TinyEXIF 1.2.0

Added

• MaxApertureValue (EXIF 0x9205), the widest aperture of the lens, as an
  f-number like ApertureValue, with FIELD_ID_MaxApertureValue.
• DJI XMP flight speed (drone-dji:FlightXSpeed/YSpeed/ZSpeed) now fills
  GeoLocation.SpeedX/Y/Z when the DJI MakerNote has not, as on cameras that
  write no MakerNote speed (e.g. the Osmo 360). The MakerNote, being binary,
  still takes precedence.
• Multi-Picture Format (MPF, CIPA DC-007) index: MPImages lists the images a
  JPEG stores after its own, such as the preview of a panorama, the second view
  of a stereo (.MPO) image or an HDR gain map, with the type, flags, offset
  and length of each (FIELD_ID_MPImages). They are only located, never read;
  the README shows how to read one safely and parse it in turn. To reach the
  index, the scan no longer stops as soon as it has EXIF and XMP: it goes on to
  the image data, skipping any further EXIF or XMP segment as before.
• The README now documents every feature. That covers the input sources and
  return codes, and a reference of every field with the EXIF tag or XMP
  property it comes from. It also covers the presence API, MPF, the guarantees
  for untrusted files, building with CMake or vcpkg, and testing. The header
  comments of ShutterSpeedValue, an exposure time in seconds, and of
  LensInfo.FStopMin/FStopMax, the widest aperture at each end of the zoom
  range, described other values.

Fixed

• Shared-library ABI identity now changes with the new EXIFInfo layout: ABI
  version 2 on Linux/macOS, and TinyEXIF-2.dll on Windows. Consumers must
  rebuild with the 1.2.0 header; old binaries must not load the new library.
• XMP flight speed fills each missing component individually, preserving a
  finite DJI MakerNote speed even when another component is absent or invalid.
• MPF entry byte counts must describe complete 16-byte entries; a truncated
  count is rejected rather than rounded down to an apparently valid index.
• The release workflow now runs the sample corpus for the tagged build before
  publishing the release.
• BUILD_FUZZER=ON now links the coverage runtime for the demo and shared
  library too, so the documented build works with the demo enabled.
• A GPS receiver without a fix no longer yields a position. Cameras write
  the position tags even then, usually all zeros, which placed the image at
  0°N 0°E. With GPSStatus V (measurement void) latitude, longitude and
  altitude are now left absent, as is the DJI XMP AbsoluteAltitude when
  drone-dji:GpsStatus is Invalid.
• A latitude, longitude or altitude of 0 with a south, west or below-sea-level
  reference read back as IEEE 754 -0, which prints as "-0".
• A DJI MakerNote was skipped entirely, speeds and camera angles included, in
  files whose EXIF was rewritten in Motorola byte order: DJI writes the
  MakerNote little-endian and editors copy it unchanged, so its byte order is
  now detected rather than taken from the TIFF header.
• Big-endian CPUs could not parse EXIF at all. The "Big-endian CPU
  support" of 1.0.3 (#14) combined the TIFF byte-order marker with the host's
  byte order, but parse16()/parse32() assemble values byte by byte and were
  already host-independent, so on a big-endian host both II and MM were
  inverted: II files failed with PARSE_CORRUPT_DATA and MM files decoded
  corrupted values. The marker now selects the file's byte order alone.
  Little-endian hosts are unaffected. A new CI job runs the full corpus on
  s390x under QEMU. Reported by @cinema-ONE. (#29)
• gcc 16 -Wmaybe-uninitialized warnings on the six fetch-into-temporary
  sites. Rather than initializing each temporary, they are gone: the new
  EntryParser::FetchAs<T>() fetches a value stored as one type into a member
  of another, and also replaces FetchFloat(). Builds with
  TINYEXIF_NO_XMP_SUPPORT no longer warn about the XMP-only string helpers.
  A new CI job compiles with gcc 16 -Wall -Wextra -Werror at every
  optimization level, with and without XMP. Reported by @heitbaum. (#30)
• XMP tiff:XResolution / tiff:YResolution are rationals ("300/1"), but
  were read with a parser that stopped at the slash, so "144/2" came back as
  144 instead of 72. They now go through the same rational parser as the other
  XMP numbers.
• XMP numbers followed by other text were read as the number they start with:
  "12junk" as 12 and the rational "1e/2" as 1/2. The tiff: integers, read
  by tinyxml2, also turned "-1" into 4294967295 and truncated an orientation
  beyond 16 bits. Every XMP number, the tiff: ones included, must now be the
  whole text, surrounding whitespace aside, or it is left absent.
• Parsing from memory, parseFrom(data, length), refused a read ending exactly
  at the end of the buffer, so a file whose last segment runs to its end parsed
  differently than from a stream. The check also formed a pointer past the end
  of the buffer before comparing it, which is undefined behavior. Found by
  fuzzing the new MPF parser.
• TestSamples.py --update aborted on the corpus's own error samples, mistaking
  the demo's 253..255 exit codes for signal deaths.

Security

Three privately reported advisories, all low severity: none is a memory-safety
issue, but each lets a crafted file put an invalid value into a parsed field.
Every floating point field is now guaranteed to be finite: a value that is not
leaves the field absent, so HasField() and the hasXxx() accessors report it
as such.

• Undefined behavior converting an out-of-range ExposureIndex to
  ISOSpeedRatings: a negative index now leaves ISOSpeedRatings absent and
  one above 65535 is clamped to 65535, as EXIF records ISO. The sanitizer job
  now enables float-cast-overflow, which gcc leaves out of
  -fsanitize=undefined. Reported by @ruben0315R, crediting @Magashwarahan
  for the variant discovery. (GHSA-jqfq-fwfw-xc8x)
• XMP rationals with a zero denominator ("1/0", "0/0") divided to inf or
  NaN, which read back as present through the DBL_MAX sentinels and pass any
  range check, as every comparison with NaN is false. The same values also
  came from the "nan", "inf" and out-of-range literals that strtod and
  tinyxml2 accept, including in DewarpData and tiff:X/YResolution; every
  XMP number now goes through one parser that rejects all of them. Reported by
  Ruben Arumugam Gunavathy (@ruben0315R). (GHSA-rvfv-gvcj-m5f5)
• The APEX conversions of ShutterSpeedValue and ApertureValue overflowed to
  +inf for values far outside any camera's range, e.g. a shutter speed of
  INT32_MIN. Found in the same audit: a DJI MakerNote FLOAT is raw IEEE 754
  bits and could carry NaN or inf straight into the GeoLocation speed and
  orientation fields. Reported by Ruben Arumugam Gunavathy (@ruben0315R).
  (GHSA-pgrm-rwjh-553p)