsets and counts must fit 32 bits)
+constexpr std::uint64_t image_limit = 0xFFFFFFF0u;
+
+/// Copy the current structure of an edited document into nodes in document
+/// order, as the parser would have written them. Text written by edits is
+/// appended to text_tail (number tokens) and arena_tail (strings); floats that
+/// are not finite become null, as dump() writes them.
+inline void compact_nodes(const document_data& d, std::size_t arena_size, std::vector<node>& out, std::string& text_tail, std::string& arena_tail)
+{
+    struct frame
+    {
+        const node* cur;
+        const node* end;
+        std::size_t index; ///< the container's node in out
+        std::uint32_t count;
+        bool object;
+    };
+    std::vector<frame> stack;
+    const auto string_node = [&](const node & s)
+    {
+        node r = s;
+        r.extra = 0;
+        r.flags = static_cast<std::uint8_t>(s.flags & node_flags::storage);
+        if (r.flags == node_flags::edited)
+        {
+            r.off = static_cast<std::uint32_t>(arena_size + arena_tail.size());
+            arena_tail.append(d.str(s), s.len);
+            r.flags = node_flags::escaped;
+        }
+        return r;
+    };
+    const auto emit = [&](const node * v)
+    {
+        node r = *v;
+        switch (static_cast<value_t>(v->kind))
+        {
+            case value_t::object:
+            case value_t::array:
+                r.flags = 0;
+                r.extra = 0;
+                r.off = (v->flags & (node_flags::moved | node_flags::is_new)) != 0 ? 0 : v->off;
+                r.len = 0;  // counted below
+                r.next = 0; // set when the container is complete
+                stack.push_back(frame{d.first_child_edited(v), d.child_end_edited(v), out.size(), 0, v->kind == static_cast<std::uint8_t>(value_t::object)});
+                break;
+            case value_t::string:
+                r = string_node(*v);
+                break;
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+                if ((v->flags & node_flags::storage) == node_flags::edited)
+                {
+                    r.off = static_cast<std::uint32_t>(d.size + text_tail.size());
+                    text_tail.append(d.str(*v), number_length(*v));
+                }
+                r.flags = 0;
+                break;
+            case value_t::number_float:
+                if ((v->flags & node_flags::storage) == node_flags::edited)
+                {
+                    const char* const t = d.str(*v);
+                    if (t[0] == 'n' || t[0] == 'i' || (v->len > 1 && t[1] == 'i'))
+                    {
+                        r = node{}; // nan and infinity: null, as dump() writes them
+                        r.kind = static_cast<std::uint8_t>(value_t::null);
+                        break;
+                    }
+                    r.off = static_cast<std::uint32_t>(d.size + text_tail.size());
+                    text_tail.append(t, v->len);
+                    r.extra = 0xFFFFu; // the digit layout is not recorded
+                }
+                r.flags = 0;
+                break;
+            case value_t::boolean:
+                r.flags = static_cast<std::uint8_t>(v->flags & node_flags::is_true);
+                break;
+            case value_t::null:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+                r.flags = 0;
+                break;
+        }
+        out.push_back(r);
+    };
+    emit(d.tape);
+    while (!stack.empty())
+    {
+        frame& top = stack.back();
+        if (top.cur == top.end)
+        {
+            node& c = out[top.index];
+            c.len = top.count;
+            c.next = static_cast<std::uint32_t>(out.size() - top.index);
+            stack.pop_back();
+            continue;
+        }
+        ++top.count;
+        const node* v = nullptr;
+        if (top.object)
+        {
+            out.push_back(string_node(*top.cur));
+            v = document_data::deref(top.cur + 1);
+            top.cur = document_data::after(top.cur + 1);
+        }
+        else
+        {
+            v = document_data::deref(top.cur);
+            top.cur = document_data::after(top.cur);
+        }
+        emit(v); // may grow the stack (top is not used afterwards)
+    }
+}
+
+/// the document as an image
+inline std::vector<std::uint8_t> save_image(const document_data& d)
+{
+#if !NLOHMANN_VIEW_LITTLE_ENDIAN
+    throw_type_error(320, "json_document images need a little-endian target"); // LCOV_EXCL_LINE
+#endif
+    const std::size_t arena_size = d.arena_size;
+    const node* nodes = d.tape;
+    std::size_t count = d.tape_size;
+    std::vector<node> compacted;
+    std::string text_tail;
+    std::string arena_tail;
+    if (d.edits)
+    {
+        compact_nodes(d, arena_size, compacted, text_tail, arena_tail);
+        nodes = compacted.data();
+        count = compacted.size();
+    }
+    const std::size_t text_size = d.size + text_tail.size();
+    const std::size_t total_arena = arena_size + arena_tail.size();
+    if (NLOHMANN_VIEW_UNLIKELY(text_size >= image_limit || total_arena >= image_limit || count >= image_limit))
+    {
+        // LCOV_EXCL_START (4 GiB)
+        throw_out_of_range(416, "images of 4 GiB or more are not supported by json_document");
+        // LCOV_EXCL_STOP
+    }
+    image_header h{};
+    h.magic = {{'N', 'J', 'V', 'I'}};
+    h.version = image_version;
+    h.node_count = count;
+    h.text_size = text_size;
+    h.arena_size = total_arena;
+    std::vector<std::uint8_t> image(sizeof(h) + (count * sizeof(node)) + text_size + 1 + total_arena + 1);
+    std::uint8_t* o = image.data();
+    std::memcpy(o, &h, sizeof(h));
+    o += sizeof(h);
+    std::memcpy(o, nodes, count * sizeof(node));
+    // the hash indexes are rebuilt by load()
+    for (std::size_t i = 0; i < count; ++i)
+    {
+        if (nodes[i].kind == static_cast<std::uint8_t>(value_t::object) && nodes[i].extra != 0)
+        {
+            node n = nodes[i];
+            n.extra = 0;
+            std::memcpy(o + (i * sizeof(node)), &n, sizeof(node));
+        }
+    }
+    o += count * sizeof(node);
+    const auto append = [&o](const char* s, std::size_t n)
+    {
+        if (n != 0)
+        {
+            std::memcpy(o, s, n);
+            o += n;
+        }
+    };
+    append(d.src, d.size);
+    append(text_tail.data(), text_tail.size());
+    *o++ = 0;
+    append(d.base[1], arena_size);
+    append(arena_tail.data(), arena_tail.size());
+    *o = 0;
+    return image;
+}
+
+/// whether a number node matches its token the way the parser records it
+/// (after the bounds check)
+inline bool check_number(const node& n, const unsigned char* text)
+{
+    const std::size_t len = number_length(n);
+    const unsigned char* const s = text + n.off;
+    const unsigned char* const e = s + len;
+    const unsigned char* p = s;
+    const bool negative = *p == '-';
+    p += negative ? 1 : 0;
+    const unsigned char* const int_start = p;
+    if (p == e)
+    {
+        return false;
+    }
+    if (*p == '0')
+    {
+        ++p;
+    }
+    else if (*p >= '1' && *p <= '9')
+    {
+        while (p != e && is_digit(*p))
+        {
+            ++p;
+        }
+    }
+    else
+    {
+        return false;
+    }
+    const auto int_digits = static_cast<std::size_t>(p - int_start);
+    std::size_t frac_digits = 0;
+    bool is_float = false;
+    if (p != e && *p == '.')
+    {
+        const unsigned char* const f0 = ++p;
+        while (p != e && is_digit(*p))
+        {
+            ++p;
+        }
+        if (p == f0)
+        {
+            return false;
+        }
+        frac_digits = static_cast<std::size_t>(p - f0);
+        is_float = true;
+    }
+    std::int64_t exponent = 0;
+    if (p != e && (*p | 0x20u) == 'e')
+    {
+        ++p;
+        const bool exp_negative = p != e && *p == '-';
+        p += (p != e && (*p == '+' || *p == '-')) ? 1 : 0;
+        if (p == e || !is_digit(*p))
+        {
+            return false;
+        }
+        while (p != e && is_digit(*p))
+        {
+            exponent = exponent < 100000 ? (exponent * 10) + (*p - '0') : exponent;
+            ++p;
+        }
+        exponent = exp_negative ? -exponent : exponent;
+        is_float = true;
+    }
+    if (p != e)
+    {
+        return false;
+    }
+    if (n.kind == static_cast<std::uint8_t>(value_t::number_float))
+    {
+        // the digit layout the parser records (or "many", as compaction
+        // writes it), and a finite value
+        const auto layout = static_cast<std::uint16_t>((int_digits < 255 ? int_digits : 255) | ((frac_digits < 255 ? frac_digits : 255) << 8u));
+        if (n.extra != layout && n.extra != 0xFFFFu)
+        {
+            return false;
+        }
+        // parse() rejects floats that overflow; as there, only a number whose
+        // magnitude could reach 1e308 needs the conversion
+        if (static_cast<std::int64_t>(int_digits) + exponent > 300)
+        {
+            const auto v = float_value<double>(reinterpret_cast<const char*>(s), n); // NOLINT(cppcoreguidelines-pro-type-reinterpret-cast)
+            return v <= (std::numeric_limits<double>::max)() && v >= -(std::numeric_limits<double>::max)();
+        }
+        return true;
+    }
+    // integers: the token's value is the stored one; number_integer nodes of
+    // edits can be non-negative (as basic_json keeps the type of a value)
+    const bool integer = n.kind == static_cast<std::uint8_t>(value_t::number_integer);
+    if (is_float || int_digits > 20 || (negative && !integer))
+    {
+        return false;
+    }
+    // (at most 19 digits cannot overflow; 20 digits are compared with 2^64 - 1)
+    if (int_digits == 20 && std::memcmp(int_start, "18446744073709551615", 20) > 0)
+    {
+        return false;
+    }
+    std::uint64_t m = 0;
+    for (const unsigned char* d = int_start; d != int_start + int_digits; ++d)
+    {
+        m = (m * 10) + static_cast<std::uint64_t>(*d - '0');
+    }
+    if (integer && m > (negative ? std::uint64_t{1} << 63u : (std::uint64_t{1} << 63u) - 1))
+    {
+        return false;
+    }
+    return integer_bits(n) == (negative ? 0 - m : m);
+}
+
+/// Check the nodes of a loaded image against its text and decoded strings:
+/// kinds, flags, and `extra`; extents and element counts of arrays and
+/// objects; keys; bounds; string contents (source strings as the parser
+/// leaves them: no quotes, backslashes, or control characters; all strings
+/// valid UTF-8); and number tokens.
+inline bool check_image(const node* nodes, std::size_t count, const unsigned char* text, std::size_t text_size,
+                        const unsigned char* arena, std::size_t arena_size, bool full)
+{
+    struct frame
+    {
+        std::size_t end;
+        std::uint32_t len;
+        std::uint32_t seen;
+        bool object;
+        bool expect_key;
+    };
+    std::vector<frame> stack;
+    const auto check_string = [&](const node & n) -> bool
+    {
+        if ((n.flags & ~node_flags::escaped) != 0 || n.extra != 0)
+        {
+            return false;
+        }
+        const bool decoded = (n.flags & node_flags::escaped) != 0;
+        const unsigned char* const base = decoded ? arena : text;
+        const std::size_t limit = decoded ? arena_size : text_size;
+        if (n.off > limit || n.len > limit - n.off)
+        {
+            return false;
+        }
+        if (!full)
+        {
+            return true;
+        }
+        const unsigned char* const b = base + n.off;
+        return decoded ? valid_utf8_prefix(b, n.len) == n.len : scan_string_run(b, b + n.len) == b + n.len;
+    };
+    // bounds of a number token; the recorded digit layout must lie within it
+    const auto number_in_bounds = [&](const node & n) -> bool
+    {
+        const std::size_t len = number_length(n);
+        if (len == 0 || n.off > text_size || len > text_size - n.off)
+        {
+            return false;
+        }
+        if (n.kind != static_cast<std::uint8_t>(value_t::number_float))
+        {
+            return (n.extra >> 8u) == 0;
+        }
+        // float_value() reads the sign, the integer digits, and the point and
+        // fraction digits the layout records (a layout of more than 19 digits
+        // means the general conversion, which stays within the token)
+        const std::size_t int_digits = n.extra & 0xFFu;
+        const std::size_t frac_digits = n.extra >> 8u;
+        const std::size_t need = (text[n.off] == '-' ? 1u : 0u) + int_digits + (frac_digits != 0 ? frac_digits + 1 : 0);
+        return int_digits + frac_digits > 19 || need <= len;
+    };
+    std::size_t i = 0;
+    for (;;)
+    {
+        // close finished arrays and objects
+        while (!stack.empty() && i == stack.back().end)
+        {
+            const frame f = stack.back();
+            if (f.seen != f.len || (f.object && !f.expect_key))
+            {
+                return false;
+            }
+            stack.pop_back();
+            if (!stack.empty())
+            {
+                ++stack.back().seen;
+                stack.back().expect_key = true;
+            }
+        }
+        if (i == count)
+        {
+            return stack.empty();
+        }
+        if (i != 0 && stack.empty())
+        {
+            return false; // nodes after the root
+        }
+        const node& n = nodes[i];
+        if (!stack.empty() && stack.back().object && stack.back().expect_key)
+        {
+            if (n.kind != static_cast<std::uint8_t>(value_t::string) || !check_string(n))
+            {
+                return false;
+            }
+            stack.back().expect_key = false;
+            ++i;
+            continue;
+        }
+        bool complete = true;
+        switch (static_cast<value_t>(n.kind))
+        {
+            case value_t::null:
+                // (the offset of a literal is read to size the output of dump())
+                if (n.flags != 0 || n.extra != 0 || n.off > text_size)
+                {
+                    return false;
+                }
+                break;
+            case value_t::boolean:
+                if ((n.flags & ~node_flags::is_true) != 0 || n.extra != 0 || n.off > text_size)
+                {
+                    return false;
+                }
+                break;
+            case value_t::string:
+                if (!check_string(n))
+                {
+                    return false;
+                }
+                break;
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+                if (n.flags != 0 || !number_in_bounds(n) || (full && !check_number(n, text)))
+                {
+                    return false;
+                }
+                break;
+            case value_t::array:
+            case value_t::object:
+            {
+                const std::size_t limit = stack.empty() ? count : stack.back().end;
+                if (n.flags != 0 || n.extra != 0 || n.next == 0 || n.next > limit - i || n.off > text_size)
+                {
+                    return false;
+                }
+                stack.push_back(frame{i + n.next, n.len, 0, n.kind == static_cast<std::uint8_t>(value_t::object), true});
+                complete = false;
+                break;
+            }
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+                return false;
+        }
+        ++i;
+        if (complete && !stack.empty())
+        {
+            ++stack.back().seen;
+            stack.back().expect_key = true;
+        }
+    }
+}
+
+[[noreturn]] NLOHMANN_VIEW_NOINLINE inline void throw_invalid_image(const char* what)
+{
+    throw_parse_error(116, concat("invalid json_document image: ", what));
+}
+
+/// Read an image into d. The text and the decoded strings stay in the image;
+/// the nodes are copied (so that they are aligned, and edits can change them).
+inline void load_image(document_data& d, const std::uint8_t* image, std::size_t size, image_check check)
+{
+#if !NLOHMANN_VIEW_LITTLE_ENDIAN
+    throw_type_error(320, "json_document images need a little-endian target"); // LCOV_EXCL_LINE
+#endif
+    if (image == nullptr || size < sizeof(image_header))
+    {
+        throw_invalid_image("too short");
+    }
+    image_header h{};
+    std::memcpy(&h, image, sizeof(h));
+    // (the reserved fields are for later versions)
+    if (std::memcmp(h.magic.data(), "NJVI", 4) != 0 || h.version != image_version
+            || (h.reserved[0] | h.reserved[1] | h.reserved[2] | h.reserved[3]) != 0)
+    {
+        throw_invalid_image("unknown format");
+    }
+    const std::size_t room = size - sizeof(h);
+    if (h.node_count == 0 || h.node_count > room / sizeof(node) || h.node_count >= image_limit || h.text_size >= image_limit || h.arena_size >= image_limit)
+    {
+        throw_invalid_image("sizes out of range");
+    }
+    const auto count = static_cast<std::size_t>(h.node_count);
+    const auto text_size = static_cast<std::size_t>(h.text_size);
+    const auto arena_size = static_cast<std::size_t>(h.arena_size);
+    const std::size_t text_at = sizeof(h) + (count * sizeof(node));
+    // the text, a NUL, the decoded strings, a NUL, and nothing after them
+    if (size - text_at < 2 || text_size > size - text_at - 2 || arena_size != size - text_at - text_size - 2
+            || image[text_at + text_size] != 0 || image[size - 1] != 0)
+    {
+        throw_invalid_image("sizes out of range");
+    }
+
+    d.discarded = true;
+    d.edits.reset();
+    d.base[2] = nullptr;
+    d.owned.clear();
+    if (d.owned_image.empty() || image != d.owned_image.data())
+    {
+        d.owned_image.clear();
+    }
+    d.arena.clear();
+    d.indexes.clear();
+    d.index_slots.clear();
+    d.large_objects.clear();
+    d.tape_size = 0;
+    d.reserve(count);
+    std::memcpy(d.tape, image + sizeof(h), count * sizeof(node));
+    d.tape_size = count;
+    const std::uint8_t* const text = image + text_at;
+    const std::uint8_t* const arena = text + text_size + 1;
+    d.src = reinterpret_cast<const char*>(text); // NOLINT(cppcoreguidelines-pro-type-reinterpret-cast)
+    d.size = text_size;
+    d.base[0] = d.src;
+    d.base[1] = reinterpret_cast<const char*>(arena); // NOLINT(cppcoreguidelines-pro-type-reinterpret-cast)
+    d.arena_size = arena_size;
+    if (check != image_check::none && !check_image(d.tape, count, text, text_size, arena, arena_size, check == image_check::full))
+    {
+        throw_invalid_image("the check failed");
+    }
+    // the hash indexes of large objects, as after parsing
+    for (std::size_t i = 0; i < count; ++i)
+    {
+        node& n = d.tape[i];
+        if (n.kind == static_cast<std::uint8_t>(value_t::object))
+        {
+            n.extra = 0;
+            if (n.len >= document_data::index_min_members)
+            {
+                d.large_objects.push_back(static_cast<std::uint32_t>(i));
+            }
+        }
+    }
+    build_object_indexes(d);
+    d.discarded = false;
+}
+
+}  // namespace view
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
@@ -29,43 +29,76 @@ namespace detail
 namespace view
 {

+/*!
+@brief locate the decimal point and the end of the mantissa of a float token
+
+Also checks that the token is a JSON number. Tokens of the parser and of edits
+always are; an image loaded with image_check::bounds can hold any bytes, which
+must not reach the conversion (it expects a well-formed token).
+*/
+inline bool float_token_layout(const char* first, const char* last, std::size_t& dot, std::size_t& mantissa_end) noexcept
+{
+    const auto digit = [last](const char* q)
+    {
+        return q != last && is_digit(static_cast<unsigned char>(*q));
+    };
+    const char* p = first;
+    p += (p != last && *p == '-') ? 1 : 0;
+    if (!digit(p) || (*p == '0' && digit(p + 1)))
+    {
+        return false;
+    }
+    while (digit(p))
+    {
+        ++p;
+    }
+    dot = std::string::npos;
+    if (p != last && *p == '.')
+    {
+        dot = static_cast<std::size_t>(p - first);
+        if (!digit(++p))
+        {
+            return false;
+        }
+        while (digit(p))
+        {
+            ++p;
+        }
+    }
+    mantissa_end = static_cast<std::size_t>(p - first);
+    if (p != last && (*p == 'e' || *p == 'E'))
+    {
+        ++p;
+        p += (p != last && (*p == '+' || *p == '-')) ? 1 : 0;
+        if (!digit(p))
+        {
+            return false;
+        }
+        while (digit(p))
+        {
+            ++p;
+        }
+    }
+    return p == last;
+}
+
 /*!
@brief the value of the float token of a node, as parse() converts it

 Uses the lexer's conversion (detail::convert_float_fast, then the locale-aware
-strtod fallback), so that the values are bit-identical to parse(). The digit
-layout recorded while parsing locates the decimal point and the exponent
-without scanning the token.
+strtod fallback), so that the values are bit-identical to parse(). A token
+that is not a JSON number (only in a damaged image loaded with
+image_check::bounds) yields 0.
 */
 template<typename FloatType>
 NLOHMANN_VIEW_NOINLINE FloatType float_value(const char* first, const node& n)
 {
    const char* const last = first + n.len;
-    const std::size_t neg = first[0] == '-' ? 1 : 0;
-    const std::size_t int_digits = n.extra & 0xFFu;
-    const std::size_t frac_digits = n.extra >> 8u;
-    std::size_t dot = std::string::npos;
-    std::size_t mantissa_end = n.len;
-    if (int_digits != 255 && frac_digits != 255)
+    std::size_t dot = 0;
+    std::size_t mantissa_end = 0;
+    if (NLOHMANN_VIEW_UNLIKELY(!float_token_layout(first, last, dot, mantissa_end)))
    {
-        dot = frac_digits != 0 ? neg + int_digits : std::string::npos;
-        mantissa_end = neg + int_digits + (frac_digits != 0 ? 1 + frac_digits : 0);
-    }
-    else
-    {
-        // more digits than the layout records: locate them
-        for (std::size_t i = 0; i < n.len; ++i)
-        {
-            if (first[i] == '.')
-            {
-                dot = i;
-            }
-            else if (first[i] == 'e' || first[i] == 'E')
-            {
-                mantissa_end = i;
-                break;
-            }
-        }
+        return FloatType{};
    }
    FloatType v{};
    if (!convert_float_fast(first, last, dot, mantissa_end, v))
@@ -104,16 +137,20 @@ NLOHMANN_VIEW_ALWAYS_INLINE double layout_double(const unsigned char* p, const u
    }
    if (p != e)
    {
-        // [eE][+-]digits; huge exponents saturate (the parser rejected overflow)
+        // [eE][+-]digits; huge exponents saturate (the parser rejected
+        // overflow). The token is not read beyond e, and the digits are taken
+        // as unsigned, so that a token that is not well-formed (a damaged
+        // image loaded with image_check::bounds) yields a wrong value, but no
+        // overflow.
        ++p;
-        const bool exp_negative = *p == '-';
-        p += (*p == '-' || *p == '+') ? 1 : 0;
+        const bool exp_negative = p != e && *p == '-';
+        p += (p != e && (*p == '-' || *p == '+')) ? 1 : 0;
        std::int64_t exp_value = 0;
        for (; p != e; ++p)
        {
            if (exp_value < 0x10000000)
            {
-                exp_value = (exp_value * 10) + (*p - '0');
+                exp_value = (exp_value * 10) + static_cast<unsigned char>(*p - '0');
            }
        }
        q += exp_negative ? -exp_value : exp_value;
@@ -75,6 +75,23 @@ class output_buffer
        m_pos += n;
    }

+    /// the write position and the end of the writable space, for a writer
+    /// that keeps the position in a local variable (set_cursor() hands it back)
+    char* cursor() const noexcept
+    {
+        return m_pos;
+    }
+
+    char* limit() const noexcept
+    {
+        return m_end;
+    }
+
+    void set_cursor(char* p) noexcept
+    {
+        m_pos = p;
+    }
+
  private:
    static StringType& sized(StringType& out, std::size_t estimate)
    {
@@ -95,6 +112,105 @@ class output_buffer
    char* m_end;
 };

+/// The length of the run at s that dump() writes unchanged without
+/// ensure_ascii: all bytes but quotes, backslashes, and control characters.
+/// Unlike detail::string_bulk_run(), non-ASCII bytes are not validated: the
+/// strings of a document are valid UTF-8 (a damaged image loaded with
+/// image_check::bounds can have others, which are then written unchanged).
+inline std::size_t plain_output_run(const unsigned char* s, std::size_t n) noexcept
+{
+    constexpr std::uint64_t ones = 0x0101010101010101ull;
+    constexpr std::uint64_t high = 0x8080808080808080ull;
+    std::size_t i = 0;
+    for (; i + 8 <= n; i += 8)
+    {
+        const std::uint64_t v = read_eight_bytes(s + i);
+        const std::uint64_t q = v ^ 0x2222222222222222ull; // '"'
+        const std::uint64_t b = v ^ 0x5C5C5C5C5C5C5C5Cull; // '\\'
+        const std::uint64_t stop = (((q - ones) & ~q) | ((b - ones) & ~b) | ((v - 0x2020202020202020ull) & ~v)) & high;
+        if (stop != 0)
+        {
+            // the lowest flagged byte is the first stop: borrows only flag bytes above a true one
+            return i + (static_cast<std::size_t>(count_trailing_zeros(stop)) / 8);
+        }
+    }
+    for (; i < n; ++i)
+    {
+        if (s[i] == '"' || s[i] == '\\' || s[i] < 0x20)
+        {
+            return i;
+        }
+    }
+    return n;
+}
+
+/// A stack that starts in a buffer of the caller (a local array) and moves to
+/// the heap (a vector of the caller) only when that is full, so that dumps of
+/// shallow documents need no allocation. The top is a pointer, as in
+/// std::vector. The address of the stack never escapes (the growth gets the
+/// vector and returns the new storage), so its pointers stay in registers.
+template<typename T>
+class small_stack
+{
+  public:
+    small_stack(T* buffer, std::size_t capacity, std::vector<T>& heap) noexcept
+        : m_begin(buffer), m_top(buffer), m_end(buffer + capacity), m_heap(&heap)
+    {}
+    small_stack(const small_stack&) = delete;
+    small_stack(small_stack&&) = delete;
+    small_stack& operator=(const small_stack&) = delete;
+    small_stack& operator=(small_stack&&) = delete;
+    ~small_stack() = default;
+
+    NLOHMANN_VIEW_ALWAYS_INLINE void push_back(const T& x)
+    {
+        if (NLOHMANN_VIEW_UNLIKELY(m_top == m_end))
+        {
+            const std::size_t used = size();
+            const std::size_t capacity = 2 * static_cast<std::size_t>(m_end - m_begin);
+            m_begin = grow(*m_heap, m_begin, used, capacity);
+            m_top = m_begin + used;
+            m_end = m_begin + capacity;
+        }
+        *m_top++ = x;
+    }
+
+    NLOHMANN_VIEW_ALWAYS_INLINE T& back() noexcept
+    {
+        return m_top[-1];
+    }
+
+    NLOHMANN_VIEW_ALWAYS_INLINE void pop_back() noexcept
+    {
+        --m_top;
+    }
+
+    NLOHMANN_VIEW_ALWAYS_INLINE bool empty() const noexcept
+    {
+        return m_top == m_begin;
+    }
+
+    NLOHMANN_VIEW_ALWAYS_INLINE std::size_t size() const noexcept
+    {
+        return static_cast<std::size_t>(m_top - m_begin);
+    }
+
+  private:
+    /// the used entries moved to heap storage of the given capacity
+    NLOHMANN_VIEW_NOINLINE static T* grow(std::vector<T>& heap, const T* begin, std::size_t used, std::size_t capacity)
+    {
+        std::vector<T> bigger(capacity);
+        std::copy(begin, begin + used, bigger.begin());
+        heap.swap(bigger);
+        return heap.data();
+    }
+
+    T* m_begin;
+    T* m_top;
+    T* m_end;
+    std::vector<T>* m_heap;
+};
+
 /// how the view's dump() writes a value
 struct dump_style
 {
@@ -129,6 +245,18 @@ class view_serializer

    void dump(const node* root)
    {
+        if (!m_style.pretty && !m_style.ensure_ascii)
+        {
+            if (m_style.source_numbers)
+            {
+                dump_compact<true>(root);
+            }
+            else
+            {
+                dump_compact<false>(root);
+            }
+            return;
+        }
        struct frame
        {
            const node* pos; ///< next element, or key of the next member
@@ -136,7 +264,9 @@ class view_serializer
            bool object;
            bool first;  ///< nothing written yet
        };
-        std::vector<frame> stack;
+        std::array<frame, 32> buffer; // NOLINT(cppcoreguidelines-pro-type-member-init,hicpp-member-init): written before read
+        std::vector<frame> heap;
+        small_stack<frame> stack(buffer.data(), buffer.size(), heap);
        const node* n = root;
        for (;;)
        {
@@ -207,6 +337,284 @@ class view_serializer
    }

  private:
+    /*!
+    @brief the compact output without ensure_ascii (the default dump())
+
+    The same walk as dump(), with the write position in a local variable
+    (stores through char pointers would otherwise force a reload of the
+    buffer's members after each one), and with strings and number tokens of
+    the source copied by fixed-size moves of 32 bytes where the source has
+    that many bytes left, instead of a library call per token. The buffer
+    keeps 64 bytes of slack for the overshoot.
+    */
+    /// a string that is not a plain string of the source (decoded, or written
+    /// by an edit), without ensure_ascii: runs without characters to escape
+    /// are copied
+    NLOHMANN_VIEW_NOINLINE void write_decoded(const node& n)
+    {
+        const auto* const s = reinterpret_cast<const unsigned char*>(m_doc.str(n)); // NOLINT(cppcoreguidelines-pro-type-reinterpret-cast)
+        m_out.put('"');
+        for (std::size_t i = 0; i < n.len;)
+        {
+            const std::size_t run = plain_output_run(s + i, n.len - i);
+            if (run != 0)
+            {
+                m_out.put(reinterpret_cast<const char*>(s + i), run); // NOLINT(cppcoreguidelines-pro-type-reinterpret-cast)
+                i += run;
+                continue;
+            }
+            write_codepoint<false>(s[i], s + i, 1); // a quote, a backslash, or a control character
+            ++i;
+        }
+        m_out.put('"');
+    }
+
+    /// the copies of dump_compact() that are not fixed-size moves (long
+    /// strings, or near the end of the source); out of line, so that the
+    /// compiler does not merge the fixed-size moves into this call
+    NLOHMANN_VIEW_NOINLINE static void copy_long(char* to, const char* from, std::size_t n) noexcept
+    {
+        std::memcpy(to, from, n);
+    }
+
+    template<bool SourceNumbers>
+    void dump_compact(const node* root)
+    {
+        struct frame
+        {
+            const node* pos; ///< (editable documents) next element, or key of the next member
+            const node* end;
+            bool object;
+        };
+        std::array<frame, 32> buffer; // NOLINT(cppcoreguidelines-pro-type-member-init,hicpp-member-init): written before read
+        std::vector<frame> heap;
+        small_stack<frame> stack(buffer.data(), buffer.size(), heap);
+        const char* const src = m_doc.src;
+        const char* const src_end = src + m_doc.size;
+        char* w = m_out.cursor();
+        char* lim = m_out.limit();
+        // room for n bytes and the slack
+        const auto room = [&](std::size_t n)
+        {
+            if (NLOHMANN_VIEW_UNLIKELY(static_cast<std::size_t>(lim - w) < n + 64))
+            {
+                m_out.set_cursor(w);
+                m_out.reserve(n + 64);
+                w = m_out.cursor();
+                lim = m_out.limit();
+            }
+        };
+        // copy n bytes of the source (after room(n))
+        const auto copy = [&](const char* from, std::size_t n)
+        {
+            if (n <= 32 && src_end - from >= 32)
+            {
+                std::memcpy(w, from, 32);
+            }
+            else if (n <= 256 && src_end - from >= static_cast<std::ptrdiff_t>(n) + 32)
+            {
+                for (std::size_t i = 0; i < n; i += 32)
+                {
+                    std::memcpy(w + i, from + i, 32);
+                }
+            }
+            else
+            {
+                copy_long(w, from, n);
+            }
+            w += n;
+        };
+        // a literal of n bytes (after room(n))
+        const auto literal = [&](const char* text, std::size_t n)
+        {
+            std::memcpy(w, text, n);
+            w += n;
+        };
+        // a string that is not a plain string of the source (out of line, so
+        // that the cursor stays in a register here)
+        const auto escaped = [&](const node & n)
+        {
+            m_out.set_cursor(w);
+            write_decoded(n);
+            w = m_out.cursor();
+            lim = m_out.limit();
+        };
+
+        // Read-only documents: the elements of a container follow it in the
+        // node array, so the walk goes through the array in order, and a
+        // frame only needs the end of its container. Editable documents: the
+        // elements of a moved container live elsewhere, so a frame keeps the
+        // position of the next element (see navigation).
+        // The innermost open container is kept in registers (cur; end ==
+        // nullptr: none), the stack holds the ones around it.
+        frame cur{nullptr, nullptr, false};
+        const node* n = root;
+        for (;;)
+        {
+            // write the value at n (read-only documents: and advance n)
+            bool opened = false;
+            switch (static_cast<value_t>(n->kind))
+            {
+                case value_t::string:
+                    if ((n->flags & node_flags::storage) == 0)
+                    {
+                        room(n->len + 2);
+                        *w++ = '"';
+                        copy(src + n->off, n->len);
+                        *w++ = '"';
+                    }
+                    else
+                    {
+                        escaped(*n);
+                    }
+                    break;
+                case value_t::number_integer:
+                case value_t::number_unsigned:
+                {
+                    const std::uint32_t len = number_length(*n);
+                    room(len);
+                    if (Editable && (n->flags & node_flags::storage) != 0)
+                    {
+                        copy_long(w, m_doc.str(*n), len); // a canonical token written by an edit
+                        w += len;
+                        break;
+                    }
+                    const char* const token = src + n->off;
+                    if (!SourceNumbers && NLOHMANN_VIEW_UNLIKELY(len == 2 && token[0] == '-' && token[1] == '0'))
+                    {
+                        *w++ = '0'; // parse() reads -0 as the integer 0
+                    }
+                    else
+                    {
+                        copy(token, len);
+                    }
+                    break;
+                }
+                case value_t::number_float:
+                    if (SourceNumbers && (n->flags & node_flags::storage) != node_flags::edited)
+                    {
+                        room(n->len);
+                        copy(src + n->off, n->len);
+                    }
+                    else
+                    {
+                        m_out.set_cursor(w);
+                        write_float(float_value<number_float_t>(m_doc, *n));
+                        w = m_out.cursor();
+                        lim = m_out.limit();
+                    }
+                    break;
+                case value_t::boolean:
+                    room(8);
+                    if ((n->flags & node_flags::is_true) != 0)
+                    {
+                        literal("true", 4);
+                    }
+                    else
+                    {
+                        literal("false", 5);
+                    }
+                    break;
+                case value_t::object:
+                case value_t::array:
+                {
+                    const bool object = n->kind == static_cast<std::uint8_t>(value_t::object);
+                    room(8);
+                    if (n->len == 0)
+                    {
+                        literal(object ? "{}" : "[]", 2);
+                    }
+                    else
+                    {
+                        *w++ = object ? '{' : '[';
+                        stack.push_back(cur);
+                        if (Editable)
+                        {
+                            cur = frame{nav::first(m_doc, n), nav::end(m_doc, n), object};
+                        }
+                        else
+                        {
+                            cur = frame{nullptr, n + n->next, object};
+                        }
+                        opened = true;
+                    }
+                    break;
+                }
+                case value_t::null:
+                    room(8);
+                    literal("null", 4);
+                    break;
+                case value_t::binary:    // LCOV_EXCL_LINE (not in a document)
+                case value_t::discarded: // LCOV_EXCL_LINE
+                default:                 // LCOV_EXCL_LINE
+                    break;               // LCOV_EXCL_LINE
+            }
+            if (!Editable)
+            {
+                ++n; // the next node: the first element of an opened container, or the node after a scalar
+            }
+
+            // go to the next value: close finished containers, then separate
+            // (a container just opened has an element)
+            if (!opened)
+            {
+                for (;;)
+                {
+                    if (cur.end == nullptr)
+                    {
+                        m_out.set_cursor(w);
+                        m_out.finish();
+                        return;
+                    }
+                    if ((Editable ? cur.pos : n) != cur.end)
+                    {
+                        break;
+                    }
+                    room(1);
+                    *w++ = cur.object ? '}' : ']';
+                    cur = stack.back();
+                    stack.pop_back();
+                }
+                room(1);
+                *w++ = ',';
+            }
+            const node* const at = Editable ? cur.pos : n;
+            if (cur.object)
+            {
+                const node& key = *at;
+                if ((key.flags & node_flags::storage) == 0)
+                {
+                    room(key.len + 3);
+                    *w++ = '"';
+                    copy(src + key.off, key.len);
+                    w[0] = '"';
+                    w[1] = ':';
+                    w += 2;
+                }
+                else
+                {
+                    escaped(key);
+                    room(1);
+                    *w++ = ':';
+                }
+                if (Editable)
+                {
+                    n = nav::value(at + 1);
+                    cur.pos = document_data::after(at + 1);
+                }
+                else
+                {
+                    ++n;
+                }
+            }
+            else if (Editable)
+            {
+                n = nav::value(at);
+                cur.pos = document_data::after(at);
+            }
+        }
+    }
+
    void newline(std::size_t level)
    {
        if (m_style.pretty)
@@ -317,7 +725,9 @@ class view_serializer
        m_out.put('"');
    }

-    /// as serializer::dump_escaped() for valid UTF-8 (the view has no other)
+    /// as serializer::dump_escaped(); strings of a document are valid UTF-8,
+    /// except in a damaged image loaded with image_check::bounds, for which
+    /// this throws what basic_json::dump() throws for the string
    template<bool EnsureAscii>
    void write_escaped(const unsigned char* s, std::size_t n)
    {
@@ -341,12 +751,13 @@ class view_serializer
            }
            std::uint32_t codepoint = s[i];
            std::size_t len = 1;
-            if (codepoint >= 0xC0)
+            if (codepoint >= 0x80)
            {
-                len = 2;
-                if (codepoint >= 0xE0)
+                len = validate_one_utf8(s + i, n - i);
+                if (NLOHMANN_VIEW_UNLIKELY(len == 0))
                {
-                    len = codepoint >= 0xF0 ? 4 : 3;
+                    invalid_utf8(s, n);
+                    return;
                }
                codepoint &= 0xFFu >> (len + 1);
                for (std::size_t k = 1; k < len; ++k)
@@ -359,6 +770,13 @@ class view_serializer
        }
    }

+    /// throw what basic_json::dump() throws for a string that is not valid UTF-8
+    NLOHMANN_VIEW_NOINLINE static void invalid_utf8(const unsigned char* s, std::size_t n)
+    {
+        const string_t dumped = BasicJsonType(string_t(reinterpret_cast<const char*>(s), n)).dump(); // NOLINT(cppcoreguidelines-pro-type-reinterpret-cast)
+        static_cast<void>(dumped);
+    }
+
    template<bool EnsureAscii>
    void write_codepoint(std::uint32_t codepoint, const unsigned char* bytes, std::size_t len)
    {
@@ -25,7 +25,7 @@
 #define INCLUDE_NLOHMANN_JSON_VIEW_HPP_

 #include <cstddef> // size_t
-#include <cstdint> // uint32_t
+#include <cstdint> // uint8_t, uint32_t
 #include <cstring> // memcpy, strlen
 #include <iterator> // distance, input_iterator_tag, iterator_traits
 #include <map> // map
@@ -53,6 +53,7 @@
 #include <nlohmann/detail/view/edit.hpp>
 #include <nlohmann/detail/view/edit_storage.hpp>
 #include <nlohmann/detail/view/errors.hpp>
+#include <nlohmann/detail/view/image.hpp>
 #include <nlohmann/detail/view/input.hpp>
 #include <nlohmann/detail/view/iterator.hpp>
 #include <nlohmann/detail/view/lookup.hpp>
@@ -591,8 +592,10 @@ class basic_json_view
        style.indent_char = indent_char;
        style.ensure_ascii = ensure_ascii;
        style.source_numbers = numbers == number_format::source;
-        // the compact text is about as long as the source text of the value
-        const std::size_t estimate = source_extent() + (style.pretty ? source_extent() / 2 : 0) + 64;
+        // the compact text is about as long as the source text of the value;
+        // the compact writer keeps 64 bytes of slack, so that it does not grow
+        // the buffer just before the end
+        const std::size_t estimate = source_extent() + (style.pretty ? source_extent() / 2 : 0) + 160;
        detail::view::view_serializer<BasicJsonType, Editable>(*m_doc, out, estimate, style).dump(m_node);
        return out;
    }
@@ -933,7 +936,7 @@ class basic_json_document
    /// whether the document holds its own copy of the text
    bool owns_source() const noexcept
    {
-        return m_data && !m_data->owned.empty() && m_data->src == m_data->owned.data();
+        return m_data && ((!m_data->owned.empty() && m_data->src == m_data->owned.data()) || !m_data->owned_image.empty());
    }

    /// number of index nodes (values plus object keys)
@@ -942,7 +945,7 @@ class basic_json_document
        return m_data ? m_data->tape_size : 0;
    }

-    /// bytes held by the document (index, decoded strings, owned text)
+    /// bytes held by the document (index, decoded strings, owned text or image)
    std::size_t memory_usage() const noexcept
    {
        if (!m_data)
@@ -951,7 +954,7 @@ class basic_json_document
        }
        return sizeof(document_data) + (m_data->inline_cap * sizeof(detail::view::node))
               + (m_data->tape != m_data->inline_tape ? m_data->tape_cap * sizeof(detail::view::node) : 0)
-               + m_data->arena.capacity() + m_data->owned.capacity()
+               + m_data->arena.capacity() + m_data->owned.capacity() + m_data->owned_image.capacity()
               + (m_data->indexes.capacity() * sizeof(document_data::object_index)) + (m_data->index_slots.capacity() * sizeof(std::uint32_t))
               + (m_data->large_objects.capacity() * sizeof(std::uint32_t))
               + (m_data->edits != nullptr ? m_data->edits->bytes : 0);
@@ -971,8 +974,10 @@ class basic_json_document

        // allocate everything first, so that an exception leaves the document
        // unchanged
+        // (the decoded strings of a loaded image stay in the image)
+        const bool arena_in_use = d.base[1] == d.arena.data();
        const bool shrink_arena = d.arena.capacity() > d.arena.size();
-        std::string arena(shrink_arena ? d.arena : std::string());
+        std::string arena(shrink_arena && arena_in_use ? d.arena : std::string());
        // (edits link to the nodes of the index, which then stays in place)
        const bool shrink_tape = d.tape != d.inline_tape && d.tape_size != d.tape_cap && d.edits == nullptr;
        const bool into_header = d.tape_size <= d.inline_cap;
@@ -988,10 +993,62 @@ class basic_json_document
        if (shrink_arena)
        {
            d.arena.swap(arena);
-            d.base[1] = d.arena.data();
+            if (arena_in_use)
+            {
+                d.base[1] = d.arena.data();
+            }
        }
    }

+    ////////////
+    // images //
+    ////////////
+
+    /// how load() checks an image (full, bounds, or none)
+    using image_check = detail::view::image_check;
+
+    /// The document as an image that load() reads without parsing: the node
+    /// index, the text, and the decoded strings. An edited document is
+    /// written in its current state (floats that are not finite become null,
+    /// as in dump()).
+    std::vector<std::uint8_t> save() const
+    {
+        if (NLOHMANN_VIEW_UNLIKELY(!m_data || m_data->discarded))
+        {
+            detail::view::throw_type_error(320, "cannot save a discarded json_document");
+        }
+        return detail::view::save_image(*m_data);
+    }
+
+    /// Read an image written by save(). The image is borrowed: it must stay
+    /// alive and unchanged while the document is used.
+    NLOHMANN_VIEW_NODISCARD
+    static basic_json_document load(const std::uint8_t* image, std::size_t size, const image_check check = image_check::full)
+    {
+        basic_json_document d;
+        d.ensure_data(nullptr, 0);
+        detail::view::load_image(*d.m_data, image, size, check);
+        return d;
+    }
+
+    /// read an image (borrowed)
+    NLOHMANN_VIEW_NODISCARD
+    static basic_json_document load(const std::vector<std::uint8_t>& image, const image_check check = image_check::full)
+    {
+        return load(image.data(), image.size(), check);
+    }
+
+    /// read an image and keep it (no copy)
+    NLOHMANN_VIEW_NODISCARD
+    static basic_json_document load(std::vector<std::uint8_t>&& image, const image_check check = image_check::full)
+    {
+        basic_json_document d;
+        d.ensure_data(nullptr, 0);
+        d.m_data->owned_image = std::move(image);
+        detail::view::load_image(*d.m_data, d.m_data->owned_image.data(), d.m_data->owned_image.size(), check);
+        return d;
+    }
+
    ///////////
    // edits //
    ///////////
@@ -1175,6 +1232,7 @@ class basic_json_document
        {
            d.owned.clear();
        }
+        d.owned_image.clear();
        d.src = src;
        d.size = size;
        d.tape_size = 0;
@@ -1199,6 +1257,7 @@ class basic_json_document
        {
            d.base[0] = d.src;
            d.base[1] = d.arena.data();
+            d.arena_size = d.arena.size();
            detail::view::build_object_indexes(d);
            d.discarded = false;
            return;
@@ -23918,11 +23918,240 @@ NLOHMANN_JSON_NAMESPACE_END

 #include <array> // array
 #include <cmath>   // signbit, isfinite
+#include <cstddef> // size_t
 #include <cstdint> // intN_t, uintN_t
 #include <cstring> // memcpy, memmove
 #include <limits> // numeric_limits
 #include <type_traits> // conditional

+#ifdef _MSC_VER
+    #include <cstdlib> // _byteswap_uint64
+#endif
+
+// #include <nlohmann/detail/conversions/zmij.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.12.0
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2025 Victor Zverovich <https://github.com/vitaut/zmij>
+// SPDX-FileCopyrightText: 2013-2026 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <array> // array
+#include <cstddef> // size_t
+#include <cstdint> // uint32_t, uint64_t
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+// #include <nlohmann/detail/bit_ops.hpp>
+
+// #include <nlohmann/detail/input/pow5_table.hpp>
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+/*!
+@brief the shortest decimal representation of a double
+
+A C++11 port of the conversion of Zmij by Victor Zverovich
+(https://github.com/vitaut/zmij, MIT license): the shortest decimal in the
+rounding interval of a double, the closest one if there are several. Zmij
+credits Xiang JunBo (producing the shorter candidate without a division) and
+Dougall Johnson (the compressed powers of ten). The powers of ten are taken
+from the table for number parsing (pow5_table.hpp) where it holds them, and
+computed from the compressed tables of Zmij beyond it.
+*/
+namespace zmij
+{
+
+/// significand * 10^exponent
+struct decimal
+{
+    std::uint64_t significand;
+    int exponent;
+};
+
+/// the compressed powers of ten of Zmij
+inline const std::array<std::uint64_t, 28>& pow10_minor() noexcept
+{
+    static const std::array<std::uint64_t, 28> table =
+    {
+        {
+            0x8000000000000000u, 0xa000000000000000u, 0xc800000000000000u, 0xfa00000000000000u, 0x9c40000000000000u,
+            0xc350000000000000u, 0xf424000000000000u, 0x9896800000000000u, 0xbebc200000000000u, 0xee6b280000000000u,
+            0x9502f90000000000u, 0xba43b74000000000u, 0xe8d4a51000000000u, 0x9184e72a00000000u, 0xb5e620f480000000u,
+            0xe35fa931a0000000u, 0x8e1bc9bf04000000u, 0xb1a2bc2ec5000000u, 0xde0b6b3a76400000u, 0x8ac7230489e80000u,
+            0xad78ebc5ac620000u, 0xd8d726b7177a8000u, 0x878678326eac9000u, 0xa968163f0a57b400u, 0xd3c21bcecceda100u,
+            0x84595161401484a0u, 0xa56fa5b99019a5c8u, 0xcecb8f27f4200f3au
+        }
+    };
+    return table;
+}
+
+/// (high, low) pairs
+inline const std::array<std::uint64_t, 50>& pow10_major() noexcept
+{
+    static const std::array<std::uint64_t, 50> table =
+    {
+        {
+            0xaddcb9e83c6b1793u, 0xdf4abe242a1bbf3eu, 0xaf8e5410288e1b6fu, 0x07ecf0ae5ee44ddau, 0xb1442798f49ffb4au, 0x99cd11cfdf41779du,
+            0xb2fe3f0b8599ef07u, 0x861fa7e6dcb4aa15u, 0xb4bca50b065abe63u, 0x0fed077a756b53aau, 0xb67f6455292cbf08u, 0x1a3bc84c17b1d543u,
+            0xb84687c269ef3bfbu, 0x3d5d514f40eea742u, 0xba121a4650e4ddebu, 0x92f34d62616ce413u, 0xbbe226efb628afeau, 0x890489f70a55368cu,
+            0xbdb6b8e905cb600fu, 0x5400e987bbc1c921u, 0xbf8fdb78849a5f96u, 0xde98520472bdd034u, 0xc16d9a0095928a27u, 0x75b7053c0f178294u,
+            0xc350000000000000u, 0x0000000000000000u, 0xc5371912364ce305u, 0x6c28000000000000u, 0xc722f0ef9d80aad6u, 0x424d3ad2b7b97ef6u,
+            0xc913936dd571c84cu, 0x03bc3a19cd1e38eau, 0xcb090c8001ab551cu, 0x5cadf5bfd3072cc6u, 0xcd036837130890a1u, 0x36dba887c37a8c10u,
+            0xcf02b2c21207ef2eu, 0x94f967e45e03f4bcu, 0xd106f86e69d785c7u, 0xe13336d701beba52u, 0xd31045a8341ca07cu, 0x1ede48111209a051u,
+            0xd51ea6fa85785631u, 0x552a74227f3ea566u, 0xd732290fbacaf133u, 0xa97c177947ad4096u, 0xd94ad8b1c7380874u, 0x18375281ae7822bdu,
+            0xdb68c2ca82ed2a05u, 0xa67398db9f6820e1u
+        }
+    };
+    return table;
+}
+
+/// one bit per power: whether the computed value is one unit too large
+inline const std::array<std::uint32_t, 21>& pow10_fixups() noexcept
+{
+    static const std::array<std::uint32_t, 21> table =
+    {
+        {
+            0x8d8fc810u, 0x06100293u, 0x19000000u, 0x00100000u, 0x00000908u, 0x00000000u, 0x04e00300u, 0x3807e0b2u, 0x3d83d793u, 0x0006f5ccu,
+            0x00000000u, 0xffff0000u, 0x8076337du, 0x4ff45ba0u, 0x09405033u, 0x034376d9u, 0x09000000u, 0x4e100501u, 0x076d14dcu, 0xf964f45eu,
+            0x0000003du
+        }
+    };
+    return table;
+}
+
+/// the 128-bit significand of 10^k, rounded down, for k in [-307, 341]
+/// (compute_pow10 of Zmij)
+inline uint128_parts compute_pow10(int k) noexcept
+{
+    const auto i = static_cast<unsigned>(k + 307);
+    const std::uint64_t m = pow10_minor()[(i + 24) % 28];
+    const std::size_t j = 2 * static_cast<std::size_t>((i + 24) / 28);
+    const std::uint64_t h_hi = pow10_major()[j];
+    const std::uint64_t h_lo = pow10_major()[j + 1];
+    const std::uint64_t h1 = full_multiplication(h_lo, m).high;
+    const std::uint64_t c0 = h_lo * m;
+    const std::uint64_t c1 = h1 + (h_hi * m);
+    const std::uint64_t c2 = (c1 < h1 ? 1u : 0u) + full_multiplication(h_hi, m).high;
+    uint128_parts r{};
+    if ((c2 >> 63u) != 0)
+    {
+        r.high = c2;
+        r.low = c1;
+    }
+    else
+    {
+        r.high = (c2 << 1u) | (c1 >> 63u);
+        r.low = (c1 << 1u) | (c0 >> 63u);
+    }
+    r.low -= (pow10_fixups()[i >> 5u] >> (i & 31u)) & 1u;
+    return r;
+}
+
+/// The 128-bit significand of 10^k, rounded down, for k in [-342, 341].
+/// Up to 10^308, the table for number parsing holds the same significands
+/// (those of 5^k), except for k in [-27, -1], where it holds them one unit
+/// larger (as the Eisel-Lemire algorithm needs them).
+inline uint128_parts pow10(int k) noexcept
+{
+    if (k > pow5_128_largest_power)
+    {
+        return compute_pow10(k); // (only for the smallest doubles)
+    }
+    const auto i = 2 * static_cast<std::size_t>(k - pow5_128_smallest_power);
+    uint128_parts r{pow5_128()[i + 1], pow5_128()[i]};
+    const std::uint64_t adjust = static_cast<unsigned>(k + 27) < 27u ? 1u : 0u;
+    r.high -= r.low < adjust ? 1u : 0u;
+    r.low -= adjust;
+    return r;
+}
+
+/// (x_hi * 2^64 + x_lo) * y >> 64, as 128 bits
+inline uint128_parts umul192_hi128(std::uint64_t x_hi, std::uint64_t x_lo, std::uint64_t y) noexcept
+{
+    const uint128_parts p = full_multiplication(x_hi, y);
+    uint128_parts r{};
+    r.low = p.low + full_multiplication(x_lo, y).high;
+    r.high = p.high + (r.low < p.low ? 1u : 0u);
+    return r;
+}
+
+/// (x * y + c) >> 64
+inline std::uint64_t umul128_add_hi64(std::uint64_t x, std::uint64_t y, std::uint64_t c) noexcept
+{
+    const uint128_parts p = full_multiplication(x, y);
+    return p.high + (p.low + c < p.low ? 1u : 0u);
+}
+
+/// The shortest decimal in the rounding interval of a positive finite double
+/// given by its bits, the closest one if there are several (to_decimal of
+/// Zmij). The significand can end in zeros.
+inline decimal to_decimal(std::uint64_t bits) noexcept
+{
+    constexpr int extra_shift = 9;
+    const auto raw_exp = static_cast<int>((bits >> 52u) & 0x7FFu);
+    std::uint64_t bin_sig = bits & ((std::uint64_t{1} << 52u) - 1);
+    // a power of two has a narrower interval below (except the smallest normal)
+    const bool regular = bin_sig != 0 || raw_exp <= 1;
+    const int bin_exp = (raw_exp == 0 ? 1 : raw_exp) - 1075;
+    if (raw_exp != 0)
+    {
+        bin_sig |= std::uint64_t{1} << 52u;
+    }
+    // floor(log10(2^bin_exp)), or floor(log10(3/4 * 2^bin_exp)) for the irregular case
+    const int dec_exp = ((bin_exp * 315653) - (regular ? 0 : 131072)) >> 20;
+    // scaled by 10^(-dec_exp - 1): the integral part is the shorter candidate
+    const int shift = bin_exp + ((-(dec_exp + 1) * 217707) >> 16) + 1 + extra_shift;
+    const uint128_parts p10 = pow10(-dec_exp - 1);
+    const uint128_parts p = umul192_hi128(p10.high, p10.low, bin_sig << static_cast<unsigned>(shift));
+    std::uint64_t integral = p.high >> static_cast<unsigned>(extra_shift);
+    const std::uint64_t fractional = (p.high << static_cast<unsigned>(64 - extra_shift)) | (p.low >> static_cast<unsigned>(extra_shift));
+    std::uint64_t digit = 0;
+    bool round_up = false;
+    bool round_down = false;
+    if (JSON_HEDLEY_LIKELY(regular))
+    {
+        const std::uint64_t half_ulp = (p10.high >> static_cast<unsigned>(extra_shift + 1 - shift)) + (1 - (bin_sig & 1u));
+        round_up = fractional + half_ulp < fractional;
+        round_down = half_ulp > fractional;
+        // the last digit of the longer candidate, rounded to nearest
+        digit = umul128_add_hi64(fractional, 10, (std::uint64_t{1} << 63u) + 6);
+        if (fractional == (std::uint64_t{1} << 62u))
+        {
+            digit = 2; // 2.5 rounds to 2
+        }
+    }
+    else
+    {
+        const std::uint64_t half_ulp = p10.high >> static_cast<unsigned>(extra_shift + 1 - shift);
+        round_up = half_ulp > ~std::uint64_t{0} - fractional;
+        round_down = (half_ulp >> 1u) > fractional;
+        digit = umul128_add_hi64(fractional, 10, (std::uint64_t{1} << 63u) - 1);
+        const std::uint64_t lowest = umul128_add_hi64(fractional - (half_ulp >> 1u), 10, ~std::uint64_t{0});
+        digit = digit < lowest ? lowest : digit;
+    }
+    integral += round_up ? 1u : 0u;
+    if (!round_up && !round_down)
+    {
+        // the shorter candidate is outside the rounding interval: one digit more
+        return decimal{(integral * 10) + digit, dec_exp};
+    }
+    return decimal{integral, dec_exp + 1};
+}
+
+}  // namespace zmij
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
 // #include <nlohmann/detail/macro_scope.hpp>


@@ -24826,6 +25055,87 @@ void grisu2(char* buf, int& len, int& decimal_exponent, FloatType value)
    grisu2(buf, len, decimal_exponent, w.minus, w.w, w.plus);
 }

+/*!
+@brief the shortest digits of a positive finite float (other than double): Grisu2
+*/
+template<typename FloatType>
+JSON_HEDLEY_NON_NULL(1)
+void shortest_digits(char* buf, int& len, int& decimal_exponent, FloatType value)
+{
+    grisu2(buf, len, decimal_exponent, value);
+}
+
+/*!
+@brief the shortest digits of a positive finite double: the conversion of
+Zmij (see zmij.hpp), which always finds the shortest digits that read back as
+the same value (Grisu2 does not for about one double in a thousand), and the
+closest of them if there are several
+
+v = buf * 10^decimal_exponent, as for grisu2()
+*/
+JSON_HEDLEY_NON_NULL(1)
+inline void shortest_digits(char* buf, int& len, int& decimal_exponent, double value)
+{
+    static_assert(std::numeric_limits<double>::is_iec559 && std::numeric_limits<double>::digits == 53,
+                  "internal error: the conversion of Zmij needs IEEE 754 binary64 doubles");
+    JSON_ASSERT(std::isfinite(value));
+    JSON_ASSERT(value > 0);
+
+    std::uint64_t bits = 0;
+    std::memcpy(&bits, &value, sizeof(bits));
+    zmij::decimal d = zmij::to_decimal(bits);
+    // without trailing zeros (up to 16): 8, 4, 2, 1 at a time
+    while (d.significand % 100000000 == 0)
+    {
+        d.significand /= 100000000;
+        d.exponent += 8;
+    }
+    if (d.significand % 10000 == 0)
+    {
+        d.significand /= 10000;
+        d.exponent += 4;
+    }
+    if (d.significand % 100 == 0)
+    {
+        d.significand /= 100;
+        d.exponent += 2;
+    }
+    if (d.significand % 10 == 0)
+    {
+        d.significand /= 10;
+        d.exponent += 1;
+    }
+    // at most 17 digits, written from the back two at a time
+    static constexpr const char* pairs =
+        "00010203040506070809101112131415161718192021222324252627282930313233343536373839"
+        "40414243444546474849505152535455565758596061626364656667686970717273747576777879"
+        "8081828384858687888990919293949596979899";
+    std::array<char, 20> digits{};
+    std::size_t n = digits.size();
+    while (d.significand >= 100)
+    {
+        const auto i = static_cast<std::size_t>(d.significand % 100) * 2;
+        d.significand /= 100;
+        n -= 2;
+        digits[n] = pairs[i];
+        digits[n + 1] = pairs[i + 1];
+    }
+    if (d.significand >= 10)
+    {
+        const auto i = static_cast<std::size_t>(d.significand) * 2;
+        n -= 2;
+        digits[n] = pairs[i];
+        digits[n + 1] = pairs[i + 1];
+    }
+    else
+    {
+        digits[--n] = static_cast<char>('0' + d.significand);
+    }
+    len = static_cast<int>(digits.size() - n);
+    std::memcpy(buf, digits.data() + n, static_cast<std::size_t>(len));
+    decimal_exponent = d.exponent;
+}
+
 /*!
@brief appends a decimal representation of e to buf
@return a pointer to the element following the exponent.
@@ -24955,6 +25265,177 @@ inline char* format_buffer(char* buf, int len, int decimal_exponent,
    return append_exponent(buf, n - 1);
 }

+/// eight decimal digits (a value below 10^8) as bytes 0..9, the first digit
+/// in the most significant byte: three steps that divide all lanes at once
+/// by a multiplication (the conversion of Xiang JunBo, as in Zmij)
+inline std::uint64_t eight_digit_bytes(std::uint64_t abcdefgh) noexcept
+{
+    const std::uint64_t abcd_efgh = abcdefgh + (((std::uint64_t{1} << 32u) - 10000u) * ((abcdefgh * (((std::uint64_t{1} << 40u) / 10000u) + 1u)) >> 40u));
+    const std::uint64_t ab_cd_ef_gh = abcd_efgh + (((std::uint64_t{1} << 16u) - 100u) * (((abcd_efgh * (((std::uint64_t{1} << 19u) / 100u) + 1u)) >> 19u) & 0x7F0000007Fu));
+    return ab_cd_ef_gh + (((std::uint64_t{1} << 8u) - 10u) * (((ab_cd_ef_gh * (((std::uint64_t{1} << 10u) / 10u) + 1u)) >> 10u) & 0x000F000F000F000Fu));
+}
+
+/// store the bytes of v, the most significant one first (one byte swap and
+/// one store where the byte order is known: compilers do not reliably merge
+/// the byte stores once this is inlined)
+inline void store_msb_first(char* p, std::uint64_t v) noexcept
+{
+#if defined(__BYTE_ORDER__) && defined(__ORDER_LITTLE_ENDIAN__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+    v = __builtin_bswap64(v);
+    std::memcpy(p, &v, sizeof(v));
+#elif defined(__BYTE_ORDER__) && defined(__ORDER_BIG_ENDIAN__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
+    std::memcpy(p, &v, sizeof(v));
+#elif defined(_MSC_VER) // (little-endian on all its targets)
+    v = _byteswap_uint64(v);
+    std::memcpy(p, &v, sizeof(v));
+#else
+    for (unsigned i = 0; i < 8; ++i)
+    {
+        p[i] = static_cast<char>(v >> (56u - (8u * i)));
+    }
+#endif
+}
+
+/*!
+@brief digits * 10^exp for a double, in the layout of format_buffer()
+
+The layout is that of format_buffer() with min_exp -4 and max_exp 15 (the
+digits10 of double). The digits are converted eight at a time and placed
+with fixed-size moves instead of per-digit loops and moves of the buffer.
+
+@param[in] digits  the digits (not 0, at most 17 digits; trailing zeros allowed)
+@param[in] exp     the decimal exponent of the last digit
+@return a pointer past the text; up to 41 bytes at @a first are written
+        (some beyond the returned end)
+*/
+JSON_HEDLEY_NON_NULL(1)
+JSON_HEDLEY_RETURNS_NON_NULL
+inline char* write_decimal(char* first, std::uint64_t digits, int exp) noexcept
+{
+    JSON_ASSERT(digits != 0 && digits < 100000000000000000u);
+    const std::uint64_t upper = digits / 100000000u;
+    const std::uint64_t b0 = upper / 100000000u; // (one digit: it is its own byte)
+    const std::uint64_t b1 = eight_digit_bytes(upper % 100000000u);
+    const std::uint64_t b2 = eight_digit_bytes(digits % 100000000u);
+    // leading and trailing zero digits: zero bytes, counted without division
+    int leading = 16;
+    int zeros = 16;
+    if (b0 != 0)
+    {
+        leading = count_leading_zeros(b0) / 8;
+    }
+    else if (b1 != 0)
+    {
+        leading = 8 + (count_leading_zeros(b1) / 8);
+    }
+    else
+    {
+        leading += count_leading_zeros(b2) / 8;
+    }
+    if (b2 != 0)
+    {
+        zeros = count_trailing_zeros(b2) / 8;
+    }
+    else if (b1 != 0)
+    {
+        zeros = 8 + (count_trailing_zeros(b1) / 8);
+    }
+    // (else: 16, b0 is the one digit that is not 0)
+    // the digits as text at text + leading, then '0's, so that fixed-size
+    // moves need not check how many digits there are
+    std::array<char, 64> text; // NOLINT(cppcoreguidelines-pro-type-member-init,hicpp-member-init): written before read
+    store_msb_first(text.data(), b0 + 0x3030303030303030u);
+    store_msb_first(text.data() + 8, b1 + 0x3030303030303030u);
+    store_msb_first(text.data() + 16, b2 + 0x3030303030303030u);
+    std::memset(text.data() + 24, '0', 40);
+    const int k = 24 - leading - zeros; // significant digits
+    const int n = k + exp + zeros;      // position of the decimal point after the first digit
+    const char* const s0 = text.data() + leading;
+
+    if (-4 < n && n <= 15)
+    {
+        // "0.[000]digits" (n <= 0) is the digits after 1 - n leading '0's
+        // with the point after the first; "digits[000].0" (n >= k) and
+        // "dig.its" put the point after n characters
+        const int pad = n <= 0 ? 1 - n : 0;
+        const char* const s = s0 - pad;
+        const int len = k + pad;
+        const int point = n + pad;
+        std::memcpy(first, s, 16);
+        std::memcpy(first + point + 1, s + point, 24);
+        first[point] = '.';
+        return first + (point >= len ? point + 2 : len + 1);
+    }
+
+    // d.igitse+XX, with at least two exponent digits (as append_exponent())
+    std::memcpy(first, s0, 16);
+    std::memcpy(first + 2, s0 + 1, 16);
+    first[1] = '.';
+    char* const end = first + (k == 1 ? 1 : k + 1);
+    const int e = n - 1;
+    const auto ea = static_cast<unsigned>(e < 0 ? -e : e);
+    const bool three = ea >= 100;
+    end[0] = 'e';
+    end[1] = e < 0 ? '-' : '+';
+    end[2] = static_cast<char>('0' + (three ? ea / 100 : (ea / 10) % 10));
+    end[3] = static_cast<char>('0' + (three ? (ea / 10) % 10 : ea % 10));
+    end[4] = static_cast<char>('0' + (ea % 10));
+    return end + (three ? 5 : 4);
+}
+
+/// a positive finite float (other than double): Grisu2 and format_buffer()
+template<typename FloatType>
+JSON_HEDLEY_NON_NULL(1, 2)
+JSON_HEDLEY_RETURNS_NON_NULL
+char* write_positive(char* first, const char* last, FloatType value)
+{
+    JSON_ASSERT(last - first >= std::numeric_limits<FloatType>::max_digits10);
+
+    // Compute v = buffer * 10^decimal_exponent.
+    // The decimal digits are stored in the buffer, which needs to be interpreted
+    // as an unsigned decimal integer.
+    // len is the length of the buffer, i.e., the number of decimal digits.
+    int len = 0;
+    int decimal_exponent = 0;
+    shortest_digits(first, len, decimal_exponent, value);
+
+    JSON_ASSERT(len <= std::numeric_limits<FloatType>::max_digits10);
+
+    // Format the buffer like printf("%.*g", prec, value)
+    constexpr int kMinExp = -4;
+    // Use digits10 here to increase compatibility with version 2.
+    constexpr int kMaxExp = std::numeric_limits<FloatType>::digits10;
+
+    JSON_ASSERT(last - first >= kMaxExp + 2);
+    JSON_ASSERT(last - first >= 2 + (-kMinExp - 1) + std::numeric_limits<FloatType>::max_digits10);
+    JSON_ASSERT(last - first >= std::numeric_limits<FloatType>::max_digits10 + 6);
+
+    return format_buffer(first, len, decimal_exponent, kMinExp, kMaxExp);
+}
+
+/// a positive finite double: the shortest digits (Zmij), laid out by
+/// write_decimal() (through a local buffer if [first, last) is shorter than
+/// the 41 bytes it may write)
+JSON_HEDLEY_NON_NULL(1, 2)
+JSON_HEDLEY_RETURNS_NON_NULL
+inline char* write_positive(char* first, const char* last, double value)
+{
+    static_assert(std::numeric_limits<double>::is_iec559 && std::numeric_limits<double>::digits == 53,
+                  "internal error: the conversion of Zmij needs IEEE 754 binary64 doubles");
+    std::uint64_t bits = 0;
+    std::memcpy(&bits, &value, sizeof(bits));
+    const zmij::decimal d = zmij::to_decimal(bits);
+    if (JSON_HEDLEY_LIKELY(last - first >= 41))
+    {
+        return write_decimal(first, d.significand, d.exponent);
+    }
+    std::array<char, 64> buf; // NOLINT(cppcoreguidelines-pro-type-member-init,hicpp-member-init): written before read
+    const auto len = static_cast<std::size_t>(write_decimal(buf.data(), d.significand, d.exponent) - buf.data());
+    JSON_ASSERT(static_cast<std::size_t>(last - first) >= len);
+    std::memcpy(first, buf.data(), len);
+    return first + len;
+}
+
 }  // namespace dtoa_impl

 /*!
@@ -24972,7 +25453,6 @@ JSON_HEDLEY_NON_NULL(1, 2)
 JSON_HEDLEY_RETURNS_NON_NULL
 char* to_chars(char* first, const char* last, FloatType value)
 {
-    static_cast<void>(last); // maybe unused - fix warning
    JSON_ASSERT(std::isfinite(value));

    // Use signbit(value) instead of (value < 0) since signbit works for -0.
@@ -24998,28 +25478,7 @@ char* to_chars(char* first, const char* last, FloatType value)
    JSON_HEDLEY_DIAGNOSTIC_POP
 #endif

-    JSON_ASSERT(last - first >= std::numeric_limits<FloatType>::max_digits10);
-
-    // Compute v = buffer * 10^decimal_exponent.
-    // The decimal digits are stored in the buffer, which needs to be interpreted
-    // as an unsigned decimal integer.
-    // len is the length of the buffer, i.e., the number of decimal digits.
-    int len = 0;
-    int decimal_exponent = 0;
-    dtoa_impl::grisu2(first, len, decimal_exponent, value);
-
-    JSON_ASSERT(len <= std::numeric_limits<FloatType>::max_digits10);
-
-    // Format the buffer like printf("%.*g", prec, value)
-    constexpr int kMinExp = -4;
-    // Use digits10 here to increase compatibility with version 2.
-    constexpr int kMaxExp = std::numeric_limits<FloatType>::digits10;
-
-    JSON_ASSERT(last - first >= kMaxExp + 2);
-    JSON_ASSERT(last - first >= 2 + (-kMinExp - 1) + std::numeric_limits<FloatType>::max_digits10);
-    JSON_ASSERT(last - first >= std::numeric_limits<FloatType>::max_digits10 + 6);
-
-    return dtoa_impl::format_buffer(first, len, decimal_exponent, kMinExp, kMaxExp);
+    return dtoa_impl::write_positive(first, last, value);
 }

 }  // namespace detail
@@ -10,7 +10,7 @@ CXXFLAGS += -std=c++11
 CPPFLAGS += -I ../single_include

 FUZZER_ENGINE = src/fuzzer-driver_afl.cpp
-FUZZERS = parse_afl_fuzzer parse_bson_fuzzer parse_cbor_fuzzer parse_msgpack_fuzzer parse_ubjson_fuzzer parse_bjdata_fuzzer parse_bon8_fuzzer parse_json_view_fuzzer
+FUZZERS = parse_afl_fuzzer parse_bson_fuzzer parse_cbor_fuzzer parse_msgpack_fuzzer parse_ubjson_fuzzer parse_bjdata_fuzzer parse_bon8_fuzzer parse_json_view_fuzzer json_view_image_fuzzer
 fuzzers: $(FUZZERS)

 parse_afl_fuzzer:
@@ -19,6 +19,9 @@ parse_afl_fuzzer:
 parse_json_view_fuzzer:
 	$(CXX) $(CXXFLAGS) $(CPPFLAGS) $(FUZZER_ENGINE) src/fuzzer-parse_json_view.cpp -o $@

+json_view_image_fuzzer:
+	$(CXX) $(CXXFLAGS) $(CPPFLAGS) $(FUZZER_ENGINE) src/fuzzer-json_view_image.cpp -o $@
+
 parse_bson_fuzzer:
 	$(CXX) $(CXXFLAGS) $(CPPFLAGS) $(FUZZER_ENGINE) src/fuzzer-parse_bson.cpp -o $@

@@ -10,6 +10,12 @@ produces, and that a rejected input makes both parsers throw with an identical `
 reuses the `corpus_json` corpus (or, for the `make fuzz_testing_json_view` target below, `tests/data/json_tests`) rather
 than a format of its own.

+`json_view_image_fuzzer` (`tests/src/fuzzer-json_view_image.cpp`) tests the images of `json_document` (`save()` and
+`load()`). It uses each input twice: as an image, which `load()` must either reject with `parse_error.116` or read
+safely (with `image_check::full`, the document must also serialize to the JSON it reads as), and as a JSON text, whose
+image must load and serialize to the same text. A corpus of images can be made from JSON files with a small program
+that calls `json_document::parse(text).save()`; plain JSON files work as well.
+
 ## Corpus creation

 For most effective fuzzing, a [corpus](https://llvm.org/docs/LibFuzzer.html#corpus) should be provided. A corpus is a
@@ -0,0 +1,89 @@
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  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
+
+/*
+This file implements a test of json_document images suitable for fuzz
+testing. The input is used twice:
+
+- as an image: json_document::load() with image_check::full must either throw
+  a parse_error or yield a document that serializes to the JSON text it reads
+  as; with image_check::bounds, reading and serializing must be safe (checked
+  by the sanitizers), and serializing may only throw type_error.316
+- as a JSON text: if json_document::parse() accepts it, the image of the
+  document must load (with every check) and serialize to the same text
+
+The provided function `LLVMFuzzerTestOneInput` can be used in different fuzzer
+drivers.
+*/
+
+#include <cassert>
+#include <cstdint>
+#include <string>
+#include <vector>
+#include <nlohmann/json.hpp>
+#include <nlohmann/json_view.hpp>
+
+// the checks below are assertions; NDEBUG would compile them away
+#ifdef NDEBUG
+    #error "the fuzzer drivers must be built without NDEBUG"
+#endif
+
+using json = nlohmann::json;
+using json_document = nlohmann::json_document;
+using image_check = json_document::image_check;
+
+// see http://llvm.org/docs/LibFuzzer.html
+extern "C" int LLVMFuzzerTestOneInput(const uint8_t* data, size_t size)
+{
+    // the input as an image
+    for (const image_check check : {image_check::full, image_check::bounds})
+    {
+        json_document d;
+        try
+        {
+            d = json_document::load(data, size, check);
+        }
+        catch (const json::parse_error& e)
+        {
+            assert(e.id == 116);
+            continue;
+        }
+        std::string dumped;
+        try
+        {
+            dumped = d.root().dump();
+        }
+        catch (const json::type_error& e)
+        {
+            // invalid UTF-8 can only pass the bounds check
+            assert(check == image_check::bounds && e.id == 316);
+            continue;
+        }
+        const json j = d.root().materialize();
+        if (check == image_check::full)
+        {
+            assert(json::parse(dumped) == j);
+            // an image of the loaded document is the input
+            assert(d.save() == std::vector<std::uint8_t>(data, data + size));
+        }
+    }
+
+    // the input as a JSON text
+    const std::string text(reinterpret_cast<const char*>(data), size); // NOLINT(cppcoreguidelines-pro-type-reinterpret-cast)
+    const json_document parsed = json_document::parse(text, false);
+    if (!parsed.is_discarded())
+    {
+        const std::vector<std::uint8_t> image = parsed.save();
+        for (const image_check check : {image_check::full, image_check::bounds, image_check::none})
+        {
+            const json_document loaded = json_document::load(image, check);
+            assert(loaded.root().dump() == parsed.root().dump());
+        }
+    }
+    return 0;
+}
@@ -33,7 +33,7 @@ TEST_CASE("Binary Formats" * doctest::skip())
        const auto ubjson_2_size = json::to_ubjson(j, true).size();
        const auto ubjson_3_size = json::to_ubjson(j, true, true).size();

-        CHECK(json_size == 2090303);
+        CHECK(json_size == 2090234);
        CHECK(bjdata_1_size == 1112030);
        CHECK(bjdata_2_size == 1224148);
        CHECK(bjdata_3_size == 1224148);
@@ -46,16 +46,16 @@ TEST_CASE("Binary Formats" * doctest::skip())
        CHECK(ubjson_3_size == 1169069);

        CHECK((100.0 * double(json_size) / double(json_size)) == Approx(100.0));
-        CHECK((100.0 * double(bjdata_1_size) / double(json_size)) == Approx(53.199));
-        CHECK((100.0 * double(bjdata_2_size) / double(json_size)) == Approx(58.563));
-        CHECK((100.0 * double(bjdata_3_size) / double(json_size)) == Approx(58.563));
-        CHECK((100.0 * double(bon8_size) / double(json_size)) == Approx(50.509));
-        CHECK((100.0 * double(bson_size) / double(json_size)) == Approx(85.849));
-        CHECK((100.0 * double(cbor_size) / double(json_size)) == Approx(50.497));
-        CHECK((100.0 * double(msgpack_size) / double(json_size)) == Approx(50.526));
-        CHECK((100.0 * double(ubjson_1_size) / double(json_size)) == Approx(53.199));
-        CHECK((100.0 * double(ubjson_2_size) / double(json_size)) == Approx(58.563));
-        CHECK((100.0 * double(ubjson_3_size) / double(json_size)) == Approx(55.928));
+        CHECK((100.0 * double(bjdata_1_size) / double(json_size)) == Approx(53.201));
+        CHECK((100.0 * double(bjdata_2_size) / double(json_size)) == Approx(58.565));
+        CHECK((100.0 * double(bjdata_3_size) / double(json_size)) == Approx(58.565));
+        CHECK((100.0 * double(bon8_size) / double(json_size)) == Approx(50.511));
+        CHECK((100.0 * double(bson_size) / double(json_size)) == Approx(85.853));
+        CHECK((100.0 * double(cbor_size) / double(json_size)) == Approx(50.499));
+        CHECK((100.0 * double(msgpack_size) / double(json_size)) == Approx(50.528));
+        CHECK((100.0 * double(ubjson_1_size) / double(json_size)) == Approx(53.201));
+        CHECK((100.0 * double(ubjson_2_size) / double(json_size)) == Approx(58.565));
+        CHECK((100.0 * double(ubjson_3_size) / double(json_size)) == Approx(55.930));
    }

    SECTION("twitter.json")
@@ -1089,6 +1089,9 @@ TEST_CASE("json_view dump")
        CHECK(d.root().dump() == json::parse(text).dump());
        CHECK(d.root().dump() == "[1.5,100.0,0,-0.0,1.2345678901234568e+29,18446744073709551615,-9223372036854775808,0.1,1e-07,5e-324]");
        CHECK(d.root().dump(-1, ' ', false, json_view::number_format::source) == "[1.50,1E2,-0,-0.0,123456789012345678901234567890,18446744073709551615,-9223372036854775808,0.1,1e-7,5e-324]");
+        // also indented, and with ensure_ascii
+        CHECK(d.root().dump(0, ' ', false, json_view::number_format::source) == "[\n1.50,\n1E2,\n-0,\n-0.0,\n123456789012345678901234567890,\n18446744073709551615,\n-9223372036854775808,\n0.1,\n1e-7,\n5e-324\n]");
+        CHECK(d.root().dump(-1, ' ', true, json_view::number_format::source) == "[1.50,1E2,-0,-0.0,123456789012345678901234567890,18446744073709551615,-9223372036854775808,0.1,1e-7,5e-324]");

        // random doubles, written as parse() and dump() would
        std::mt19937_64 rng(1170); // NOLINT(cert-msc32-c,cert-msc51-cpp,bugprone-random-generator-seed)
@@ -0,0 +1,791 @@
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  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
+
+#include "doctest_compatibility.h"
+
+#include <nlohmann/json_view.hpp>
+using nlohmann::json;
+using nlohmann::ordered_json;
+using nlohmann::json_document;
+using nlohmann::json_editable_document;
+using nlohmann::ordered_json_document;
+using nlohmann::ordered_json_editable_document;
+using image_check = json_document::image_check;
+using nlohmann::detail::view::node;
+
+#include <array>
+#include <cstdint>
+#include <cstring>
+#include <fstream>
+#include <functional>
+#include <limits>
+#include <random>
+#include <sstream>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include <test_data.hpp>
+
+#if !(defined(__BYTE_ORDER__) && defined(__ORDER_BIG_ENDIAN__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
+
+namespace
+{
+std::string exception_of(const std::function<void()>& f)
+{
+    try
+    {
+        f();
+    }
+    catch (const json::exception& e)
+    {
+        return e.what();
+    }
+    return "";
+}
+
+const char* const check_failed = "[json.exception.parse_error.116] parse error: invalid json_document image: the check failed";
+
+std::string read_file(const std::string& name)
+{
+    std::ifstream f(std::string(TEST_DATA_DIRECTORY) + name, std::ios::binary);
+    std::stringstream ss;
+    ss << f.rdbuf();
+    return ss.str();
+}
+
+// the offsets of the parts of an image
+constexpr std::size_t header_size = 64;
+
+std::uint64_t header_field(const std::vector<std::uint8_t>& image, std::size_t offset)
+{
+    std::uint64_t v = 0;
+    std::memcpy(&v, image.data() + offset, sizeof(v));
+    return v;
+}
+
+void set_header_field(std::vector<std::uint8_t>& image, std::size_t offset, std::uint64_t v)
+{
+    std::memcpy(image.data() + offset, &v, sizeof(v));
+}
+
+std::size_t node_count(const std::vector<std::uint8_t>& image)
+{
+    return static_cast<std::size_t>(header_field(image, 8));
+}
+
+std::size_t text_at(const std::vector<std::uint8_t>& image)
+{
+    return header_size + (node_count(image) * sizeof(node));
+}
+
+node node_at(const std::vector<std::uint8_t>& image, std::size_t i)
+{
+    node n{};
+    std::memcpy(&n, image.data() + header_size + (i * sizeof(node)), sizeof(node));
+    return n;
+}
+
+void set_node(std::vector<std::uint8_t>& image, std::size_t i, const node& n)
+{
+    std::memcpy(image.data() + header_size + (i * sizeof(node)), &n, sizeof(node));
+}
+
+/// the result of loading an image with a check: "" or the exception message
+std::string load_result(const std::vector<std::uint8_t>& image, image_check check)
+{
+    return exception_of([&]
+    {
+        const json_document d = json_document::load(image, check);
+        static_cast<void>(d);
+    });
+}
+
+/// a copy of the image with node i changed by f
+template<typename F>
+std::vector<std::uint8_t> corrupted(const std::vector<std::uint8_t>& image, std::size_t i, F f)
+{
+    std::vector<std::uint8_t> b = image;
+    node n = node_at(b, i);
+    f(n);
+    set_node(b, i, n);
+    return b;
+}
+
+/// a document and the documents loaded from its image must be equal
+template<typename Document>
+void check_round_trip(const Document& d)
+{
+    const std::vector<std::uint8_t> image = d.save();
+    for (const image_check check :
+            {
+                image_check::full, image_check::bounds, image_check::none
+            })
+    {
+        const json_document l = json_document::load(image, check);
+        CHECK(l.root().dump() == d.root().dump());
+        CHECK(l.root().dump(2) == d.root().dump(2));
+        CHECK(l.root().materialize() == json(d.root().materialize()));
+        // an image of a loaded document is the same image
+        CHECK(l.save() == image);
+    }
+    // an editable document can be loaded, too
+    const ordered_json_editable_document e = ordered_json_editable_document::load(image);
+    CHECK(e.root().dump() == d.root().dump());
+}
+
+std::uint32_t rng()
+{
+    static std::mt19937 generator(5295); // NOLINT(cert-msc32-c,cert-msc51-cpp,bugprone-random-generator-seed): reproducible
+    return generator();
+}
+} // namespace
+
+TEST_CASE("json_view images: round trips")
+{
+    SECTION("small documents")
+    {
+        for (const char* text :
+                {
+                    "null", "true", "false", "0", "-0", "42", "-42", "18446744073709551615", "-9223372036854775808",
+                    "123456789012345678901234567890", "1.5", "-1.25e-300", "1E308", "0.1000000000000000000000000001",
+                    "\"\"", "\"text\"", R"("esc\"aped\n\u00e9\ud83d\ude00")", "\"\xc3\xa9\xe3\x81\x82\"",
+                    "[]", "{}", "[[]]", "[{}]", "{\"\":{}}",
+                    R"({"a": [1, 2.5, "x\ty", true, null, {"b": []}], "c": {"d": -3, "eA": "f"}})",
+                    R"({"k": 1, "k": 2, "l": [], "k": 3})",
+                    "  [1 ,  2 ]  "
+                })
+        {
+            CAPTURE(text);
+            check_round_trip(json_document::parse(text));
+            check_round_trip(ordered_json_document::parse(text));
+        }
+    }
+
+    SECTION("files")
+    {
+        for (const char* name :
+                {
+                    "/json_testsuite/sample.json", "/nativejson-benchmark/canada.json", "/nativejson-benchmark/citm_catalog.json",
+                    "/nativejson-benchmark/twitter.json", "/json_tests/pass1.json", "/json_tests/pass2.json", "/json_tests/pass3.json"
+                })
+        {
+            CAPTURE(name);
+            const std::string text = read_file(name);
+            const json_document d = json_document::parse(text);
+            check_round_trip(d);
+            // what a loaded document reads is what parse() produces
+            CHECK(json_document::load(d.save()).root().materialize() == json::parse(text));
+        }
+    }
+
+    SECTION("images are deterministic")
+    {
+        const std::string text = R"({"b": [1, 2, {"c": "\u00e9"}], "a": 1.5})";
+        const json_document d = json_document::parse(text);
+        CHECK(d.save() == json_document::parse(text).save());
+        CHECK(d.save() == json_editable_document::parse(text).save());
+        CHECK(d.save() == ordered_json_document::parse(text).save());
+        const json_document copy = json_document::parse_copy(text);
+        CHECK(copy.save() == d.save());
+    }
+
+    SECTION("large objects get their hash index again")
+    {
+        std::string text = "{";
+        for (int i = 0; i < 1000; ++i)
+        {
+            text += (i != 0 ? ",\"k" : "\"k") + std::to_string(i) + "\":" + std::to_string(i);
+        }
+        text += R"(,"k7":"a duplicate","inner":{)";
+        for (int i = 0; i < 200; ++i)
+        {
+            text += (i != 0 ? ",\"m" : "\"m") + std::to_string(i) + "\":" + std::to_string(-i);
+        }
+        text += "}}";
+        const json_document d = json_document::parse(text);
+        const std::vector<std::uint8_t> image = d.save();
+        for (const image_check check :
+                {
+                    image_check::full, image_check::none
+                })
+        {
+            const json_document l = json_document::load(image, check);
+            for (int i = 0; i < 1000; ++i)
+            {
+                CHECK(l.root()["k" + std::to_string(i)] == d.root()["k" + std::to_string(i)]);
+            }
+            CHECK(l.root()["k7"].get<int>() == 7); // the first of duplicate keys
+            CHECK(l.root()["inner"]["m199"].get<int>() == -199);
+            CHECK(!l.root().contains("k1000"));
+            // the index is not part of the image
+            CHECK(l.save() == image);
+        }
+        // the nodes of objects in the image do not carry the number of an index
+        CHECK(node_at(image, 0).extra == 0);
+    }
+}
+
+TEST_CASE("json_view images: edited documents")
+{
+    const std::string text = R"({"name": "x", "n": 1, "f": 2.5, "list": [1, 2, 3], "obj": {"a": "\u00e9", "b": [true]}, "s": "a\"b"})";
+
+    SECTION("every kind of edit")
+    {
+        ordered_json_editable_document d = ordered_json_editable_document::parse(text);
+        d.set(d.root()["name"], "a new \"name\"");      // string in the edit arena
+        d.set(d.root()["n"], -17);                     // negative integer
+        d.set(d.root(), "p", 5);                       // non-negative number_integer
+        d.set(d.root(), "u", 18446744073709551615u);   // unsigned
+        d.set(d.root()["f"], 0.1);                     // float token
+        d.set(d.root(), "nan", std::numeric_limits<double>::quiet_NaN());
+        d.set(d.root(), "inf", -std::numeric_limits<double>::infinity());
+        d.push_back(d.root()["list"], "pushed");       // moved array
+        d.insert(d.root()["list"], 0, ordered_json::object({{"new", {1, 2}}}));
+        d.erase(d.root()["list"], 2);
+        d.erase(d.root(), "s");
+        d.set(d.root()["obj"], "c", ordered_json::array({1, "two", 3.5, nullptr, false})); // new object member with a new array
+        d.set(d.root(), "copy", d.root()["obj"]);     // a copy of a subtree
+        d.set(d.root(), "key \xc3\xa9", true);        // a key in the edit arena
+
+        const std::vector<std::uint8_t> image = d.save();
+        const ordered_json expected = ordered_json::parse(d.root().dump());
+        for (const image_check check :
+                {
+                    image_check::full, image_check::bounds, image_check::none
+                })
+        {
+            const ordered_json_document l = ordered_json_document::load(image, check);
+            CHECK(l.root().dump() == d.root().dump());
+            CHECK(l.root().materialize() == expected);
+            CHECK(l.root()["nan"].is_null());
+            CHECK(l.root()["inf"].is_null());
+            CHECK(l.root()["p"].is_number_integer());
+            CHECK(l.root()["p"].get<int>() == 5);
+            CHECK(l.root()["f"].get<double>() == 0.1);
+            CHECK(l.root()["u"].get<std::uint64_t>() == 18446744073709551615u);
+        }
+        // the node index is in document order again: an image of the loaded
+        // document is the same image
+        CHECK(ordered_json_document::load(image).save() == image);
+        // number tokens of edits follow the source; the text is the source's
+        // prefix
+        const ordered_json_document l = ordered_json_document::load(image);
+        REQUIRE(l.source().size() > text.size());
+        CHECK(std::string(l.source().data(), text.size()) == text);
+    }
+
+    SECTION("a loaded document can be edited and saved again")
+    {
+        const std::vector<std::uint8_t> first = json_editable_document::parse(text).save();
+        json_editable_document d = json_editable_document::load(first);
+        d.set(d.root()["obj"]["a"], "changed");
+        d.push_back(d.root()["list"], 4);
+        d.set(d.root(), "z", json::array({json::object()}));
+        const std::vector<std::uint8_t> second = d.save();
+        const json_document l = json_document::load(second);
+        CHECK(l.root().dump() == d.root().dump());
+        CHECK(l.root()["obj"]["a"] == "changed");
+        CHECK(l.root()["list"].size() == 4);
+    }
+
+    SECTION("the root replaced")
+    {
+        json_editable_document d = json_editable_document::parse(text);
+        d.set(d.root(), json::array({1, "x"}));
+        check_round_trip(d);
+        d.set(d.root(), 3.5);
+        check_round_trip(d);
+        d.set(d.root(), "text");
+        check_round_trip(d);
+    }
+}
+
+TEST_CASE("json_view images: ownership")
+{
+    const std::string text = R"({"a": "esc\u00e9aped", "b": [1, 2]})";
+    const std::vector<std::uint8_t> image = json_document::parse(text).save();
+
+    SECTION("borrowed")
+    {
+        const json_document d = json_document::load(image);
+        CHECK(!d.owns_source());
+        CHECK(d.root()["a"] == "esc\xc3\xa9" "aped");
+        const json_document p = json_document::load(image.data(), image.size());
+        CHECK(!p.owns_source());
+        CHECK(p.root() == d.root());
+        // the text is the image's
+        CHECK(d.source().data() == reinterpret_cast<const char*>(image.data() + text_at(image)));
+    }
+
+    SECTION("owned")
+    {
+        std::vector<std::uint8_t> copy = image;
+        const std::uint8_t* const data = copy.data();
+        json_document d = json_document::load(std::move(copy));
+        CHECK(d.owns_source());
+        CHECK(d.source().data() == reinterpret_cast<const char*>(data + text_at(image)));
+        CHECK(d.memory_usage() >= image.size());
+        CHECK(d.root()["b"][1] == 2);
+        // read() replaces the image
+        d.read(std::string("[1]"));
+        CHECK(d.owns_source());
+        CHECK(d.root().dump() == "[1]");
+        const std::string borrowed = "[2]";
+        d.read(borrowed);
+        CHECK(!d.owns_source());
+    }
+
+    SECTION("shrink_to_fit keeps the decoded strings of the image")
+    {
+        json_document d = json_document::parse(R"(["\u00e9\u00e9\u00e9\u00e9\u00e9\u00e9\u00e9\u00e9\u00e9\u00e9\u00e9\u00e9\u00e9\u00e9\u00e9\u00e9"])");
+        d.shrink_to_fit();
+        const std::vector<std::uint8_t> img = d.save();
+        json_document l = json_document::load(img);
+        l.shrink_to_fit();
+        CHECK(l.root().dump() == d.root().dump());
+        CHECK(l.save() == img);
+    }
+}
+
+TEST_CASE("json_view images: errors")
+{
+    SECTION("a literal as the root: dump() after loading")
+    {
+        for (const char* text :
+                {
+                    "null", "true", "false"
+                })
+        {
+            std::vector<std::uint8_t> image = json_document::parse(text).save();
+            node n = node_at(image, 0);
+            n.off = static_cast<std::uint32_t>(image.size());
+            set_node(image, 0, n);
+            CHECK(load_result(image, image_check::full) == check_failed);
+            CHECK(load_result(image, image_check::bounds) == check_failed);
+        }
+    }
+
+    SECTION("saving a discarded document")
+    {
+        const json_document empty;
+        CHECK(exception_of([&] { static_cast<void>(empty.save()); }) == "[json.exception.type_error.320] cannot save a discarded json_document");
+        const json_document failed = json_document::parse("[1,", false);
+        CHECK(exception_of([&] { static_cast<void>(failed.save()); }) == "[json.exception.type_error.320] cannot save a discarded json_document");
+    }
+
+    const std::vector<std::uint8_t> image = json_document::parse(R"({"a": [1, "\u00e9"]})").save();
+    const std::string prefix = "[json.exception.parse_error.116] parse error: invalid json_document image: ";
+
+    SECTION("header and sizes")
+    {
+        CHECK(exception_of([]
+        {
+            const json_document d = json_document::load(nullptr, 0);
+            static_cast<void>(d);
+        }) == prefix + "too short");
+        CHECK(exception_of([&]
+        {
+            const json_document d = json_document::load(image.data(), 63);
+            static_cast<void>(d);
+        }) == prefix + "too short");
+
+        std::vector<std::uint8_t> bad = image;
+        bad[0] = 'X';
+        CHECK(load_result(bad, image_check::full) == prefix + "unknown format");
+        bad = image;
+        bad[4] = 2; // version
+        CHECK(load_result(bad, image_check::full) == prefix + "unknown format");
+        for (std::size_t reserved = 32; reserved < 64; reserved += 8)
+        {
+            bad = image;
+            bad[reserved + 3] = 1;
+            CHECK(load_result(bad, image_check::none) == prefix + "unknown format");
+        }
+
+        const auto sizes = [&](std::size_t offset, std::uint64_t v)
+        {
+            std::vector<std::uint8_t> b = image;
+            set_header_field(b, offset, v);
+            return load_result(b, image_check::none);
+        };
+        CHECK(sizes(8, 0) == prefix + "sizes out of range");                 // no nodes
+        CHECK(sizes(8, 1000) == prefix + "sizes out of range");              // more nodes than bytes
+        CHECK(sizes(8, 0xFFFFFFF0u) == prefix + "sizes out of range");
+        CHECK(sizes(16, 0xFFFFFFF0u) == prefix + "sizes out of range");      // text size
+        CHECK(sizes(16, header_field(image, 16) + 1) == prefix + "sizes out of range");
+        CHECK(sizes(16, image.size()) == prefix + "sizes out of range");
+        CHECK(sizes(24, 0xFFFFFFF0u) == prefix + "sizes out of range");      // decoded string size
+        CHECK(sizes(24, header_field(image, 24) - 1) == prefix + "sizes out of range");
+
+        // the NULs after the text and the decoded strings
+        bad = image;
+        bad[text_at(image) + header_field(image, 16)] = 'x';
+        CHECK(load_result(bad, image_check::none) == prefix + "sizes out of range");
+        bad = image;
+        bad.back() = 'x';
+        CHECK(load_result(bad, image_check::none) == prefix + "sizes out of range");
+        // nothing after the image
+        bad = image;
+        bad.push_back(0);
+        CHECK(load_result(bad, image_check::none) == prefix + "sizes out of range");
+        // nodes, but not even room for the NULs
+        bad.assign(image.begin(), image.begin() + static_cast<std::ptrdiff_t>(text_at(image)));
+        CHECK(load_result(bad, image_check::none) == prefix + "sizes out of range");
+    }
+}
+
+TEST_CASE("json_view images: check")
+{
+    // nodes: 0 {  1 "s" 2 "x\"y" (escaped)  3 "i" 4 -12  5 "u" 6 7  7 "f" 8 1.5e300  9 "b" 10 true  11 "n" 12 null
+    //        13 "a" 14 [  15 "t"  16 {}  ]
+    const std::string text = R"({"s":"x\"y","i":-12,"u":7,"f":1.5e300,"b":true,"n":null,"a":["t",{}]})";
+    const std::vector<std::uint8_t> image = json_document::parse(text).save();
+    REQUIRE(load_result(image, image_check::full).empty());
+    REQUIRE(node_count(image) == 17);
+
+    // bounds: rejected by both checks; content: only by the full one
+    const auto rejected = [&](const std::vector<std::uint8_t>& b, bool bounds)
+    {
+        CHECK(load_result(b, image_check::full) == check_failed);
+        CHECK(load_result(b, image_check::bounds) == (bounds ? check_failed : ""));
+    };
+
+    SECTION("kinds")
+    {
+        const std::array<std::uint8_t, 4> kinds = {{8, 9, 10, 200}}; // binary, discarded, link, unknown
+        for (const std::uint8_t kind : kinds)
+        {
+            rejected(corrupted(image, 12, [&](node & n)
+            {
+                n.kind = kind;
+            }), true);
+        }
+        // a key that is not a string
+        rejected(corrupted(image, 1, [](node & n)
+        {
+            n.kind = 0;
+            n.len = 0;
+            n.off = 0;
+        }), true);
+    }
+
+    SECTION("flags and extra")
+    {
+        rejected(corrupted(image, 12, [](node & n)
+        {
+            n.flags = 4;
+        }), true);
+        rejected(corrupted(image, 12, [](node & n)
+        {
+            n.extra = 1;
+        }), true);
+        rejected(corrupted(image, 10, [](node & n)
+        {
+            n.flags = 5;
+        }), true);
+        rejected(corrupted(image, 10, [](node & n)
+        {
+            n.extra = 1;
+        }), true);
+        rejected(corrupted(image, 1, [](node & n)
+        {
+            n.flags = 2; // a string in the edit arena
+        }), true);
+        rejected(corrupted(image, 1, [](node & n)
+        {
+            n.extra = 3;
+        }), true);
+        rejected(corrupted(image, 4, [](node & n)
+        {
+            n.flags = 2;
+        }), true);
+        rejected(corrupted(image, 4, [](node & n)
+        {
+            n.extra = static_cast<std::uint16_t>(n.extra | 0x100u); // an integer with fraction digits
+        }), true);
+        rejected(corrupted(image, 0, [](node & n)
+        {
+            n.flags = 8; // moved
+        }), true);
+        rejected(corrupted(image, 0, [](node & n)
+        {
+            n.extra = 1; // a hash index
+        }), true);
+    }
+
+    SECTION("bounds")
+    {
+        const std::size_t text_size = header_field(image, 16);
+        const std::size_t arena_size = header_field(image, 24);
+        rejected(corrupted(image, 1, [&](node & n)
+        {
+            n.off = static_cast<std::uint32_t>(text_size + 1);
+        }), true);
+        rejected(corrupted(image, 1, [&](node & n)
+        {
+            n.len = static_cast<std::uint32_t>(text_size);
+        }), true);
+        rejected(corrupted(image, 2, [&](node & n)
+        {
+            n.len = static_cast<std::uint32_t>(arena_size + 1);
+        }), true);
+        rejected(corrupted(image, 6, [&](node & n)
+        {
+            n.off = static_cast<std::uint32_t>(text_size);
+        }), true);
+        rejected(corrupted(image, 6, [&](node & n)
+        {
+            n.off = static_cast<std::uint32_t>(text_size + 5);
+        }), true);
+        rejected(corrupted(image, 6, [](node & n)
+        {
+            n.extra = 0; // no digits
+        }), true);
+        rejected(corrupted(image, 8, [&](node & n)
+        {
+            n.len = static_cast<std::uint32_t>(text_size);
+        }), true);
+        rejected(corrupted(image, 8, [](node & n)
+        {
+            n.len = 2; // shorter than the recorded digits
+        }), true);
+        rejected(corrupted(image, 14, [&](node & n)
+        {
+            n.off = static_cast<std::uint32_t>(text_size + 1);
+        }), true);
+        // literals: their offset sizes the output of dump()
+        rejected(corrupted(image, 10, [&](node & n)
+        {
+            n.off = static_cast<std::uint32_t>(text_size + 1);
+        }), true);
+        rejected(corrupted(image, 12, [&](node & n)
+        {
+            n.off = 0xFFFFFFFFu;
+        }), true);
+    }
+
+    SECTION("structure")
+    {
+        rejected(corrupted(image, 0, [](node & n)
+        {
+            n.next = 0;
+        }), true);
+        rejected(corrupted(image, 0, [](node & n)
+        {
+            n.next = 18; // beyond the image
+        }), true);
+        rejected(corrupted(image, 14, [](node & n)
+        {
+            n.next = 4; // beyond the enclosing object
+        }), true);
+        rejected(corrupted(image, 0, [](node & n)
+        {
+            n.len = 6; // member count
+        }), true);
+        rejected(corrupted(image, 14, [](node & n)
+        {
+            n.len = 3; // element count
+        }), true);
+        rejected(corrupted(image, 0, [](node & n)
+        {
+            n.next = 14; // the object ends after the key "a"
+            n.len = 7;
+        }), true);
+        rejected(corrupted(image, 0, [](node & n)
+        {
+            n.next = 13; // nodes after the root
+            n.len = 6;
+        }), true);
+        rejected(corrupted(image, 0, [](node & n)
+        {
+            n.kind = 2; // an array: the "keys" are values, and the counts do not match
+        }), true);
+        const std::vector<std::uint8_t> as_array = corrupted(image, 16, [](node & n)
+        {
+            n.kind = 2; // {} as []: fine
+        });
+        CHECK(load_result(as_array, image_check::full).empty());
+        CHECK(json_document::load(as_array).root().dump() == R"({"s":"x\"y","i":-12,"u":7,"f":1.5e+300,"b":true,"n":null,"a":["t",[]]})");
+    }
+
+    SECTION("strings")
+    {
+        // a quote in a source string (the full check only)
+        std::vector<std::uint8_t> b = image;
+        const std::size_t t = text_at(image);
+        const node t15 = node_at(image, 15);
+        b[t + t15.off] = '"';
+        rejected(b, false);
+        // a control character
+        b[t + t15.off] = '\n';
+        rejected(b, false);
+        // invalid UTF-8 in a decoded string
+        b = image;
+        const node s2 = node_at(image, 2);
+        b[t + header_field(image, 16) + 1 + s2.off] = 0xFF;
+        rejected(b, false);
+    }
+
+    SECTION("numbers")
+    {
+        const std::size_t t = text_at(image);
+        const node i4 = node_at(image, 4);
+        const node u6 = node_at(image, 6);
+        const node f8 = node_at(image, 8);
+        const auto at_token = [&](const node & n, std::size_t k, std::uint8_t c)
+        {
+            std::vector<std::uint8_t> b = image;
+            b[t + n.off + k] = c;
+            return b;
+        };
+        rejected(at_token(i4, 1, 'x'), false);      // -x2
+        rejected(at_token(i4, 1, '0'), false);      // -02
+        rejected(at_token(i4, 0, '1'), false);      // 112 != -12
+        rejected(at_token(f8, 1, 'x'), false);      // 1x5e300
+        rejected(at_token(f8, 2, 'e'), false);      // 1.ee300
+        rejected(at_token(f8, 4, 'x'), false);      // 1.5ex00
+        rejected(at_token(f8, 3, '0'), false);      // 1.50300: another layout
+        rejected(at_token(f8, 4, '9'), false);      // 1.5e900: overflow
+        rejected(at_token(f8, 0, 'x'), false);
+        rejected(at_token(u6, 0, '8'), false);      // 8 != 7
+        rejected(corrupted(image, 4, [](node & n)
+        {
+            n.kind = 6; // "-12" as unsigned: a sign
+            n.extra = 3;
+        }), false);
+        // a non-negative number_integer (as edits write it): fine
+        const std::vector<std::uint8_t> positive = corrupted(image, 6, [](node & n)
+        {
+            n.kind = 5;
+            n.extra = 0;
+        });
+        CHECK(load_result(positive, image_check::full).empty());
+        CHECK(json_document::load(positive).root()["u"].is_number_integer());
+        rejected(corrupted(image, 8, [](node & n)
+        {
+            n.kind = 6; // a float token as integer
+            n.extra = 7;
+        }), false);
+    }
+
+    SECTION("float tokens of an image checked for bounds only")
+    {
+        // A float node whose layout records "many" digits is converted from
+        // its token alone; a token that is not a JSON number reads as 0.
+        const std::vector<std::uint8_t> img = json_document::parse("[1.5e300,2]").save();
+        const std::size_t t = text_at(img);
+        for (const char* token :
+                {
+                    "x.5e300", "01.5e30", "1.xe300", "1.5ex00", "1.5e+x0", "1.5e30x", "-.5e300", "1.5E300"
+                })
+        {
+            CAPTURE(token);
+            std::vector<std::uint8_t> b = img;
+            node n = node_at(b, 1);
+            n.extra = 0xFFFFu;
+            set_node(b, 1, n);
+            std::memcpy(b.data() + t + n.off, token, n.len);
+            const json_document d = json_document::load(b, image_check::bounds);
+            const auto v = d.root()[0].get<double>();
+            CHECK(v == (std::string(token) == "1.5E300" ? 1.5e300 : 0.0));
+            CHECK(load_result(b, image_check::full) == (std::string(token) == "1.5E300" ? "" : check_failed));
+        }
+    }
+
+    SECTION("integer ranges")
+    {
+        // tokens of many digits, which the parser stores as floats
+        const std::string big = R"([123456789012345678901234, 99999999999999999999, 9223372036854775808])";
+        const std::vector<std::uint8_t> img = json_document::parse(big).save();
+        const auto as_integer = [&](std::size_t i, std::uint8_t kind, std::uint16_t extra)
+        {
+            std::vector<std::uint8_t> b = img;
+            node n = node_at(b, i);
+            n.kind = kind;
+            n.extra = extra;
+            set_node(b, i, n);
+            return load_result(b, image_check::full);
+        };
+        CHECK(as_integer(1, 6, 24) == check_failed); // more than 20 digits
+        CHECK(as_integer(2, 6, 20) == check_failed); // more than 2^64 - 1
+        CHECK(as_integer(3, 5, 18) == check_failed); // more than 2^63 - 1 as number_integer
+    }
+}
+
+TEST_CASE("json_view images: damaged images")
+{
+    // A damaged image must be rejected, or read safely; with the full check,
+    // it also serializes to the JSON it reads as.
+    const std::vector<std::string> texts =
+    {
+        R"({"a": [1, -2, 3.25, "x\u00e9y", true, null], "b": {"c": "\"q\"", "d": 1e10}, "e": ""})",
+        R"([[[[]]], {"k": {"k": {"k": 12345678901234567890}}}, "\ud83d\ude00", -0.0, 0])",
+    };
+    for (const std::string& text : texts)
+    {
+        const std::vector<std::uint8_t> image = json_document::parse(text).save();
+        for (int round = 0; round < 3000; ++round)
+        {
+            std::vector<std::uint8_t> b = image;
+            const std::uint32_t flips = 1 + (rng() % 3);
+            for (std::uint32_t k = 0; k < flips; ++k)
+            {
+                // mostly the nodes, where the damage matters most
+                const std::size_t at = rng() % 4 != 0 ? header_size + (rng() % (b.size() - header_size)) : rng() % b.size();
+                b[at] = static_cast<std::uint8_t>(rng() % 3 == 0 ? rng() : b[at] ^ (1u << (rng() % 8)));
+            }
+            for (const image_check check :
+                    {
+                        image_check::full, image_check::bounds
+                    })
+            {
+                json_document d;
+                try
+                {
+                    d = json_document::load(b, check);
+                }
+                catch (const json::parse_error& e)
+                {
+                    CHECK(e.id == 116);
+                    continue;
+                }
+                std::string dumped;
+                std::string dumped_ascii;
+                try
+                {
+                    dumped = d.root().dump();
+                    dumped_ascii = d.root().dump(-1, ' ', true);
+                }
+                catch (const json::type_error& e)
+                {
+                    // invalid UTF-8 (the bounds check only)
+                    CHECK(check == image_check::bounds);
+                    CHECK(e.id == 316);
+                    continue;
+                }
+                const json j = d.root().materialize();
+                if (check == image_check::full)
+                {
+                    CHECK(json::parse(dumped) == j);
+                    CHECK(json::parse(dumped_ascii) == j);
+                }
+            }
+        }
+    }
+}
+
+#else
+
+TEST_CASE("json_view images: big-endian targets")
+{
+    const json_document d = json_document::parse("[1]");
+    CHECK_THROWS_WITH_AS(d.save(), "[json.exception.type_error.320] json_document images need a little-endian target", json::type_error&);
+}
+
+#endif
@@ -15,6 +15,20 @@
 #include <nlohmann/json.hpp>
 using nlohmann::detail::dtoa_impl::reinterpret_bits;

+#include <array>
+#include <cmath>
+#include <cstdint>
+#include <cstdio>
+#include <cstdlib>
+#include <limits>
+#include <random>
+#include <string>
+#include <utility>
+#include <vector>
+#if defined(JSON_HAS_CPP_17)
+    #include <charconv>
+#endif
+
 namespace
 {
 float make_float(uint32_t sign_bit, uint32_t biased_exponent, uint32_t significand)
@@ -450,7 +464,7 @@ TEST_CASE("formatting")
        check_double(  1.2345e+18,   "1.2345e+18"             ); //  1.2345e+18                1.2345e+18                1.2345e18
        check_double(  1.2345e+19,   "1.2345e+19"             ); //  1.2345e+19                1.2345e+19                1.2345e19
        check_double(  1.2345e+20,   "1.2345e+20"             ); //  1.2345e+20                1.2345e+20                1.2345e20
-        check_double(  1.2345e+21,   "1.2344999999999999e+21" ); //  1.2345e+21                1.2344999999999999e+21    1.2345e21
+        check_double(  1.2345e+21,   "1.2345e+21"             ); //  1.2345e+21                1.2344999999999999e+21    1.2345e21
        check_double(  1.2345e+22,   "1.2345e+22"             ); //  1.2345e+22                1.2345e+22                1.2345e22
    }

@@ -514,3 +528,243 @@ TEST_CASE("formatting")
        check_integer(1000000000000000000LL, "1000000000000000000");
    }
 }
+
+namespace
+{
+// a small unsigned big integer (32-bit limbs, least significant first), to
+// recompute the powers of ten of the shortest double conversion
+using big = std::vector<std::uint32_t>;
+
+void big_mul_small(big& x, std::uint32_t m)
+{
+    std::uint64_t carry = 0;
+    for (auto& limb : x)
+    {
+        const std::uint64_t v = (static_cast<std::uint64_t>(limb) * m) + carry;
+        limb = static_cast<std::uint32_t>(v);
+        carry = v >> 32u;
+    }
+    if (carry != 0)
+    {
+        x.push_back(static_cast<std::uint32_t>(carry));
+    }
+}
+
+void big_div_small(big& x, std::uint32_t d)
+{
+    std::uint64_t rest = 0;
+    for (std::size_t i = x.size(); i-- > 0;)
+    {
+        const std::uint64_t v = (rest << 32u) | x[i];
+        x[i] = static_cast<std::uint32_t>(v / d);
+        rest = v % d;
+    }
+    while (!x.empty() && x.back() == 0)
+    {
+        x.pop_back();
+    }
+}
+
+std::size_t big_bit_length(const big& x)
+{
+    std::size_t n = 32 * x.size();
+    for (std::uint32_t top = x.back(); (top & 0x80000000u) == 0; top <<= 1u)
+    {
+        --n;
+    }
+    return n;
+}
+
+bool big_bit(const big& x, std::size_t i)
+{
+    return ((x[i / 32] >> (i % 32)) & 1u) != 0;
+}
+
+/// the 128 most significant bits of x (floor), shifted left if x has fewer bits
+std::pair<std::uint64_t, std::uint64_t> big_top128(const big& x)
+{
+    const std::size_t n = big_bit_length(x);
+    std::uint64_t high = 0;
+    std::uint64_t low = 0;
+    for (std::size_t k = 0; k < 128; ++k)
+    {
+        const bool bit = k < n && big_bit(x, n - 1 - k);
+        if (k < 64)
+        {
+            high = (high << 1u) | (bit ? 1u : 0u);
+        }
+        else
+        {
+            low = (low << 1u) | (bit ? 1u : 0u);
+        }
+    }
+    return {high, low};
+}
+
+/// the digits (without trailing zeros) and the decimal exponent of a
+/// representation "[-]d[.ddd][e[+-]x]"
+std::pair<std::string, int> digits_and_exponent(const std::string& s)
+{
+    std::string digits;
+    int point = -1;
+    int exponent = 0;
+    for (std::size_t i = 0; i < s.size(); ++i)
+    {
+        const char c = s[i];
+        if (c >= '0' && c <= '9')
+        {
+            digits += c;
+        }
+        else if (c == '.')
+        {
+            point = static_cast<int>(digits.size());
+        }
+        else if (c == 'e' || c == 'E')
+        {
+            exponent = std::stoi(s.substr(i + 1));
+            break;
+        }
+    }
+    int e = exponent + (point < 0 ? static_cast<int>(digits.size()) : point) - static_cast<int>(digits.size());
+    const std::size_t first = digits.find_first_not_of('0');
+    digits = first == std::string::npos ? "0" : digits.substr(first);
+    while (digits.size() > 1 && digits.back() == '0')
+    {
+        digits.pop_back();
+        ++e;
+    }
+    return {digits, e};
+}
+
+/// whether the decimal digits * 10^e reads back as v
+bool reads_back(const std::string& digits, int e, double v)
+{
+    const std::string text = digits + "e" + std::to_string(e);
+    return std::strtod(text.c_str(), nullptr) == v;
+}
+
+/// Check the representation of a positive finite double: it reads back as
+/// the same value, and no representation with fewer digits does.
+void check_shortest(double v)
+{
+    std::array<char, 33> buf{};
+    char* end = nlohmann::detail::to_chars(buf.data(), buf.data() + 32, v);
+    const std::string text(buf.data(), end);
+    CAPTURE(text);
+    CHECK(std::strtod(text.c_str(), nullptr) == v);
+    // the layout is that of format_buffer() for the same digits
+    std::array<char, 64> reference{};
+    int len = 0;
+    int exponent = 0;
+    nlohmann::detail::dtoa_impl::shortest_digits(reference.data(), len, exponent, v);
+    const char* const reference_end = nlohmann::detail::dtoa_impl::format_buffer(reference.data(), len, exponent, -4, 15);
+    CHECK(text == std::string(reference.data(), static_cast<std::size_t>(reference_end - reference.data())));
+    const auto de = digits_and_exponent(text);
+    const std::string& digits = de.first;
+    if (digits.size() > 1)
+    {
+        // the decimals of one digit fewer next to the value
+        std::array<char, 64> shorter{};
+        const int n = std::snprintf(shorter.data(), shorter.size(), "%.*e", static_cast<int>(digits.size()) - 2, v); // NOLINT(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+        const auto near = digits_and_exponent(std::string(shorter.data(), static_cast<std::size_t>(n)));
+        // as an integer with digits.size() - 1 digits
+        std::string m = near.first;
+        int e = near.second;
+        while (m.size() < digits.size() - 1)
+        {
+            m += '0';
+            --e;
+        }
+        const std::uint64_t mid = std::stoull(m);
+        for (const std::uint64_t candidate :
+                {
+                    mid - 1, mid, mid + 1
+                })
+        {
+            CAPTURE(candidate);
+            CHECK(!reads_back(std::to_string(candidate), e, v));
+        }
+    }
+#if defined(JSON_HAS_CPP_17) && defined(__cpp_lib_to_chars)
+    // the closest of the shortest representations, as std::to_chars finds it
+    std::array<char, 64> std_text{};
+    const auto r = std::to_chars(std_text.data(), std_text.data() + std_text.size(), v, std::chars_format::scientific);
+    CHECK(digits_and_exponent(std::string(std_text.data(), r.ptr)) == de);
+#endif
+}
+} // namespace
+
+TEST_CASE("shortest digits of doubles")
+{
+    SECTION("powers of ten")
+    {
+        // the 128-bit significands of 10^k, rounded down, recomputed
+        for (int k = -342; k <= 341; ++k)
+        {
+            CAPTURE(k);
+            big x{1};
+            if (k >= 0)
+            {
+                for (int i = 0; i < k; ++i)
+                {
+                    big_mul_small(x, 10);
+                }
+            }
+            else
+            {
+                // floor(2^b / 10^-k) for a b that leaves more than 128 bits
+                const int b = 128 + 64 + (4 * -k);
+                x.assign(static_cast<std::size_t>(b / 32) + 1, 0);
+                x.back() = 1u << (b % 32);
+                for (int i = 0; i < -k; ++i)
+                {
+                    big_div_small(x, 10);
+                }
+            }
+            const auto expected = big_top128(x);
+            const auto actual = nlohmann::detail::zmij::pow10(k);
+            CHECK(actual.high == expected.first);
+            CHECK(actual.low == expected.second);
+        }
+    }
+
+    SECTION("boundary values")
+    {
+        for (const double v :
+                {
+                    std::numeric_limits<double>::min(), std::numeric_limits<double>::max(), std::numeric_limits<double>::denorm_min(),
+                    std::nextafter(std::numeric_limits<double>::min(), 0.0), 1.0, 2.0, 0.1, 0.3, 1e21, 1e22, 1e23, 5e-324, 9007199254740993.0,
+                    1.2345e+21, 2.2250738585072014e-308, 1.7976931348623157e308, 4.9406564584124654e-324, 123456789012345680.0
+                })
+        {
+            check_shortest(v);
+        }
+        // all powers of two (their rounding interval is narrower below)
+        for (int e = -1074; e <= 1023; ++e)
+        {
+            check_shortest(std::ldexp(1.0, e));
+        }
+        // powers of ten and their neighbors
+        for (int e = -323; e <= 308; ++e)
+        {
+            const double p = std::strtod(("1e" + std::to_string(e)).c_str(), nullptr);
+            check_shortest(p);
+            check_shortest(std::nextafter(p, 0.0));
+            check_shortest(std::nextafter(p, std::numeric_limits<double>::infinity()));
+        }
+    }
+
+    SECTION("random doubles")
+    {
+        std::mt19937_64 rng(5295); // NOLINT(cert-msc32-c,cert-msc51-cpp,bugprone-random-generator-seed): reproducible
+        for (int i = 0; i < 100000; ++i)
+        {
+            const std::uint64_t bits = rng() & 0x7FFFFFFFFFFFFFFFu;
+            const auto v = reinterpret_bits<double>(bits);
+            if (std::isfinite(v) && v != 0)
+            {
+                check_shortest(v);
+            }
+        }
+    }
+}
Author	SHA1	Message	Date
Niels Lohmann	a04aa3d095	Write doubles with the shortest digits (Zmij) dump() writes doubles with the conversion of Zmij by Victor Zverovich (https://github.com/vitaut/zmij, MIT), ported to C++11 in detail/conversions/zmij.hpp: the shortest decimal in the rounding interval, the closest one if there are several. Grisu2 does not always find the shortest digits; about 0.14% of random doubles are now written differently (0.08% with fewer digits, 0.06% with the closest last digit); short decimals such as 0.1 or 2555.56 are not affected. float keeps Grisu2. The layout of doubles is unchanged, but written differently: the digits are converted eight at a time (the BCD conversion of Xiang JunBo, as in Zmij) and stored with one byte swap per eight digits; leading and trailing zeros are counted from those bytes; and the layouts of format_buffer() are written with fixed-size moves instead of per-digit loops and moves of the buffer (to_chars() uses a local buffer if the caller's is shorter than the 41 bytes this may write). The powers of ten come from the table for number parsing, adjusted where it holds them rounded up, and from the compressed tables of Zmij beyond 10^308. json::dump() gets faster on floats: canada -53%, numbers -46%, mesh -37%, marine_ik -30%. Tests: the powers of ten recomputed with a small big-integer; for random doubles, all powers of two and of ten and their neighbors, and boundary values: the output reads back as the same value, no decimal with one digit fewer does, the layout equals that of format_buffer() for the same digits, and (C++17) the digits equal those of std::to_chars. The size ratios of canada.json in unit-binary_formats.cpp and one expectation in unit-to_chars.cpp change with the shorter output. Signed-off-by: Niels Lohmann <mail@nlohmann.me>	2026-09-29 14:03:25 +02:00
Niels Lohmann	cbdc502fbf	Write compact dumps of json_view without a library call per token The default dump() (no indentation, no ensure_ascii) gets its own writer: the same walk and output, with the write position in a local variable (stores through char pointers would otherwise force a reload of the buffer's members after each one), strings and number tokens of the source copied by fixed-size moves of 32 bytes where the source has that many bytes left (the buffer keeps 64 bytes of slack), and decoded strings copied in runs up to the next quote, backslash, or control character. Documents that are not edited are walked through the node array in order, so that a frame only needs the end of its container, and integer tokens are read from the source directly. The innermost open container is kept in local variables, and the stack holds only the ones around it; the stack starts in a local array of 32 and moves to the heap only for deeper nesting (its address does not escape, so its pointers stay in registers). Dumps of shallow documents thus allocate only the output, whose first size includes the slack, so it does not grow just before the end. The long copies are out of line: otherwise, the compiler merges the fixed-size moves into the same library call. These techniques come from the prototype; the writer lost them when the view was split into pull requests, which made dump() 2 to 3 times slower. Signed-off-by: Niels Lohmann <mail@nlohmann.me>	2026-09-29 14:03:24 +02:00
Niels Lohmann	300ab011b0	Document how images store the node index of json_view The node index section on the architecture page says how save() writes the nodes and that a change of their layout must raise the image version, and save's format note links to it. Signed-off-by: Niels Lohmann <mail@nlohmann.me>	2026-09-29 14:03:23 +02:00
Niels Lohmann	8d9088ab05	Address the cpplint findings of json_document images The exponent of the overflow check is an std::int64_t instead of a long (runtime/int). Signed-off-by: Niels Lohmann <mail@nlohmann.me>	2026-09-29 14:03:22 +02:00
Niels Lohmann	5a936a2ef3	Document the images of json_documents Add API pages for basic_json_document::save() and load(), including the image_check enumeration and its three levels. Add examples that show caching a parsed document as an image, loading it without parsing, the difference between a borrowed and an owned image, and a full check rejecting a damaged image that a bounds check still reads safely. Add an "Images" section to the json_view feature page, register the new pages in mkdocs.yml and docSet.sql, group basic_json_document's member list by parsing/access/images/edits, and document parse_error.116 and type_error.320 on the exceptions page, extending out_of_range.416 for images. Signed-off-by: Niels Lohmann <mail@nlohmann.me>	2026-09-29 14:03:21 +02:00
Niels Lohmann	5876a22712	Add images of json_documents: save() and load() An image is a document stored so that loading it needs no parsing: a 64-byte header, the nodes, the text, and the decoded strings (little-endian; version 1). - save() writes an edited document in its current state, in document order (floats that are not finite become null, as in dump()); the same document always gives the same bytes - load(pointer, size) and load(const vector&) borrow the image; load(vector&&) keeps it without a copy. The nodes are copied (aligned, and editable); the hash indexes of large objects are rebuilt. - image_check::full checks everything the parser guarantees (structure, bounds, UTF-8, strings of the source, number tokens and their values); bounds checks structure and bounds, so that reading and serializing stay safe; none trusts the image. A malformed image or a failed check throws the new parse_error.116; saving a discarded document (or images on a big-endian target) throws the new type_error.320; images of 4 GiB or more out_of_range.416. As images checked for bounds only can hold any bytes, the general float conversion now checks the token's grammar (and locates the point and the exponent itself), the exponent loop of the layout conversion takes digits as unsigned, and the serializer validates each non-ASCII sequence it decodes, throwing what basic_json::dump() throws for invalid UTF-8. Parsed and edited documents are not affected. The idea of images comes from zero-copy formats such as FlatBuffers and YaFF, the check from FlatBuffers' Verifier; no code is taken from them. Tests: round trips with every check (small documents, test files, large objects, edited documents with every kind of edit), ownership, all errors, one corruption per rejection branch of the check, and 12,000 seeded random corruptions, which must be rejected or read safely. The fuzzer json_view_image_fuzzer uses each input as an image and as a JSON text. Signed-off-by: Niels Lohmann <mail@nlohmann.me>	2026-09-29 14:03:20 +02:00