exiftool_rs/

exiftool.rs

//! Core ExifTool struct and public API.
//!
//! This is the main entry point for reading metadata from files.
//! Mirrors ExifTool.pm's ImageInfo/ExtractInfo/GetInfo pipeline.

use std::collections::HashMap;
use std::fs;
use std::path::Path;

use crate::error::{Error, Result};
use crate::file_type::{self, FileType};
use crate::formats;
use crate::metadata::exif::ByteOrderMark;
use crate::tag::{Tag, TagGroup, MAIN_DOCUMENT};
use crate::value::Value;
use crate::writer::{
    exif_writer, iptc_writer, jpeg_writer, matroska_writer, mp4_writer, pdf_writer, png_writer,
    psd_writer, tiff_writer, webp_writer, xmp_writer,
};

/// Processing options for metadata extraction.
#[derive(Debug, Clone)]
pub struct Options {
    /// Include duplicate tags (different groups may have same tag name).
    pub duplicates: bool,
    /// Apply print conversions (human-readable values).
    pub print_conv: bool,
    /// Fast scan level: 0=normal, 1=skip composite, 2=skip maker notes, 3=skip thumbnails.
    pub fast_scan: u8,
    /// Only extract these tag names (empty = all).
    pub requested_tags: Vec<String>,
    /// Extract embedded documents/data (video frames, etc.). Level: 0=off, 1=-ee, 2=-ee2, 3=-ee3.
    pub extract_embedded: u8,
    /// Show unknown tags: 0=off, 1=-u (show unknown), 2=-U (show unknown + binary data).
    pub show_unknown: u8,
    /// Process compressed data in files (-z option).
    pub process_compressed: bool,
    /// Use MWG (Metadata Working Group) composite tags for reading/writing.
    pub use_mwg: bool,
    /// Reverse-geocode `Geolocation*` tags from GPS coordinates
    /// (ExifTool's `Geolocation` API option). Off by default, like ExifTool.
    pub geolocation: bool,
}

impl Default for Options {
    fn default() -> Self {
        Self {
            duplicates: false,
            print_conv: true,
            fast_scan: 0,
            requested_tags: Vec::new(),
            extract_embedded: 0,
            show_unknown: 0,
            process_compressed: false,
            use_mwg: false,
            geolocation: false,
        }
    }
}

/// The main ExifTool struct. Create one and use it to extract metadata from files.
///
/// # Example
/// ```no_run
/// use exiftool_rs::ExifTool;
///
/// let mut et = ExifTool::new();
/// let info = et.image_info("photo.jpg").unwrap();
/// for (name, value) in &info {
///     println!("{}: {}", name, value);
/// }
/// ```
/// A queued tag change for writing.
#[derive(Debug, Clone)]
pub struct NewValue {
    /// Tag name (e.g., "Artist", "Copyright", "XMP:Title")
    pub tag: String,
    /// Group prefix if specified (e.g., "EXIF", "XMP", "IPTC")
    pub group: Option<String>,
    /// New value (None = delete tag)
    pub value: Option<String>,
}

/// The main ExifTool engine — read, write, and edit metadata.
///
/// # Reading metadata
/// ```no_run
/// use exiftool_rs::ExifTool;
///
/// let et = ExifTool::new();
///
/// // Full tag structs
/// let tags = et.extract_info("photo.jpg").unwrap();
/// for tag in &tags {
///     println!("[{}] {}: {}", tag.group.family0, tag.name, tag.print_value);
/// }
///
/// // Simple name→value map
/// let info = et.image_info("photo.jpg").unwrap();
/// println!("Camera: {}", info.get("Model").unwrap_or(&String::new()));
/// ```
///
/// # Writing metadata
/// ```no_run
/// use exiftool_rs::ExifTool;
///
/// let mut et = ExifTool::new();
/// et.set_new_value("Artist", Some("John Doe"));
/// et.set_new_value("Copyright", Some("2024"));
/// et.write_info("input.jpg", "output.jpg").unwrap();
/// ```
pub struct ExifTool {
    options: Options,
    new_values: Vec<NewValue>,
}

/// Result of metadata extraction: maps tag names to display values.
pub type ImageInfo = HashMap<String, String>;

impl ExifTool {
    /// Create a new ExifTool instance with default options.
    pub fn new() -> Self {
        Self {
            options: Options::default(),
            new_values: Vec::new(),
        }
    }

    /// Create a new ExifTool instance with custom options.
    pub fn with_options(options: Options) -> Self {
        Self {
            options,
            new_values: Vec::new(),
        }
    }

    /// Get a mutable reference to the options.
    pub fn options_mut(&mut self) -> &mut Options {
        &mut self.options
    }

    /// Get a reference to the options.
    pub fn options(&self) -> &Options {
        &self.options
    }

    // ================================================================
    // Writing API
    // ================================================================

    /// Queue a new tag value for writing.
    ///
    /// Call this one or more times, then call `write_info()` to apply changes.
    ///
    /// # Arguments
    /// * `tag` - Tag name, optionally prefixed with group (e.g., "Artist", "XMP:Title", "EXIF:Copyright")
    /// * `value` - New value, or None to delete the tag
    ///
    /// # Example
    /// ```no_run
    /// use exiftool_rs::ExifTool;
    /// let mut et = ExifTool::new();
    /// et.set_new_value("Artist", Some("John Doe"));
    /// et.set_new_value("Copyright", Some("2024 John Doe"));
    /// et.set_new_value("XMP:Title", Some("My Photo"));
    /// et.write_info("photo.jpg", "photo_out.jpg").unwrap();
    /// ```
    pub fn set_new_value(&mut self, tag: &str, value: Option<&str>) {
        let (group, tag_name) = if let Some(colon_pos) = tag.find(':') {
            (
                Some(tag[..colon_pos].to_string()),
                tag[colon_pos + 1..].to_string(),
            )
        } else {
            (None, tag.to_string())
        };

        self.new_values.push(NewValue {
            tag: tag_name,
            group,
            value: value.map(|v| v.to_string()),
        });
    }

    /// Clear all queued new values.
    pub fn clear_new_values(&mut self) {
        self.new_values.clear();
    }

    /// Copy tags from a source file, queuing them as new values.
    ///
    /// Reads all tags from `src_path` and queues them for writing.
    /// Optionally filter by tag names.
    pub fn set_new_values_from_file<P: AsRef<Path>>(
        &mut self,
        src_path: P,
        tags_to_copy: Option<&[&str]>,
    ) -> Result<u32> {
        let src_tags = self.extract_info(src_path)?;
        let mut count = 0u32;

        for tag in &src_tags {
            // Skip file-level tags that shouldn't be copied
            if tag.group.family0 == "File" || tag.group.family0 == "Composite" {
                continue;
            }
            // Skip binary/undefined data and empty values
            if tag.print_value.starts_with("(Binary") || tag.print_value.starts_with("(Undefined") {
                continue;
            }
            if tag.print_value.is_empty() {
                continue;
            }

            // Filter by requested tags
            if let Some(filter) = tags_to_copy {
                let name_lower = tag.name.to_lowercase();
                if !filter.iter().any(|f| f.to_lowercase() == name_lower) {
                    continue;
                }
            }

            let _full_tag = format!("{}:{}", tag.group.family0, tag.name);
            self.new_values.push(NewValue {
                tag: tag.name.clone(),
                group: Some(tag.group.family0.clone()),
                value: Some(tag.print_value.clone()),
            });
            count += 1;
        }

        Ok(count)
    }

    /// Set a file's name based on a tag value.
    pub fn set_file_name_from_tag<P: AsRef<Path>>(
        &self,
        path: P,
        tag_name: &str,
        template: &str,
    ) -> Result<String> {
        let path = path.as_ref();
        let tags = self.extract_info(path)?;

        let tag_value = tags
            .iter()
            .find(|t| t.name.to_lowercase() == tag_name.to_lowercase())
            .map(|t| &t.print_value)
            .ok_or_else(|| Error::TagNotFound(tag_name.to_string()))?;

        // Build new filename from template
        // Template: "prefix%value%suffix.ext" or just use the tag value
        let new_name = if template.contains('%') {
            template.replace("%v", value_to_filename(tag_value).as_str())
        } else {
            // Default: use tag value as filename, keep extension
            let ext = path.extension().and_then(|e| e.to_str()).unwrap_or("");
            let clean = value_to_filename(tag_value);
            if ext.is_empty() {
                clean
            } else {
                format!("{}.{}", clean, ext)
            }
        };

        let parent = path.parent().unwrap_or(Path::new(""));
        let new_path = parent.join(&new_name);

        fs::rename(path, &new_path).map_err(Error::Io)?;
        Ok(new_path.to_string_lossy().to_string())
    }

    /// Write queued changes to a file.
    ///
    /// If `dst_path` is the same as `src_path`, the file is modified in-place
    /// (via a temporary file).
    pub fn write_info<P: AsRef<Path>, Q: AsRef<Path>>(
        &self,
        src_path: P,
        dst_path: Q,
    ) -> Result<u32> {
        let src_path = src_path.as_ref();
        let dst_path = dst_path.as_ref();
        let data = fs::read(src_path).map_err(Error::Io)?;

        let file_type = self.detect_file_type(&data, src_path)?;
        let output = self.apply_changes(&data, file_type)?;

        // Write to temp file first, then rename (atomic)
        let temp_path = dst_path.with_extension("exiftool_tmp");
        fs::write(&temp_path, &output).map_err(Error::Io)?;
        fs::rename(&temp_path, dst_path).map_err(Error::Io)?;

        Ok(self.new_values.len() as u32)
    }

    /// Apply queued changes to in-memory data.
    fn apply_changes(&self, data: &[u8], file_type: FileType) -> Result<Vec<u8>> {
        match file_type {
            FileType::Jpeg => self.write_jpeg(data),
            FileType::Png => self.write_png(data),
            FileType::Tiff
            | FileType::Dng
            | FileType::Cr2
            | FileType::Nef
            | FileType::Arw
            | FileType::Orf
            | FileType::Pef => self.write_tiff(data),
            FileType::WebP => self.write_webp(data),
            FileType::Mp4
            | FileType::QuickTime
            | FileType::M4a
            | FileType::ThreeGP
            | FileType::F4v => self.write_mp4(data),
            FileType::Psd => self.write_psd(data),
            FileType::Pdf => self.write_pdf(data),
            FileType::Heif | FileType::Avif => self.write_mp4(data),
            FileType::Mkv | FileType::WebM => self.write_matroska(data),
            FileType::Gif => {
                let comment = self
                    .new_values
                    .iter()
                    .find(|nv| nv.tag.to_lowercase() == "comment")
                    .and_then(|nv| nv.value.clone());
                crate::writer::gif_writer::write_gif(data, comment.as_deref())
            }
            FileType::Flac => {
                let changes: Vec<(&str, &str)> = self
                    .new_values
                    .iter()
                    .filter_map(|nv| Some((nv.tag.as_str(), nv.value.as_deref()?)))
                    .collect();
                crate::writer::flac_writer::write_flac(data, &changes)
            }
            FileType::Mp3 | FileType::Aiff => {
                let changes: Vec<(&str, &str)> = self
                    .new_values
                    .iter()
                    .filter_map(|nv| Some((nv.tag.as_str(), nv.value.as_deref()?)))
                    .collect();
                crate::writer::id3_writer::write_id3(data, &changes)
            }
            FileType::Jp2 | FileType::Jxl => {
                let new_xmp = if self
                    .new_values
                    .iter()
                    .any(|nv| nv.group.as_deref() == Some("XMP"))
                {
                    let refs: Vec<&NewValue> = self
                        .new_values
                        .iter()
                        .filter(|nv| nv.group.as_deref() == Some("XMP"))
                        .collect();
                    Some(self.build_new_xmp(&refs))
                } else {
                    None
                };
                crate::writer::jp2_writer::write_jp2(data, new_xmp.as_deref(), None)
            }
            FileType::PostScript => {
                let changes: Vec<(&str, &str)> = self
                    .new_values
                    .iter()
                    .filter_map(|nv| Some((nv.tag.as_str(), nv.value.as_deref()?)))
                    .collect();
                crate::writer::ps_writer::write_postscript(data, &changes)
            }
            FileType::Ogg | FileType::Opus => {
                let changes: Vec<(&str, &str)> = self
                    .new_values
                    .iter()
                    .filter_map(|nv| Some((nv.tag.as_str(), nv.value.as_deref()?)))
                    .collect();
                crate::writer::ogg_writer::write_ogg(data, &changes)
            }
            FileType::Xmp => {
                let props: Vec<xmp_writer::XmpProperty> = self
                    .new_values
                    .iter()
                    .filter_map(|nv| {
                        let val = nv.value.as_deref()?;
                        Some(xmp_writer::XmpProperty {
                            namespace: nv.group.clone().unwrap_or_else(|| "dc".into()),
                            property: nv.tag.clone(),
                            values: vec![val.to_string()],
                            prop_type: xmp_writer::XmpPropertyType::Simple,
                        })
                    })
                    .collect();
                Ok(crate::writer::xmp_sidecar_writer::write_xmp_sidecar(&props))
            }
            _ => Err(Error::UnsupportedFileType(format!(
                "writing not yet supported for {}",
                file_type
            ))),
        }
    }

    /// Returns the set of tag names (lowercase) that are writable for a given file type.
    /// Returns `None` if any tag is writable (open-ended formats like PNG, FLAC, MKV).
    /// Returns `Some(empty set)` if the format has no writer.
    pub fn writable_tags(file_type: FileType) -> Option<std::collections::HashSet<&'static str>> {
        use std::collections::HashSet;

        // EXIF tags supported by exif_writer
        const EXIF_TAGS: &[&str] = &[
            "imagedescription",
            "make",
            "model",
            "orientation",
            "xresolution",
            "yresolution",
            "resolutionunit",
            "software",
            "modifydate",
            "datetime",
            "artist",
            "copyright",
            "datetimeoriginal",
            "createdate",
            "datetimedigitized",
            "usercomment",
            "imageuniqueid",
            "ownername",
            "cameraownername",
            "serialnumber",
            "bodyserialnumber",
            "lensmake",
            "lensmodel",
            "lensserialnumber",
        ];

        // IPTC tags supported by iptc_writer
        const IPTC_TAGS: &[&str] = &[
            "objectname",
            "title",
            "urgency",
            "category",
            "supplementalcategories",
            "keywords",
            "specialinstructions",
            "datecreated",
            "timecreated",
            "by-line",
            "author",
            "byline",
            "by-linetitle",
            "authorsposition",
            "bylinetitle",
            "city",
            "sub-location",
            "sublocation",
            "province-state",
            "state",
            "provincestate",
            "country-primarylocationcode",
            "countrycode",
            "country-primarylocationname",
            "country",
            "headline",
            "credit",
            "source",
            "copyrightnotice",
            "contact",
            "caption-abstract",
            "caption",
            "description",
            "writer-editor",
            "captionwriter",
        ];

        // XMP auto-detected tags (no group prefix needed)
        const XMP_AUTO_TAGS: &[&str] = &[
            "title",
            "description",
            "subject",
            "creator",
            "rights",
            "keywords",
            "rating",
            "label",
            "hierarchicalsubject",
        ];

        // ID3 tags
        const ID3_TAGS: &[&str] = &[
            "title",
            "artist",
            "album",
            "year",
            "date",
            "track",
            "genre",
            "comment",
            "composer",
            "albumartist",
            "encoder",
            "encodedby",
            "publisher",
            "copyright",
            "bpm",
            "lyrics",
        ];

        // MP4/MOV ilst tags
        const MP4_TAGS: &[&str] = &[
            "title",
            "artist",
            "album",
            "year",
            "date",
            "comment",
            "genre",
            "composer",
            "writer",
            "encoder",
            "encodedby",
            "grouping",
            "lyrics",
            "description",
            "albumartist",
            "copyright",
        ];

        // PDF Info dict tags
        const PDF_TAGS: &[&str] = &[
            "title", "author", "subject", "keywords", "creator", "producer",
        ];

        // PostScript DSC tags
        const PS_TAGS: &[&str] = &[
            "title",
            "creator",
            "author",
            "for",
            "creationdate",
            "createdate",
        ];

        match file_type {
            // Open-ended: any tag name accepted
            FileType::Png
            | FileType::Flac
            | FileType::Mkv
            | FileType::WebM
            | FileType::Ogg
            | FileType::Opus
            | FileType::Xmp => None,

            // JPEG: EXIF + IPTC + XMP auto + comment
            FileType::Jpeg => {
                let mut set: HashSet<&str> = HashSet::new();
                set.extend(EXIF_TAGS);
                set.extend(IPTC_TAGS);
                set.extend(XMP_AUTO_TAGS);
                set.insert("comment");
                Some(set)
            }

            // TIFF-based: EXIF only
            FileType::Tiff
            | FileType::Dng
            | FileType::Cr2
            | FileType::Nef
            | FileType::Arw
            | FileType::Orf
            | FileType::Pef => {
                let mut set: HashSet<&str> = HashSet::new();
                set.extend(EXIF_TAGS);
                Some(set)
            }

            // WebP: EXIF + XMP auto
            FileType::WebP => {
                let mut set: HashSet<&str> = HashSet::new();
                set.extend(EXIF_TAGS);
                set.extend(XMP_AUTO_TAGS);
                Some(set)
            }

            // MP4/MOV/HEIF: ilst + XMP auto
            FileType::Mp4
            | FileType::QuickTime
            | FileType::M4a
            | FileType::ThreeGP
            | FileType::F4v
            | FileType::Heif
            | FileType::Avif => {
                let mut set: HashSet<&str> = HashSet::new();
                set.extend(MP4_TAGS);
                set.extend(XMP_AUTO_TAGS);
                Some(set)
            }

            // PSD: IPTC + XMP auto
            FileType::Psd => {
                let mut set: HashSet<&str> = HashSet::new();
                set.extend(IPTC_TAGS);
                set.extend(XMP_AUTO_TAGS);
                Some(set)
            }

            FileType::Pdf => Some(PDF_TAGS.iter().copied().collect()),
            FileType::PostScript => Some(PS_TAGS.iter().copied().collect()),

            FileType::Mp3 | FileType::Aiff => Some(ID3_TAGS.iter().copied().collect()),

            FileType::Gif => {
                let mut set: HashSet<&str> = HashSet::new();
                set.insert("comment");
                Some(set)
            }

            // JP2/JXL: XMP only (with group prefix)
            FileType::Jp2 | FileType::Jxl => Some(XMP_AUTO_TAGS.iter().copied().collect()),

            // No writer
            _ => Some(HashSet::new()),
        }
    }

    /// Write metadata changes to JPEG data.
    fn write_jpeg(&self, data: &[u8]) -> Result<Vec<u8>> {
        // Classify new values by target group
        let mut exif_values: Vec<&NewValue> = Vec::new();
        let mut xmp_values: Vec<&NewValue> = Vec::new();
        let mut iptc_values: Vec<&NewValue> = Vec::new();
        let mut comment_value: Option<&str> = None;
        let mut remove_exif = false;
        let mut remove_xmp = false;
        let mut remove_iptc = false;
        let mut remove_comment = false;

        for nv in &self.new_values {
            let group = nv.group.as_deref().unwrap_or("");
            let group_upper = group.to_uppercase();

            // Check for group deletion
            if nv.value.is_none() && nv.tag == "*" {
                match group_upper.as_str() {
                    "EXIF" => {
                        remove_exif = true;
                        continue;
                    }
                    "XMP" => {
                        remove_xmp = true;
                        continue;
                    }
                    "IPTC" => {
                        remove_iptc = true;
                        continue;
                    }
                    _ => {}
                }
            }

            match group_upper.as_str() {
                "XMP" => xmp_values.push(nv),
                "IPTC" => iptc_values.push(nv),
                "EXIF" | "IFD0" | "EXIFIFD" | "GPS" => exif_values.push(nv),
                "" => {
                    // Auto-detect best group based on tag name
                    if nv.tag.to_lowercase() == "comment" {
                        if nv.value.is_none() {
                            remove_comment = true;
                        } else {
                            comment_value = nv.value.as_deref();
                        }
                    } else if is_xmp_tag(&nv.tag) {
                        xmp_values.push(nv);
                    } else {
                        exif_values.push(nv);
                    }
                }
                _ => exif_values.push(nv), // default to EXIF
            }
        }

        // Build new EXIF data
        let new_exif = if !exif_values.is_empty() {
            Some(self.build_new_exif(data, &exif_values)?)
        } else {
            None
        };

        // Build new XMP data
        let new_xmp = if !xmp_values.is_empty() {
            Some(self.build_new_xmp(&xmp_values))
        } else {
            None
        };

        // Build new IPTC data by merging changes into the file's existing
        // IPTC (issue #7), so writing one dataset doesn't drop the rest.
        let new_iptc_data = if iptc_values.is_empty() {
            None
        } else {
            let existing = jpeg_writer::extract_jpeg_iptc_iim(data);
            self.build_new_iptc(existing.as_deref(), &iptc_values)
        };

        // Rewrite JPEG
        jpeg_writer::write_jpeg(
            data,
            new_exif.as_deref(),
            new_xmp.as_deref(),
            new_iptc_data.as_deref(),
            comment_value,
            remove_exif,
            remove_xmp,
            remove_iptc,
            remove_comment,
        )
    }

    /// Build the IPTC-IIM block by merging queued changes into the file's
    /// existing IPTC instead of replacing it (issue #7). Datasets not being
    /// changed are preserved (including `CodedCharacterSet`); a change updates
    /// its dataset, and a `None` value deletes it. String values are encoded
    /// in the block's charset — Latin-1 by default, UTF-8 if the existing IPTC
    /// declares `CodedCharacterSet=UTF8`.
    fn build_new_iptc(&self, existing: Option<&[u8]>, values: &[&NewValue]) -> Option<Vec<u8>> {
        let mut records = existing.map(iptc_writer::parse_iim).unwrap_or_default();
        // ESC % G (1B 25 47) in the CodedCharacterSet dataset (1:90) → UTF-8.
        let utf8 = records.iter().any(|r| {
            r.record == 1 && r.dataset == 90 && r.data.windows(3).any(|w| w == [0x1B, 0x25, 0x47])
        });
        for nv in values {
            let Some((record, dataset)) = iptc_writer::tag_name_to_iptc(&nv.tag) else {
                continue;
            };
            match nv.value.as_deref() {
                // Set: update the dataset in place (ExifTool preserves the
                // file's original dataset order — it does not re-sort), drop
                // any duplicates, or append if the tag is new.
                Some(value) => {
                    let data = if utf8 {
                        value.as_bytes().to_vec()
                    } else {
                        crate::encoding::encode_latin1(value)
                    };
                    let mut updated = false;
                    records.retain_mut(|r| {
                        if r.record == record && r.dataset == dataset {
                            if updated {
                                return false; // collapse repeated datasets to one
                            }
                            r.data = data.clone();
                            updated = true;
                        }
                        true
                    });
                    if !updated {
                        records.push(iptc_writer::IptcRecord {
                            record,
                            dataset,
                            data,
                        });
                    }
                }
                // Delete: remove all datasets for this tag.
                None => records.retain(|r| !(r.record == record && r.dataset == dataset)),
            }
        }
        if records.is_empty() {
            return None;
        }
        Some(iptc_writer::build_iptc(&records))
    }

    /// Build new EXIF data by merging existing EXIF with queued changes.
    fn build_new_exif(&self, jpeg_data: &[u8], values: &[&NewValue]) -> Result<Vec<u8>> {
        let bo = ByteOrderMark::BigEndian;
        let mut ifd0_entries = Vec::new();
        let mut exif_entries = Vec::new();
        let mut gps_entries = Vec::new();

        // Step 1: Extract existing EXIF entries from the JPEG
        let existing = extract_existing_exif_entries(jpeg_data, bo);
        for entry in &existing {
            match classify_exif_tag(entry.tag) {
                ExifIfdGroup::Ifd0 => ifd0_entries.push(entry.clone()),
                ExifIfdGroup::ExifIfd => exif_entries.push(entry.clone()),
                ExifIfdGroup::Gps => gps_entries.push(entry.clone()),
            }
        }

        // Step 2: Apply queued changes (add/replace/delete)
        let deleted_tags: Vec<u16> = values
            .iter()
            .filter(|nv| nv.value.is_none())
            .filter_map(|nv| tag_name_to_id(&nv.tag))
            .collect();

        // Remove deleted tags
        ifd0_entries.retain(|e| !deleted_tags.contains(&e.tag));
        exif_entries.retain(|e| !deleted_tags.contains(&e.tag));
        gps_entries.retain(|e| !deleted_tags.contains(&e.tag));

        // Add/replace new values
        for nv in values {
            if nv.value.is_none() {
                continue;
            }
            let value_str = nv.value.as_deref().unwrap_or("");
            let group = nv.group.as_deref().unwrap_or("");

            if let Some((tag_id, format, encoded)) = encode_exif_tag(&nv.tag, value_str, group, bo)
            {
                let entry = exif_writer::IfdEntry {
                    tag: tag_id,
                    format,
                    data: encoded,
                };

                let target = match group.to_uppercase().as_str() {
                    "GPS" => &mut gps_entries,
                    "EXIFIFD" => &mut exif_entries,
                    _ => match classify_exif_tag(tag_id) {
                        ExifIfdGroup::ExifIfd => &mut exif_entries,
                        ExifIfdGroup::Gps => &mut gps_entries,
                        ExifIfdGroup::Ifd0 => &mut ifd0_entries,
                    },
                };

                // Replace existing or add new
                if let Some(existing) = target.iter_mut().find(|e| e.tag == tag_id) {
                    *existing = entry;
                } else {
                    target.push(entry);
                }
            }
        }

        // Remove sub-IFD pointers from entries (they'll be rebuilt by build_exif)
        ifd0_entries.retain(|e| e.tag != 0x8769 && e.tag != 0x8825 && e.tag != 0xA005);

        exif_writer::build_exif(&ifd0_entries, &exif_entries, &gps_entries, bo)
    }

    /// Write metadata changes to PNG data.
    fn write_png(&self, data: &[u8]) -> Result<Vec<u8>> {
        let mut new_text: Vec<(&str, &str)> = Vec::new();
        let mut remove_text: Vec<&str> = Vec::new();

        // Collect text-based changes
        // We need to hold the strings in vectors that live long enough
        let owned_pairs: Vec<(String, String)> = self
            .new_values
            .iter()
            .filter(|nv| nv.value.is_some())
            .map(|nv| (nv.tag.clone(), nv.value.clone().unwrap()))
            .collect();

        for (tag, value) in &owned_pairs {
            new_text.push((tag.as_str(), value.as_str()));
        }

        for nv in &self.new_values {
            if nv.value.is_none() {
                remove_text.push(&nv.tag);
            }
        }

        png_writer::write_png(data, &new_text, None, &remove_text)
    }

    /// Write metadata changes to PSD data.
    fn write_psd(&self, data: &[u8]) -> Result<Vec<u8>> {
        let mut iptc_values = Vec::new();
        let mut xmp_values = Vec::new();

        for nv in &self.new_values {
            let group = nv.group.as_deref().unwrap_or("").to_uppercase();
            match group.as_str() {
                "XMP" => xmp_values.push(nv),
                "IPTC" => iptc_values.push(nv),
                _ => {
                    if is_xmp_tag(&nv.tag) {
                        xmp_values.push(nv);
                    } else {
                        iptc_values.push(nv);
                    }
                }
            }
        }

        let new_iptc = if !iptc_values.is_empty() {
            let records: Vec<_> = iptc_values
                .iter()
                .filter_map(|nv| {
                    let value = nv.value.as_deref()?;
                    let (record, dataset) = iptc_writer::tag_name_to_iptc(&nv.tag)?;
                    Some(iptc_writer::IptcRecord {
                        record,
                        dataset,
                        // IPTC-IIM strings use the internal charset (Latin-1
                        // by default); writing raw UTF-8 double-encodes
                        // accented characters. See issue #6.
                        data: crate::encoding::encode_latin1(value),
                    })
                })
                .collect();
            if records.is_empty() {
                None
            } else {
                Some(iptc_writer::build_iptc(&records))
            }
        } else {
            None
        };

        let new_xmp = if !xmp_values.is_empty() {
            let refs: Vec<&NewValue> = xmp_values.to_vec();
            Some(self.build_new_xmp(&refs))
        } else {
            None
        };

        psd_writer::write_psd(data, new_iptc.as_deref(), new_xmp.as_deref())
    }

    /// Write metadata changes to Matroska (MKV/WebM) data.
    fn write_matroska(&self, data: &[u8]) -> Result<Vec<u8>> {
        let changes: Vec<(&str, &str)> = self
            .new_values
            .iter()
            .filter_map(|nv| {
                let value = nv.value.as_deref()?;
                Some((nv.tag.as_str(), value))
            })
            .collect();

        matroska_writer::write_matroska(data, &changes)
    }

    /// Write metadata changes to PDF data.
    fn write_pdf(&self, data: &[u8]) -> Result<Vec<u8>> {
        let changes: Vec<(&str, &str)> = self
            .new_values
            .iter()
            .filter_map(|nv| {
                let value = nv.value.as_deref()?;
                Some((nv.tag.as_str(), value))
            })
            .collect();

        pdf_writer::write_pdf(data, &changes)
    }

    /// Write metadata changes to MP4/MOV data.
    fn write_mp4(&self, data: &[u8]) -> Result<Vec<u8>> {
        let mut ilst_tags: Vec<([u8; 4], String)> = Vec::new();
        let mut xmp_values: Vec<&NewValue> = Vec::new();

        for nv in &self.new_values {
            if nv.value.is_none() {
                continue;
            }
            let group = nv.group.as_deref().unwrap_or("").to_uppercase();
            if group == "XMP" {
                xmp_values.push(nv);
            } else if let Some(key) = mp4_writer::tag_to_ilst_key(&nv.tag) {
                ilst_tags.push((key, nv.value.clone().unwrap()));
            }
        }

        let tag_refs: Vec<(&[u8; 4], &str)> =
            ilst_tags.iter().map(|(k, v)| (k, v.as_str())).collect();

        let new_xmp = if !xmp_values.is_empty() {
            let refs: Vec<&NewValue> = xmp_values.to_vec();
            Some(self.build_new_xmp(&refs))
        } else {
            None
        };

        mp4_writer::write_mp4(data, &tag_refs, new_xmp.as_deref())
    }

    /// Write metadata changes to WebP data.
    fn write_webp(&self, data: &[u8]) -> Result<Vec<u8>> {
        let mut exif_values: Vec<&NewValue> = Vec::new();
        let mut xmp_values: Vec<&NewValue> = Vec::new();
        let mut remove_exif = false;
        let mut remove_xmp = false;

        for nv in &self.new_values {
            let group = nv.group.as_deref().unwrap_or("").to_uppercase();
            if nv.value.is_none() && nv.tag == "*" {
                if group == "EXIF" {
                    remove_exif = true;
                }
                if group == "XMP" {
                    remove_xmp = true;
                }
                continue;
            }
            match group.as_str() {
                "XMP" => xmp_values.push(nv),
                _ => exif_values.push(nv),
            }
        }

        let new_exif = if !exif_values.is_empty() {
            let bo = ByteOrderMark::BigEndian;
            let mut entries = Vec::new();
            for nv in &exif_values {
                if let Some(ref v) = nv.value {
                    let group = nv.group.as_deref().unwrap_or("");
                    if let Some((tag_id, format, encoded)) = encode_exif_tag(&nv.tag, v, group, bo)
                    {
                        entries.push(exif_writer::IfdEntry {
                            tag: tag_id,
                            format,
                            data: encoded,
                        });
                    }
                }
            }
            if !entries.is_empty() {
                Some(exif_writer::build_exif(&entries, &[], &[], bo)?)
            } else {
                None
            }
        } else {
            None
        };

        let new_xmp = if !xmp_values.is_empty() {
            Some(self.build_new_xmp(&xmp_values.to_vec()))
        } else {
            None
        };

        webp_writer::write_webp(
            data,
            new_exif.as_deref(),
            new_xmp.as_deref(),
            remove_exif,
            remove_xmp,
        )
    }

    /// Write metadata changes to TIFF data.
    fn write_tiff(&self, data: &[u8]) -> Result<Vec<u8>> {
        let bo = if data.starts_with(b"II") {
            ByteOrderMark::LittleEndian
        } else {
            ByteOrderMark::BigEndian
        };

        let mut changes: Vec<(u16, Vec<u8>)> = Vec::new();
        for nv in &self.new_values {
            if let Some(ref value) = nv.value {
                let group = nv.group.as_deref().unwrap_or("");
                if let Some((tag_id, _format, encoded)) = encode_exif_tag(&nv.tag, value, group, bo)
                {
                    changes.push((tag_id, encoded));
                }
            }
        }

        tiff_writer::write_tiff(data, &changes)
    }

    /// Build new XMP data from queued values.
    fn build_new_xmp(&self, values: &[&NewValue]) -> Vec<u8> {
        let mut properties = Vec::new();

        for nv in values {
            let value_str = match &nv.value {
                Some(v) => v.clone(),
                None => continue,
            };

            let ns = nv.group.as_deref().unwrap_or("dc").to_lowercase();
            let ns = if ns == "xmp" { "xmp".to_string() } else { ns };

            let prop_type = match nv.tag.to_lowercase().as_str() {
                "title" | "description" | "rights" => xmp_writer::XmpPropertyType::LangAlt,
                "subject" | "keywords" => xmp_writer::XmpPropertyType::Bag,
                "creator" => xmp_writer::XmpPropertyType::Seq,
                _ => xmp_writer::XmpPropertyType::Simple,
            };

            let values = if matches!(
                prop_type,
                xmp_writer::XmpPropertyType::Bag | xmp_writer::XmpPropertyType::Seq
            ) {
                value_str.split(',').map(|s| s.trim().to_string()).collect()
            } else {
                vec![value_str]
            };

            properties.push(xmp_writer::XmpProperty {
                namespace: ns,
                property: nv.tag.clone(),
                values,
                prop_type,
            });
        }

        xmp_writer::build_xmp(&properties).into_bytes()
    }

    // ================================================================
    // Reading API
    // ================================================================

    /// Extract metadata from a file and return a simple name→value map.
    ///
    /// This is the high-level one-shot API, equivalent to ExifTool's `ImageInfo()`.
    pub fn image_info<P: AsRef<Path>>(&self, path: P) -> Result<ImageInfo> {
        let tags = self.extract_info(path)?;
        Ok(self.get_info(&tags))
    }

    /// Extract all metadata tags from a file.
    ///
    /// Returns the full `Tag` structs with groups, raw values, etc.
    pub fn extract_info<P: AsRef<Path>>(&self, path: P) -> Result<Vec<Tag>> {
        let path = path.as_ref();
        // Memory-map the file instead of reading it fully into a Vec. Our format
        // readers walk container structures by offset (mp4/mov skip `mdat`, Matroska
        // stops at the first Cluster), so only the header pages are ever faulted in —
        // a multi-gigabyte video is parsed by touching a few MB, not by allocating
        // and reading the whole file. Falls back to a plain read when mapping fails.
        let data = map_file_for_read(path)?;
        self.extract_info_from_bytes(&data, path)
    }

    /// Extract metadata from in-memory data.
    pub fn extract_info_from_bytes(&self, data: &[u8], path: &Path) -> Result<Vec<Tag>> {
        // Propagate show_unknown to EXIF/MakerNotes parsers via thread-local
        crate::metadata::exif::set_show_unknown(self.options.show_unknown);
        // Propagate the Duplicates option (see `collapse_duplicates` below) to the
        // EXIF/MakerNotes reader, whose name-level pruning must not run when every
        // instance has to be reported.
        crate::metadata::exif::set_keep_duplicates(
            self.options.duplicates || self.options.extract_embedded > 0,
        );
        // Propagate process_compressed to format readers via thread-local
        crate::formats::pdf::set_process_compressed(self.options.process_compressed);

        // ExifTool's `$$self{TIFF_TYPE}`. Several EXIF tag names read it -- 0x0201
        // in IFD0 is a thumbnail offset in a JPEG and a preview offset in an ARW --
        // so it has to be in place before anything is parsed, not after.
        let file_type_result = self.detect_file_type(data, path);
        crate::metadata::exif::set_tiff_type(file_type_result.as_ref().map_or("", |ft| ft.code()));
        let (file_type, mut tags) = match file_type_result {
            Ok(ft) => {
                let t = self
                    .process_file(data, ft)
                    .or_else(|_| self.process_by_extension(data, path))?;
                (Some(ft), t)
            }
            Err(_) => {
                // File type unknown by magic/extension — try extension-based fallback
                let t = self.process_by_extension(data, path)?;
                (None, t)
            }
        };
        let file_type = file_type.unwrap_or(FileType::Zip); // placeholder for file-level tags

        // Some types refine their FileType/MIMEType/extension from the content
        // (ExifTool SetFileType): e.g. EXE -> "Win32 EXE" / "ELF executable" / Mach-O.
        let default_tags = || {
            (
                file_type.code().to_string(),
                file_type.mime_type().to_string(),
                file_type
                    .extensions()
                    .first()
                    .copied()
                    .unwrap_or("")
                    .to_string(),
            )
        };
        // Office Open XML (DOCX/XLSX/PPTX/…): ExifTool sub-detects the type from
        // [Content_Types].xml (ZIP.pm ProcessZIP -> OOXML.pm ProcessDOCX). The ZIP
        // member tags stay in the [ZIP] group; only the three File-group pseudo-tags
        // (FileType/MIMEType/FileTypeExtension) change from the generic ZIP identity.
        let ooxml = if file_type == FileType::Zip {
            crate::formats::zip::detect_ooxml_type(data, path.extension().and_then(|e| e.to_str()))
        } else {
            None
        };
        let (ft_code, mime_str, ext_str): (String, String, String) = if file_type == FileType::Exe {
            exe_subtype(data)
                .map(|(ft, mime, ext)| (ft.to_string(), mime.to_string(), ext.to_string()))
                .unwrap_or_else(default_tags)
        } else if let Some(triple) = ooxml {
            triple
        } else if let Some((code, mime)) = refine_filetype_by_content(file_type, data) {
            let (_, _, ext) = default_tags();
            (code, mime, ext)
        } else {
            default_tags()
        };

        // File-level pseudo-tags, emitted in ExifTool's own order.
        //
        // `ExtractInfo` reports them before it hands the file to a format reader:
        // ExifToolVersion (ExifTool.pm:2779), FileName (:2824), Directory (:2830),
        // FileSize (:2904), FileModifyDate (:2907), FileAccessDate (:2908),
        // FileInodeChangeDate (:2910), FilePermissions (:2921). Every format handler
        // then opens with `SetFileType`, which emits FileType, FileTypeExtension and
        // MIMEType in that order (ExifTool.pm:9713-9715). ExifByteOrder follows,
        // reported while the EXIF block itself is read.
        //
        // The order is load-bearing, not cosmetic: FoundTag arbitrates duplicates by
        // priority and then last-wins, so a pseudo-tag emitted after the format tags
        // would take a name it has to lose. They are collected here and spliced in
        // front of the format tags.
        let mut pre: Vec<Tag> = Vec::new();

        // The File/File/Other group below is only a default: the pseudo-tags that
        // ExifTool places elsewhere (FileName, Directory, the File*Date tags,
        // ExifToolVersion, …) have their groups resolved from `FILE_LEVEL_GROUPS`
        // at the end of extraction. Tags that genuinely belong to File:File, such
        // as FileTypeExtension and ExifByteOrder, keep this default.

        let file_tag = |name: &str, val: Value| -> Tag {
            Tag {
                id: crate::tag::TagId::Text(name.to_string()),
                name: name.to_string(),
                description: name.to_string(),
                group: crate::tag::TagGroup {
                    family0: "File".into(),
                    family1: "File".into(),
                    family2: "Other".into(),
                    family3: "Main".into(),
                },
                raw_value: val.clone(),
                print_value: val.to_display_string(),
                priority: 1,
            }
        };

        pre.push(file_tag(
            "ExifToolVersion",
            Value::String(crate::VERSION.to_string()),
        ));

        if let Some(fname) = path.file_name().and_then(|n| n.to_str()) {
            pre.push(file_tag("FileName", Value::String(fname.to_string())));
        }
        if let Some(dir) = path.parent().and_then(|p| p.to_str()) {
            pre.push(file_tag("Directory", Value::String(dir.to_string())));
        }

        if let Ok(metadata) = fs::metadata(path) {
            pre.push(Tag {
                id: crate::tag::TagId::Text("FileSize".into()),
                name: "FileSize".into(),
                description: "File Size".into(),
                group: crate::tag::TagGroup {
                    family0: "File".into(),
                    family1: "File".into(),
                    family2: "Other".into(),
                    family3: "Main".into(),
                },
                // String, not U32: a file may exceed 4 GB (`as u32` would silently
                // truncate). `-n` prints this verbatim, matching Perl's raw byte count.
                raw_value: Value::String(metadata.len().to_string()),
                print_value: format_file_size(metadata.len()),
                priority: 0,
            });
        }

        #[cfg(unix)]
        if let Ok(metadata) = fs::metadata(path) {
            use std::os::unix::fs::MetadataExt;
            let mode = metadata.mode();
            // Dates use ConvertUnixTime($val, 1): local time with a numeric TZ offset
            // (e.g. "2026:06:13 15:14:15+02:00"), same conversion as GZIP's ModifyDate.
            use crate::formats::gzip::gzip_unix_to_datetime;
            // FileModifyDate
            if let Ok(modified) = metadata.modified() {
                if let Ok(dur) = modified.duration_since(std::time::UNIX_EPOCH) {
                    let secs = dur.as_secs() as i64;
                    pre.push(file_tag(
                        "FileModifyDate",
                        Value::String(gzip_unix_to_datetime(secs)),
                    ));
                }
            }
            // FileAccessDate
            if let Ok(accessed) = metadata.accessed() {
                if let Ok(dur) = accessed.duration_since(std::time::UNIX_EPOCH) {
                    let secs = dur.as_secs() as i64;
                    pre.push(file_tag(
                        "FileAccessDate",
                        Value::String(gzip_unix_to_datetime(secs)),
                    ));
                }
            }
            // FileInodeChangeDate (ctime on Unix)
            let ctime = metadata.ctime();
            if ctime > 0 {
                pre.push(file_tag(
                    "FileInodeChangeDate",
                    Value::String(gzip_unix_to_datetime(ctime)),
                ));
            }

            // Port of ExifTool's FilePermissions: ValueConv is the full mode in octal
            // (`sprintf "%.3o"`, includes the file-type bits), PrintConv is the ls-style
            // "-rw-rw-r--" string built from those same bits.
            pre.push(Tag {
                id: crate::tag::TagId::Text("FilePermissions".into()),
                name: "FilePermissions".into(),
                description: "FilePermissions".into(),
                group: crate::tag::TagGroup {
                    family0: "File".into(),
                    family1: "File".into(),
                    family2: "Other".into(),
                    family3: "Main".into(),
                },
                raw_value: Value::String(format!("{:o}", mode)),
                print_value: format_file_permissions(mode),
                priority: 1,
            });
        }

        pre.push(Tag {
            id: crate::tag::TagId::Text("FileType".into()),
            name: "FileType".into(),
            description: "File Type".into(),
            group: crate::tag::TagGroup {
                family0: "File".into(),
                family1: "File".into(),
                family2: "Other".into(),
                family3: "Main".into(),
            },
            raw_value: Value::String(format!("{:?}", file_type)),
            // ExifTool's FileType value is the short code ("JPEG"), not the
            // human-readable description ("JPEG image").
            print_value: ft_code.clone(),
            priority: 1,
        });

        // Use the canonical (first) extension from the FileType, matching Perl ExifTool behavior.
        // EXE subtypes emit FileTypeExtension even when empty (ExifTool sets ext='').
        if !ext_str.is_empty() || file_type == FileType::Exe {
            pre.push(file_tag(
                "FileTypeExtension",
                Value::String(ext_str.clone()),
            ));
        }

        pre.push(Tag {
            id: crate::tag::TagId::Text("MIMEType".into()),
            name: "MIMEType".into(),
            description: "MIME Type".into(),
            group: crate::tag::TagGroup {
                family0: "File".into(),
                family1: "File".into(),
                family2: "Other".into(),
                family3: "Main".into(),
            },
            raw_value: Value::String(mime_str.clone()),
            print_value: mime_str.clone(),
            priority: 1,
        });

        // ExifByteOrder (from TIFF header)
        {
            let bo_str = if data.len() > 8 {
                // Check EXIF in JPEG or TIFF header or WebP/RIFF EXIF chunk
                let check: Option<&[u8]> = if data.starts_with(&[0xFF, 0xD8]) {
                    // JPEG: find APP1 EXIF header
                    data.windows(6)
                        .position(|w| w == b"Exif\0\0")
                        .map(|p| &data[p + 6..])
                } else if data.starts_with(b"FUJIFILMCCD-RAW") && data.len() >= 0x60 {
                    // RAF: look in the embedded JPEG for EXIF byte order
                    let jpeg_offset =
                        u32::from_be_bytes([data[0x54], data[0x55], data[0x56], data[0x57]])
                            as usize;
                    let jpeg_length =
                        u32::from_be_bytes([data[0x58], data[0x59], data[0x5A], data[0x5B]])
                            as usize;
                    if jpeg_offset > 0 && jpeg_offset + jpeg_length <= data.len() {
                        let jpeg = &data[jpeg_offset..jpeg_offset + jpeg_length];
                        jpeg.windows(6)
                            .position(|w| w == b"Exif\0\0")
                            .map(|p| &jpeg[p + 6..])
                    } else {
                        None
                    }
                } else if data.starts_with(b"RIFF") && data.len() >= 12 {
                    // RIFF/WebP: find EXIF chunk
                    let mut riff_bo: Option<&[u8]> = None;
                    let mut pos = 12usize;
                    while pos + 8 <= data.len() {
                        let cid = &data[pos..pos + 4];
                        let csz = u32::from_le_bytes([
                            data[pos + 4],
                            data[pos + 5],
                            data[pos + 6],
                            data[pos + 7],
                        ]) as usize;
                        let cstart = pos + 8;
                        let cend = (cstart + csz).min(data.len());
                        if cid == b"EXIF" && cend > cstart {
                            let exif_data = &data[cstart..cend];
                            let tiff = if exif_data.starts_with(b"Exif\0\0") {
                                &exif_data[6..]
                            } else {
                                exif_data
                            };
                            riff_bo = Some(tiff);
                            break;
                        }
                        // Also check LIST chunks
                        if cid == b"LIST" && cend >= cstart + 4 {
                            // recurse not needed for this simple scan - just advance
                        }
                        pos = cend + (csz & 1);
                    }
                    riff_bo
                } else if data.starts_with(&[0x00, 0x00, 0x00, 0x0C, b'J', b'X', b'L', b' ']) {
                    // JXL container: the Exif payload lives in an (optionally
                    // brotli-compressed) box that Jpeg2000.pm hands to ProcessTIFF,
                    // and ProcessTIFF is what raises ExifByteOrder — in box order,
                    // after the container's own tags. The JXL reader already does
                    // that, so no pre-scan value is contributed here.
                    None
                } else if data.starts_with(&[0x00, b'M', b'R', b'M']) {
                    // MRW: find TTW segment which contains TIFF/EXIF data
                    let mrw_data_offset = if data.len() >= 8 {
                        u32::from_be_bytes([data[4], data[5], data[6], data[7]]) as usize + 8
                    } else {
                        0
                    };
                    let mut mrw_bo: Option<&[u8]> = None;
                    let mut mpos = 8usize;
                    while mpos + 8 <= mrw_data_offset.min(data.len()) {
                        let seg_tag = &data[mpos..mpos + 4];
                        let seg_len = u32::from_be_bytes([
                            data[mpos + 4],
                            data[mpos + 5],
                            data[mpos + 6],
                            data[mpos + 7],
                        ]) as usize;
                        if seg_tag == b"\x00TTW" && mpos + 8 + seg_len <= data.len() {
                            mrw_bo = Some(&data[mpos + 8..mpos + 8 + seg_len]);
                            break;
                        }
                        mpos += 8 + seg_len;
                    }
                    mrw_bo
                } else {
                    Some(data)
                };
                if let Some(tiff) = check {
                    if tiff.starts_with(b"II") {
                        "Little-endian (Intel, II)"
                    } else if tiff.starts_with(b"MM") {
                        "Big-endian (Motorola, MM)"
                    } else {
                        ""
                    }
                } else {
                    ""
                }
            } else {
                ""
            };
            // Suppress ExifByteOrder for BigTIFF, Canon VRD/DR4 (Perl doesn't output it for these)
            // Also skip if already emitted by ExifReader (TIFF-based formats)
            let already_has_exifbyteorder = tags.iter().any(|t| t.name == "ExifByteOrder");
            if !bo_str.is_empty()
                && !already_has_exifbyteorder
                && file_type != FileType::Btf
                && file_type != FileType::Dr4
                && file_type != FileType::Vrd
                && file_type != FileType::Crw
            {
                pre.push(file_tag("ExifByteOrder", Value::String(bo_str.to_string())));
            }
        }

        // The pseudo-tags collected above precede every format tag.
        tags.splice(0..0, pre);

        // A format reader that overrides the file's MIME type does it by
        // assignment in ExifTool -- `$$et{VALUE}{MIMEType} = $mimeTypes[0]`
        // (Real.pm:655) -- which replaces the value in place instead of adding
        // a second tag. Keep only the first File:MIMEType, carrying the last
        // value, so the count stays one even with the Duplicates option on.
        {
            // Only the main document's: an embedded image legitimately reports
            // its own MIMEType in its own document.
            let is_mime = |t: &Tag| {
                t.name == "MIMEType"
                    && t.group.family0 == "File"
                    && t.group.family3 == crate::tag::MAIN_DOCUMENT
            };
            if tags.iter().filter(|t| is_mime(t)).count() > 1 {
                let last = tags.iter().rposition(is_mime).unwrap();
                let (value, print) = (tags[last].raw_value.clone(), tags[last].print_value.clone());
                let first = tags.iter().position(is_mime).unwrap();
                tags[first].raw_value = value;
                tags[first].print_value = print;
                let mut seen = false;
                tags.retain(|t| {
                    !is_mime(t) || {
                        let keep = !seen;
                        seen = true;
                        keep
                    }
                });
            }
        }

        // Promote authoritative specialized-source tags before computing composites,
        // so derived tags (ShutterSpeed, LightValue, ...) use the primary value.
        //
        // ExifTool builds composites from `$$self{VALUE}`, i.e. from the winner of
        // the duplicate arbitration, and Kodak's own ExposureTime/FNumber are found
        // after the ExifIFD ones, so they are what its Composite ShutterSpeed and
        // LightValue use. Here composites are computed before that arbitration, so
        // the winner has to be brought forward for these two names. The rest of the
        // list this used to hold (MinoltaRaw, Lytro) is gone: the group-blind
        // FoundTag pass now settles those by last-wins on its own.
        {
            const SPECIAL_WINS: &[(&str, &str)] =
                &[("Kodak", "FNumber"), ("Kodak", "ExposureTime")];
            //
            // Dropping the loser is only correct while duplicates are being
            // collapsed. With the Duplicates option on (which `-ee` turns on,
            // exiftool line 1030) ExifTool still lists BOTH — `-ee` on Kodak.jpg
            // prints ExifIFD 0x829a `ExposureTime: 1/180` and Kodak 0x0020
            // `ExposureTime: 1/216` — so there the promotion has to leave the
            // loser in place, in its original position, and only make the winner
            // the one the composites read.
            let keep_dups = self.options.duplicates || self.options.extract_embedded > 0;
            for (grp, name) in SPECIAL_WINS {
                if !tags
                    .iter()
                    .any(|t| t.name == *name && t.group.family1 == *grp)
                {
                    continue;
                }
                if keep_dups {
                    for t in tags.iter_mut() {
                        if t.name == *name && t.group.family1 == *grp {
                            t.priority = t.priority_rank() + 1;
                        }
                    }
                } else {
                    tags.retain(|t| t.name != *name || t.group.family1 == *grp);
                }
            }
        }

        // GPS::Composite GPSLatitude/GPSLongitude first: the other composites
        // (GPSPosition) Require them, and ExifTool resolves inter-composite
        // dependencies the same way (BuildCompositeTags defers a composite whose
        // requirements are themselves composites).
        let gps = crate::composite::gps_coordinates(&tags);
        tags.extend(gps);

        // Compute composite tags
        let composite = crate::composite::compute_composite_tags(&tags);
        tags.extend(composite);

        // Composite GPSAltitude (GPS.pm:406) claims the "GPSAltitude" name whenever
        // its Desire GPSAltitudeRef is present — FoundTag stores the composite and
        // moves the plain GPS:GPSAltitude aside to "GPSAltitude (1)". When the plain
        // altitude is a 0/0 rational the composite's ValueConv yields undef, so it
        // prints nothing; the displaced plain tag is then visible only with the
        // Duplicates option on. Model that here: with duplicates collapsed and no
        // real Composite GPSAltitude built, drop the "undef" plain tag.
        if !(self.options.duplicates || self.options.extract_embedded > 0) {
            let has_composite_alt = tags
                .iter()
                .any(|t| t.name == "GPSAltitude" && t.group.family0 == "Composite");
            let has_alt_ref = tags.iter().any(|t| t.name == "GPSAltitudeRef");
            if !has_composite_alt && has_alt_ref {
                tags.retain(|t| {
                    !(t.name == "GPSAltitude"
                        && t.group.family0 == "EXIF"
                        && t.print_value == "undef")
                });
            }
        }

        // No name filter for Composite RedBalance/BlueBalance. `%Exif::Composite`
        // declares them with `Desire` only and no `Priority` (Exif.pm:5235-5260),
        // so a manufacturer's own RedBalance is extracted too; the Composite is
        // built after the file has been read, and therefore wins the duplicate
        // competition by last-wins when duplicates are collapsed.

        // Geolocation is opt-in, matching ExifTool's `Geolocation` API option.
        if self.options.geolocation {
            if let Some(geo) = crate::composite::compute_geolocation(&tags) {
                tags.extend(geo);
            }
        }

        // MWG (Metadata Working Group) composite tags
        if self.options.use_mwg {
            let mwg = crate::composite::compute_mwg_composites(&tags);
            tags.extend(mwg);
        }

        // FLIR post-processing: remove LensID composite for FLIR cameras.
        // Perl's LensID composite requires LensType EXIF tag (not present in FLIR images),
        // and LensID-2 requires LensModel to match /(mm|\d\/F)/ (FLIR names like "FOL7"
        // don't match).  Our composite.rs uses a simpler fallback that picks up any non-empty
        // LensModel, so we remove LensID when the image is from a FLIR camera with FFF data.
        {
            let is_flir_fff = tags
                .iter()
                .any(|t| t.group.family0 == "APP1" && t.group.family1 == "FLIR");
            if is_flir_fff {
                tags.retain(|t| !(t.name == "LensID" && t.group.family0 == "Composite"));
            }
        }

        // Olympus post-processing: remove the generic "Lens" composite for Olympus cameras.
        // In Perl, the "Lens" composite tag requires Canon:MinFocalLength (Canon namespace).
        // Our composite.rs generates Lens for any manufacturer that has MinFocalLength +
        // MaxFocalLength (e.g., Olympus Equipment sub-IFD).  Remove it for non-Canon cameras.
        {
            let make = tags
                .iter()
                .find(|t| t.name == "Make")
                .map(|t| t.print_value.clone())
                .unwrap_or_default();
            if !make.to_uppercase().contains("CANON") {
                tags.retain(|t| t.name != "Lens" || t.group.family0 != "Composite");
            }
        }

        // Priority-based deduplication: when the same tag name appears multiple times,
        // keep only the one with the highest priority (e.g., EXIF over JFIF, FFF over MakerNote).
        //
        // Every pass below collapses same-named tags, which ExifTool only does when
        // the Duplicates option is off (`CombineInfo`/`GetInfo`). The exiftool CLI
        // turns Duplicates on together with ExtractEmbedded (`$mt->Options(Duplicates
        // => 1)` in the -ee branch), so -ee must keep every instance: the IFD1 copy
        // of XResolution, each ZIP member's Zip* set, every GPX track point.
        let collapse_duplicates = !self.options.duplicates && self.options.extract_embedded == 0;
        if collapse_duplicates {
            // Perl keys its extracted-info hash on the tag NAME alone, so with the
            // Duplicates option off a tag found in a sub-document is dropped as
            // soon as the file already reported that name — whatever source
            // either of them came from, since `FoundTag` never lets a tag
            // carrying a DOC_NUM override one that does not. (Below, competition
            // is keyed on the family-0 source, which cannot express that.) This is
            // why a CR3 read without `-ee` shows the Canon Timed MetaData
            // FocalLength only when the main document has no FocalLength.
            {
                let mut seen: std::collections::HashSet<&str> = tags
                    .iter()
                    .filter(|t| t.group.family3 == MAIN_DOCUMENT)
                    .map(|t| t.name.as_str())
                    .collect();
                let mut keep = Vec::with_capacity(tags.len());
                for t in &tags {
                    keep.push(t.group.family3 == MAIN_DOCUMENT || seen.insert(t.name.as_str()));
                }
                let mut it = keep.into_iter();
                tags.retain(|_| it.next().unwrap_or(true));
            }

            // Specialized-source precedence: a few container/sidecar groups are
            // authoritative for specific tags and win over a generic EXIF copy
            // (ExifTool reports the GoPro GPMF value). Applied before the priority
            // dedup so the (priority-0) specialized tag isn't pruned first.
            {
                const SPECIAL_WINS: &[(&str, &str)] = &[
                    ("GoPro", "WhiteBalance"),
                    ("GoPro", "Sharpness"),
                    ("GoPro", "ExposureCompensation"),
                    // Embedded ID3v2 Comment overrides the native container's
                    // (AIFF/...). ID3v1 is NOT in this list: its table is
                    // `PRIORITY => 0` (ID3.pm:338), so it loses to the
                    // container's tag instead of displacing it.
                    ("ID3v2_4", "Comment"),
                    ("ID3v2_3", "Comment"),
                    ("ID3v2_2", "Comment"),
                    // Minolta RAW (.mrw PRD/native block) is authoritative for these
                    // over the embedded EXIF maker note copies.
                    ("MinoltaRaw", "Contrast"),
                    ("MinoltaRaw", "Saturation"),
                    ("MinoltaRaw", "Sharpness"),
                    ("MinoltaRaw", "ISOSetting"),
                    // Kodak maker note carries more precise Exposure/FNumber than EXIF.
                    ("Kodak", "FNumber"),
                    ("Kodak", "ExposureTime"),
                    // Sigma maker note X3FillLight (int) is primary over the X3F header.
                    ("Sigma", "X3FillLight"),
                ];
                for (grp, name) in SPECIAL_WINS {
                    if tags
                        .iter()
                        .any(|t| t.name == *name && t.group.family1 == *grp)
                    {
                        tags.retain(|t| t.name != *name || t.group.family1 == *grp);
                    }
                }
            }

            let mut best_priority: HashMap<String, i32> = HashMap::new();
            for tag in &tags {
                let entry = best_priority
                    .entry(tag.name.clone())
                    .or_insert_with(|| tag.priority_rank());
                if tag.priority_rank() > *entry {
                    *entry = tag.priority_rank();
                }
            }
            tags.retain(|t| t.priority_rank() >= *best_priority.get(&t.name).unwrap_or(&0));

            // Document formats (PDF/PostScript/DjVu): their native Info metadata is the
            // LOWEST priority in ExifTool — XMP and embedded EXIF both win. Drop the
            // native copy when any non-native source provides the same tag.
            {
                // DjVu-Meta is deliberately absent: %Image::ExifTool::DjVu::Meta
                // (DjVu.pm line 132) declares no PRIORITY, so its tags meet XMP's
                // at the normal priority and FoundTag decides between them. The
                // DjVu INFO chunk is the low-priority one (`PRIORITY => 0, # first
                // INFO block takes priority`, DjVu.pm line 60).
                let is_native_doc = |g1: &str| matches!(g1, "PDF" | "PostScript" | "DjVu");
                let other_names: std::collections::HashSet<String> = tags
                    .iter()
                    .filter(|t| !is_native_doc(&t.group.family1) && !t.print_value.is_empty())
                    .map(|t| t.name.clone())
                    .collect();
                tags.retain(|t| {
                    // Trapped keeps its native value ('Unknown' vs XMP's raw '/Unknown').
                    t.name == "Trapped"
                        || !is_native_doc(&t.group.family1)
                        || !other_names.contains(&t.name)
                });
            }

            // ExifTool FoundTag rule. Among duplicates of the same tag name,
            // ExifTool keeps one primary instance decided purely by priority --
            // the comparison is group-blind (ExifTool.pm `FoundTag`, "take tag
            // with highest priority"):
            //
            //   * the incoming tag replaces the stored one iff its priority is
            //     >= the stored tag's priority;
            //   * a stored priority of 0 is PROMOTED to 1 first ("promote
            //     existing 0-priority tag so it takes precedence over a new
            //     0-tag"), so a priority-0 duplicate never displaces anything.
            //
            // With ExifTool's two usual priorities that reduces to: the LAST
            // instance wins at default priority, the FIRST wins when every
            // instance is priority 0. That single rule replaces what used to be
            // a hand-maintained list of "first-wins" container groups -- those
            // groups were simply the places where priority-0 duplicates happened
            // to have been noticed.
            //
            // The promotion has three exemptions (ExifTool.pm:9541-9548), and all
            // three collapse into the unconditional `.max(1)` used below:
            //   * the incoming tag has a DOC_NUM,
            //   * the tag is `Warning` ("never override a Warning tag because
            //     they may be added by ValueConv"),
            //   * the stored tag has no G3, i.e. it belongs to the main document.
            // Only the remaining case skips the promotion -- a main-document tag
            // arriving on top of a stored SUB-document one, which is then allowed
            // to displace it at equal priority ("don't promote sub-document tag
            // over main document"). That case cannot reach here: the pass at the
            // top of this block already drops every sub-document tag whose name
            // the main document also reports, which is the same outcome, and it
            // is also what the main condition's `not $$self{DOC_NUM} or
            // ($$self{TAG_EXTRA}{$tag}{G3} and $$self{DOC_NUM} eq ...{G3})` does
            // in the other direction -- an incoming sub-document tag never takes
            // the primary key from the main document, nor from another document.
            //
            // Like every pass in this block, the rule only applies when duplicates
            // are being collapsed — see `collapse_duplicates` above.
            {
                // Sources ExifTool gives priority 0, so that a duplicate coming
                // from them never displaces an already-stored tag:
                //   * the directories it flags LOW_PRIORITY_DIR (PreviewIFD,
                //     IFD1) -- a thumbnail describes a different image, so it
                //     must not override the main one;
                //   * XMP.pm marks its TIFF/EXIF mirror tables `PRIORITY => 0`
                //     ("not as reliable as actual EXIF tags");
                //   * an XMP property with no table entry gets a generated
                //     `{ Name, IsDefault => 1, Priority => 0 }` tagInfo;
                //   * container tables (Jpeg2000, PhotoMechanic, ...) whose
                //     stored tags ExifTool keeps first. A QuickTime track is NOT
                //     one of them: only the tkhd fields carry `Priority => 0`,
                //     and they say so tag by tag (see `parse_tkhd`), while the
                //     mdhd/hdlr/stsd tags keep the normal priority and follow the
                //     usual last-wins rule.
                //
                // VCard is deliberately absent: within one vCard, duplicate tags
                // are last-wins (TelephoneOtherVoice); the 2nd vCard is demoted
                // to priority -1 instead.
                //
                // Individual tags carrying their own `Priority => 0`, keyed by
                // the family-1 group of the table holding them. Canon::ShotInfo
                // BaseISO is the CIFF case: a .crw stores the value twice, once
                // in CanonRaw and once in the Canon MakerNotes, and ExifTool
                // keeps the CanonRaw one.
                #[rustfmt::skip]
                const LOW_PRIORITY_TAGS: &[(&str, &str)] = &[
                    ("Canon", "BaseISO"),          // Canon.pm:2789
                    // Canon::ShotInfo 22/23 carry `Priority => 0` (Canon.pm:2959,
                    // 2973, 2986) so the ExifIFD copies win. Canon::ExposureInfo
                    // has the same two names at default priority, but it is CR3
                    // timed metadata, i.e. a sub-document only `-ee` reaches, and
                    // duplicates are never collapsed under `-ee`.
                    ("Canon", "FNumber"),
                    ("Canon", "ExposureTime"),
                    // Canon::FocalLength 1: "the EXIF FocalLength is more reliable,
                    // so set this priority to zero" (Canon.pm:2709-2710). A CIFF
                    // file reaches the same table through CanonRaw 0x1029.
                    ("Canon", "FocalLength"),
                    ("CIFF", "FocalLength"),
                    // Sigma.pm:324, 337, 350, 363, 376 — the MakerNotes copies of
                    // the X3F header's picture-adjustment values are `Priority => 0`,
                    // so SigmaRaw::HeaderExt (no PRIORITY) keeps them.
                    ("Sigma", "Contrast"),
                    ("Sigma", "Shadow"),
                    ("Sigma", "Highlight"),
                    ("Sigma", "Saturation"),
                    ("Sigma", "Sharpness"),
                ];
                // Tables ExifTool declares `PRIORITY => 0` wholesale, keyed by the
                // family-1 group their tags land in:
                //   * APP12.pm:27 `%APP12::PictureInfo` — the JPEG APP12 "Picture
                //     Info" segment never displaces an EXIF value;
                //   * SigmaRaw.pm:138 `%SigmaRaw::Properties` — "(because these
                //     aren't writable like the EXIF ones)". Only the PROP tags are
                //     demoted; SigmaRaw::HeaderExt keeps the default priority, which
                //     is why the X3F header still wins Contrast and friends above;
                //   * CaptureOne COS properties are invented on the fly, and XMP.pm:3595
                //     builds those tagInfos as `{ Name, IsDefault => 1, Priority => 0 }`.
                //     They land in family-1 group XML (CaptureOne.pm:26), whose table
                //     declares one static tag only.
                const LOW_PRIORITY_GROUPS1: &[&str] = &["PictureInfo", "XML"];
                // The tag names of `%SigmaRaw::Properties` (SigmaRaw.pm:135-...),
                // which shares its family-1 group with the normal-priority
                // SigmaRaw::Header* tables and so cannot be demoted group-wide.
                #[rustfmt::skip]
                const SIGMARAW_PROPERTIES: &[&str] = &[
                    "AFArea", "AFInFocus", "ApertureDisplayed", "BracketShot",
                    "BurstShot", "CameraName", "ColorSpace", "DateTimeOriginal",
                    "DriveMode", "EvalState", "ExposureCompensation",
                    "ExposureProgram", "ExposureTime", "FNumber", "FirmwareVersion",
                    "FlashExpComp", "FlashMode", "FlashPower", "FlashTTLMode",
                    "FlashType", "FocalLength", "FocalLengthIn35mmFormat", "Focus",
                    "FocusMode", "ISO", "ImageBoardID", "ImagerBoardID",
                    "IntegrationTime", "LensApertureRange", "LensFocalRange",
                    "LensType", "Make", "MeteringMode", "Model",
                    "NetExposureCompensation", "Quality", "SceneCaptureType",
                    "SensorID", "SensorTemperature", "SerialNumber",
                    "ShutterSpeedDisplayed", "VersionBF", "WhiteBalance",
                ];
                // SceneCaptureType above is also an X3F Header2 field
                // (SigmaRaw.pm:94) at the default priority. Listing it costs
                // nothing: the header copy is emitted before the PROP one, and a
                // priority-0 tie is first-wins, so the header value still leads.
                // ExifTool.pm:4368 initialises `LOW_PRIORITY_DIR = { PreviewIFD => 1 }`
                // and only two places ever add to it: ProcessJPEG (ExifTool.pm:7317,
                // `$$self{LOW_PRIORITY_DIR}{IFD1} = 1; # lower priority of IFD1 tags`)
                // and ProcessTIFF for ARW (ExifTool.pm:8685). So IFD1 is demoted in a
                // JPEG-family file and in an ARW, but NOT in a plain TIFF/RAW, where
                // IFD1 keeps the default priority of 1 and so wins by last-wins.
                // A live dump of `%{$et->{LOW_PRIORITY_DIR}}` over the corpus confirms
                // it: only the JPEG and JPS files list IFD1.
                let ifd1_low = matches!(ft_code.as_str(), "JPEG" | "JPS" | "MPO" | "ARW");
                let is_low_priority_source = |g: &TagGroup, name: &str| -> bool {
                    let g1 = g.family1.as_str();
                    // An invented `XMP::other` tag carries `Priority => 0`
                    // whatever family 0 it reports: XMP.pm builds such a tagInfo as
                    // `{ Name, IsDefault => 1, Priority => 0 }` (XMP.pm:3595).
                    if g.family2 == "Unknown" {
                        return true;
                    }
                    // A sub-document never displaces the main document's tag:
                    // ExifTool only lets the incoming tag override when it carries
                    // no DOC_NUM, or the same one as the tag already stored.
                    if g.family3 != MAIN_DOCUMENT {
                        return true;
                    }
                    if LOW_PRIORITY_TAGS.contains(&(g1, name))
                        || LOW_PRIORITY_GROUPS1.contains(&g1)
                        || (g1 == "SigmaRaw" && SIGMARAW_PROPERTIES.contains(&name))
                    {
                        return true;
                    }
                    match g.family0.as_str() {
                        // The XMP properties ExifTool stores at priority 0 — the
                        // `PRIORITY => 0` mirror tables and every `Avoid => 1`
                        // property, which FoundTag demotes at ExifTool.pm:9472.
                        // See `scripts/gen_priority0.pl`.
                        "XMP" => {
                            crate::tags::priority0_generated::xmp_is_priority0(g1, name)
                                || crate::tags::group2::xmp_property_is_unknown(g1, name)
                        }
                        // A QuickTime track is NOT a sub-document: ProcessMOV
                        // only sets `$$et{SET_GROUP1} = 'Track'.++$track`
                        // (QuickTime.pm:10354) and never touches DOC_NUM for a
                        // track, so a track tag carries the normal priority and
                        // the LAST track wins a duplicate. The sole
                        // `PRIORITY => 0` table in QuickTime.pm is Bitrate
                        // (:1162, "often filled with zeros"), whose three tags
                        // are named here.
                        "QuickTime" => {
                            g1 == "QuickTime"
                                && matches!(name, "AverageBitrate" | "BufferSize" | "MaxBitrate")
                        }
                        // ExifTool's LOW_PRIORITY_DIR. SubIFDs are deliberately
                        // absent: ExifTool never demotes them, and a NEF's
                        // full-resolution SubIFD1 must win StripOffsets by last-wins.
                        "EXIF" | "MakerNotes" => g1 == "PreviewIFD" || (ifd1_low && g1 == "IFD1"),
                        // FujiFilm.pm:1270 declares the RAF directory table
                        // `PRIORITY => 0, # so the first RAF directory takes
                        // precedence`: a RAF file can hold two directories
                        // (header slots 0x5c and 0x78, FujiFilm.pm:1964), whose
                        // tags land in family-1 groups RAF and RAF2, and the
                        // first one wins when duplicates are collapsed.
                        "RAF" => true,
                        // IPTC.pm:1100 sets `$$et{LOW_PRIORITY_DIR}{IPTC} = 1`
                        // for an IPTC directory found outside the format's
                        // standard location, right where it numbers its family-1
                        // group (IPTC2, IPTC3, ...). The standard directory keeps
                        // the plain `IPTC` group and its normal priority.
                        "IPTC" => g1 != "IPTC",
                        // Matroska and MXF number their tracks too, but keep them
                        // all in the main document, where last-wins applies.
                        _ => matches!(g1, "Jpeg2000" | "PhotoMechanic" | "DjVu"),
                    }
                };
                // ExifTool's PRIORITY_DIR: Exif.pm 0xfe (SubfileType) and 0xff
                // (OldSubfileType) call `$self->SetPriorityDir()` when the directory
                // holds the full-resolution image, and SetPriorityDir
                // (ExifTool.pm:9636) keeps the FIRST one: `$$self{PRIORITY_DIR} =
                // $$self{DIR_NAME} unless $$self{PRIORITY_DIR}`. DIR_NAME is the
                // family-1 group name, which a live dump confirms (DNG.dng → SubIFD,
                // Nikon.nef → SubIFD1).
                let priority_dir: Option<String> = tags
                    .iter()
                    .find(|t| {
                        t.group.family0 == "EXIF"
                            && matches!(t.name.as_str(), "SubfileType" | "OldSubfileType")
                            && t.print_value == "Full-resolution image"
                    })
                    .map(|t| t.group.family1.clone());
                // QuickTime.pm:10016 — ProcessMOV runs `$$et{PRIORITY_DIR} = 'XMP'
                // unless $fileType and $fileType eq 'HEIC'` ("have XMP take
                // priority except for HEIC") before reading any box, and
                // SetPriorityDir only fills PRIORITY_DIR when it is still empty
                // (ExifTool.pm:9636), so XMP stays the priority directory for the
                // whole movie. Its effect is to promote back to 1 every XMP tag
                // ExifTool would otherwise store at priority 0.
                let xmp_is_priority_dir = matches!(
                    file_type,
                    FileType::Mp4
                        | FileType::QuickTime
                        | FileType::M4a
                        | FileType::ThreeGP
                        | FileType::Avif
                        | FileType::Cr3
                        | FileType::Crm
                        | FileType::F4v
                        | FileType::Mqv
                        | FileType::Lrv
                ) || (file_type == FileType::Heif && ft_code != "HEIC");
                use std::collections::HashMap as HM;
                // The competition is group-blind, exactly as in Perl: `$$self{VALUE}`
                // holds ONE entry per tag NAME, and FoundTag arbitrates every
                // incoming instance against it whatever directory it came from
                // (ExifTool.pm:9545-9570). So EXIF:FNumber and MakerNotes:FNumber
                // genuinely compete, and the loser is only reachable as `FNumber (1)`
                // with the Duplicates option on. The key is the name alone.
                let mut by_name: HM<&str, Vec<usize>> = HM::new();
                for (i, t) in tags.iter().enumerate() {
                    by_name.entry(t.name.as_str()).or_default().push(i);
                }
                let mut drop: std::collections::HashSet<usize> = std::collections::HashSet::new();
                for idxs in by_name.values() {
                    if idxs.len() < 2 {
                        continue;
                    }
                    // Replay FoundTag's priority comparison over the instances in
                    // extraction order, then keep only the surviving one. A tag
                    // carries priority 0 only when its source table declares it;
                    // the struct default of 0 means "unspecified", i.e. ExifTool's
                    // normal priority of 1.
                    let eff = |i: usize| -> i32 {
                        let t = &tags[i];
                        // Perl's DIR_NAME is the directory's own name, `XMP` for
                        // every XMP directory whatever family-1 group its
                        // properties end up in (XMP-dc, XMP-xmpDM, ...).
                        let in_priority_dir = priority_dir.as_deref()
                            == Some(t.group.family1.as_str())
                            || (xmp_is_priority_dir && t.group.family0 == "XMP");
                        // A priority the source table stated itself — an explicit
                        // `Priority => 0` or the `Avoid => 1` FoundTag turns into
                        // one. It bypasses the LOW_PRIORITY_DIR default and is
                        // promoted back to 1 only inside the PRIORITY_DIR
                        // (ExifTool.pm:9552-9555). A sub-document still never
                        // displaces the main document's tag.
                        // `XMP-pdf:Keywords => { Priority => -1 }`
                        // (XMP.pm line 1238), the one XMP property ExifTool puts
                        // below 0. Perl only ever promotes a priority that is
                        // FALSE, so this one is neither raised to 1 as a stored
                        // value (ExifTool.pm:9544-9551) nor promoted inside the
                        // PRIORITY_DIR (:9554): it can never take a name.
                        if t.group.family0 == "XMP"
                            && crate::tags::priority0_generated::xmp_is_below_priority0(
                                &t.group.family1,
                                &t.name,
                            )
                        {
                            return -1;
                        }
                        if t.priority == crate::tag::PRIORITY_EXPLICIT_ZERO {
                            if t.group.family3 != MAIN_DOCUMENT {
                                return 0;
                            }
                            return i32::from(in_priority_dir);
                        }
                        if t.priority == 0 && is_low_priority_source(&t.group, &t.name) {
                            // Only a table-stated `Priority => 0` is promoted in
                            // the priority directory: a LOW_PRIORITY_DIR default
                            // comes from the `elsif` branch (ExifTool.pm:9557),
                            // which never looks at PRIORITY_DIR. XMP is the one
                            // demotion here that FoundTag reaches through the
                            // `defined $priority` branch.
                            i32::from(
                                in_priority_dir
                                    && t.group.family0 == "XMP"
                                    && t.group.family3 == MAIN_DOCUMENT,
                            )
                        } else {
                            t.priority.max(1)
                        }
                    };
                    // `unless ($oldPriority) { ... $oldPriority = 1 }`
                    // (ExifTool.pm:9544-9551): a stored priority is promoted only
                    // when it is FALSE, so 0 becomes 1 and a negative one stays.
                    let promoted = |p: i32| if p == 0 { 1 } else { p };
                    let mut winner = idxs[0];
                    for &i in &idxs[1..] {
                        if eff(i) >= promoted(eff(winner)) {
                            winner = i;
                        }
                    }
                    for &i in idxs {
                        if i != winner {
                            drop.insert(i);
                        }
                    }
                }
                if !drop.is_empty() {
                    let mut i = 0usize;
                    tags.retain(|_| {
                        let keep = !drop.contains(&i);
                        i += 1;
                        keep
                    });
                }
            }
        }

        // Resolve the file-level pseudo-tags through their ExifTool table. This
        // runs last, after duplicate resolution, so re-grouping a tag can never
        // perturb which instance survives: the pass rewrites group assignment
        // only, never a tag's name or value. Family 3 is preserved, so a Warning
        // raised inside an embedded document stays in that document.
        for tag in &mut tags {
            if let Some((f0, f1, f2)) = file_level_group(&tag.name) {
                tag.group.family0 = f0.to_string();
                tag.group.family1 = f1.to_string();
                tag.group.family2 = f2.to_string();
            }
        }

        // Then re-derive family 2 from ExifTool's own group tables. Each format
        // reader picks a category as it goes, from partial knowledge; this pass
        // corrects it against the generated tables, keyed on the family 0/1 the
        // reader assigned. It runs after the pseudo-tag pass so those keep their
        // hand-picked groups, and after duplicate resolution for the same reason
        // as above: family 2 plays no part in choosing which tag survives, so
        // rewriting it here cannot move a name or a value.
        for tag in &mut tags {
            if file_level_group(&tag.name).is_some() {
                continue;
            }
            if let Some(f2) = crate::tags::group2::family2_for(
                &tag.group.family0,
                &tag.group.family1,
                &tag.name,
                &tag.group.family2,
            ) {
                if f2 != tag.group.family2 {
                    tag.group.family2 = f2.to_string();
                }
            }
        }

        // Match ExifTool's console sanitization (its `Printable`, the non-`-E`
        // path) exactly: control chars 0x01-0x1F and 0x7F become '.', NULs are
        // dropped, and *trailing* whitespace is trimmed (`s/\s+$//`) — leading
        // whitespace is preserved, as ExifTool preserves it. Runs on print
        // values only — raw values feed composites and `-n`. ASCII control chars
        // are single bytes in UTF-8, so accented/multibyte text (>= 0x80) is
        // untouched; this is a no-op for numeric print values.
        let is_ws = |c: char| c.is_ascii_whitespace();
        for tag in &mut tags {
            let pv = tag.print_value.as_str();
            let dirty = pv.ends_with(is_ws)
                || pv.chars().any(|c| {
                    let u = c as u32;
                    u == 0 || (0x01..=0x1f).contains(&u) || u == 0x7f
                });
            if !dirty {
                continue;
            }
            let mapped: String = pv
                .chars()
                .filter_map(|c| {
                    let u = c as u32;
                    if u == 0 {
                        None
                    } else if (0x01..=0x1f).contains(&u) || u == 0x7f {
                        Some('.')
                    } else {
                        Some(c)
                    }
                })
                .collect();
            tag.print_value = mapped.trim_end_matches(is_ws).to_string();
        }

        // Filter by requested tags if specified. A request is either a bare
        // tag name (`By-line`) or group-qualified (`IPTC:By-line`); the group
        // prefix matches any family 0-2 and `*` is a tag wildcard (`IPTC:*`).
        if !self.options.requested_tags.is_empty() {
            tags.retain(|t| {
                self.options
                    .requested_tags
                    .iter()
                    .any(|req| Self::tag_matches_request(t, req))
            });
        }

        Ok(tags)
    }

    /// Match a tag against a `-TAG` or `-GROUP:TAG` request (case-insensitive).
    /// The optional group prefix matches any of the tag's group families
    /// (0-2); a `*` tag name matches every tag (in the group, if given).
    fn tag_matches_request(tag: &Tag, request: &str) -> bool {
        let req = request.to_lowercase();
        let (group, name) = match req.split_once(':') {
            Some((g, n)) => (Some(g), n),
            None => (None, req.as_str()),
        };
        if name != "*" && tag.name.to_lowercase() != name {
            return false;
        }
        match group {
            None => true,
            Some(g) => {
                let grp = &tag.group;
                grp.family0.to_lowercase() == g
                    || grp.family1.to_lowercase() == g
                    || grp.family2.to_lowercase() == g
            }
        }
    }

    /// Format extracted tags into a simple name→value map.
    ///
    /// Handles duplicate tag names by appending group info.
    fn get_info(&self, tags: &[Tag]) -> ImageInfo {
        let mut info = ImageInfo::new();
        let mut seen: HashMap<String, (usize, i32)> = HashMap::new(); // (count, best priority)

        for tag in tags {
            let value = if self.options.print_conv {
                &tag.print_value
            } else {
                &tag.raw_value.to_display_string()
            };

            let entry = seen.entry(tag.name.clone()).or_insert((0, i32::MIN));
            entry.0 += 1;

            if entry.0 == 1 {
                entry.1 = tag.priority_rank();
                info.insert(tag.name.clone(), value.clone());
            } else if tag.priority_rank() > entry.1 {
                // Higher priority tag replaces the previous one
                entry.1 = tag.priority_rank();
                info.insert(tag.name.clone(), value.clone());
            } else if self.options.duplicates {
                let key = format!("{} [{}:{}]", tag.name, tag.group.family0, tag.group.family1);
                info.insert(key, value.clone());
            }
        }

        info
    }

    /// Detect file type from magic bytes and extension.
    fn detect_file_type(&self, data: &[u8], path: &Path) -> Result<FileType> {
        // Try magic bytes first
        let header_len = data.len().min(256);
        if let Some(ft) = file_type::detect_from_magic(&data[..header_len]) {
            // Override ICO to Font if extension is .dfont (Mac resource fork)
            if ft == FileType::Ico {
                if let Some(ext) = path.extension().and_then(|e| e.to_str()) {
                    if ext.eq_ignore_ascii_case("dfont") {
                        return Ok(FileType::Dfont);
                    }
                }
            }
            // Override JPEG to JPS if the file extension is .jps
            if ft == FileType::Jpeg {
                if let Some(ext) = path.extension().and_then(|e| e.to_str()) {
                    if ext.eq_ignore_ascii_case("jps") {
                        return Ok(FileType::Jps);
                    }
                }
            }
            // Override PLIST to AAE if extension is .aae
            if ft == FileType::Plist {
                if let Some(ext) = path.extension().and_then(|e| e.to_str()) {
                    if ext.eq_ignore_ascii_case("aae") {
                        return Ok(FileType::Aae);
                    }
                }
            }
            // Override XMP/XML to PLIST/AAE if extension is .plist or .aae
            if ft == FileType::Xmp || ft == FileType::Xml {
                if let Some(ext) = path.extension().and_then(|e| e.to_str()) {
                    if ext.eq_ignore_ascii_case("plist") {
                        return Ok(FileType::Plist);
                    }
                    if ext.eq_ignore_ascii_case("aae") {
                        return Ok(FileType::Aae);
                    }
                }
            }
            // Override to PhotoCD if extension is .pcd (file starts with 0xFF padding)
            if let Some(ext) = path.extension().and_then(|e| e.to_str()) {
                if ext.eq_ignore_ascii_case("pcd")
                    && data.len() >= 2056
                    && &data[2048..2055] == b"PCD_IPI"
                {
                    return Ok(FileType::PhotoCd);
                }
            }
            // Override MP3 to MPC/APE/WavPack if extension says otherwise
            if ft == FileType::Mp3 {
                if let Some(ext) = path.extension().and_then(|e| e.to_str()) {
                    if ext.eq_ignore_ascii_case("mpc") {
                        return Ok(FileType::Mpc);
                    }
                    if ext.eq_ignore_ascii_case("ape") {
                        return Ok(FileType::Ape);
                    }
                    if ext.eq_ignore_ascii_case("wv") {
                        return Ok(FileType::WavPack);
                    }
                }
            }
            // ASF is the container for WMV (video) and WMA (audio); refine by extension.
            if ft == FileType::Asf {
                if let Some(ext) = path.extension().and_then(|e| e.to_str()) {
                    if ext.eq_ignore_ascii_case("wmv") {
                        return Ok(FileType::Wmv);
                    }
                    if ext.eq_ignore_ascii_case("wma") {
                        return Ok(FileType::Wma);
                    }
                }
            }
            // Opus is an Ogg stream with the Opus codec.
            if ft == FileType::Ogg {
                if let Some(ext) = path.extension().and_then(|e| e.to_str()) {
                    if ext.eq_ignore_ascii_case("opus") {
                        return Ok(FileType::Opus);
                    }
                }
            }
            // TIFF magic covers many RAW variants (DNG, NEF, ARW, …); ExifTool refines
            // the type by extension since they share the TIFF structure.
            if ft == FileType::Tiff {
                if let Some(ext) = path.extension().and_then(|e| e.to_str()) {
                    if let Some(ext_ft) = file_type::detect_from_extension(ext) {
                        if ext_ft != FileType::Tiff && is_tiff_based(ext_ft) {
                            return Ok(ext_ft);
                        }
                    }
                }
            }
            // For ZIP files, check if it's an EIP (by extension) or OpenDocument format
            if ft == FileType::Zip {
                // Check extension first for EIP
                if let Some(ext) = path.extension().and_then(|e| e.to_str()) {
                    if ext.eq_ignore_ascii_case("eip") {
                        return Ok(FileType::Eip);
                    }
                }
                // iWork (KEY/PAGES/NUMBERS): ExifTool keys on the file extension once an
                // iWork marker member is present (ZIP.pm Process_iWork).
                if let Some(iw) = detect_iwork_type(data, path) {
                    return Ok(iw);
                }
                if let Some(od_type) = detect_opendocument_type(data) {
                    return Ok(od_type);
                }
            }
            // OLE2 compound files (DOC/XLS/PPT/FlashPix) all share the D0CF11E0 magic;
            // refine by the UTF-16 stream names in the directory.
            if ft == FileType::Doc {
                if let Some(ole) = detect_ole2_type(data) {
                    return Ok(ole);
                }
            }
            return Ok(ft);
        }

        // Fall back to extension
        if let Some(ext) = path.extension().and_then(|e| e.to_str()) {
            if let Some(ft) = file_type::detect_from_extension(ext) {
                return Ok(ft);
            }
        }

        let ext_str = path
            .extension()
            .and_then(|e| e.to_str())
            .unwrap_or("unknown");
        Err(Error::UnsupportedFileType(ext_str.to_string()))
    }

    /// Dispatch to the appropriate format reader.
    fn process_file(&self, data: &[u8], file_type: FileType) -> Result<Vec<Tag>> {
        match file_type {
            FileType::Jpeg | FileType::Jps => {
                formats::jpeg::read_jpeg_with_ee(data, self.options.extract_embedded)
            }
            FileType::Png | FileType::Mng => formats::png::read_png(data),
            // All TIFF-based formats (TIFF + most RAW formats)
            FileType::Tiff
            | FileType::Btf
            | FileType::Dng
            | FileType::Cr2
            | FileType::Nef
            | FileType::Arw
            | FileType::Sr2
            | FileType::Orf
            | FileType::Pef
            | FileType::Erf
            | FileType::Fff
            | FileType::Rwl
            | FileType::Mef
            | FileType::Srw
            | FileType::Gpr
            | FileType::Arq
            | FileType::ThreeFR
            | FileType::Dcr
            | FileType::Rw2
            | FileType::Srf => formats::tiff::read_tiff(data),
            // Phase One IIQ: TIFF + PhaseOne maker note block
            FileType::Iiq => formats::iiq::read_iiq(
                data,
                !self.options.duplicates && self.options.extract_embedded == 0,
            ),
            // Image formats
            FileType::Gif => formats::gif::read_gif(data),
            FileType::Bmp => formats::bmp::read_bmp(data),
            FileType::WebP | FileType::Avi | FileType::Wav => formats::riff::read_riff(data),
            FileType::Psd => formats::psd::read_psd(data),
            // Audio formats
            FileType::Mp3 => formats::id3::read_mp3(data),
            FileType::Flac => formats::flac::read_flac(data),
            FileType::Ogg | FileType::Opus => formats::ogg::read_ogg(data),
            FileType::Aiff => formats::aiff::read_aiff(data),
            // Video formats
            FileType::Mp4
            | FileType::QuickTime
            | FileType::M4a
            | FileType::ThreeGP
            | FileType::Heif
            | FileType::Avif
            | FileType::Cr3
            | FileType::Crm
            | FileType::F4v
            | FileType::Mqv
            | FileType::Lrv => {
                formats::quicktime::read_quicktime_with_ee(data, self.options.extract_embedded)
            }
            FileType::Mkv | FileType::WebM => formats::matroska::read_matroska(data),
            FileType::Asf | FileType::Wmv | FileType::Wma => formats::asf::read_asf(data),
            FileType::Wtv => formats::wtv::read_wtv(data),
            // RAW formats with custom containers
            FileType::Crw => formats::canon_raw::read_crw(data),
            FileType::Raf => formats::raf::read_raf(data),
            FileType::Mrw => formats::mrw::read_mrw(data),
            FileType::Mrc => formats::mrc::read_mrc(data, self.options.extract_embedded),
            // Image formats
            FileType::Jp2 => formats::jp2::read_jp2(data),
            FileType::J2c => formats::jp2::read_j2c(data),
            FileType::Jxl => formats::jp2::read_jxl(data),
            FileType::Ico => formats::ico::read_ico(data),
            FileType::Icc => formats::icc::read_icc(data),
            // Documents
            FileType::Pdf => formats::pdf::read_pdf(data, self.options.extract_embedded),
            FileType::PostScript => {
                // PFA fonts start with %!PS-AdobeFont or %!FontType1
                if data.starts_with(b"%!PS-AdobeFont") || data.starts_with(b"%!FontType1") {
                    formats::font::read_pfa(data).or_else(|_| {
                        formats::postscript::read_postscript(data, self.options.extract_embedded)
                    })
                } else {
                    formats::postscript::read_postscript(data, self.options.extract_embedded)
                }
            }
            FileType::Eip => formats::capture_one::read_eip(data, self.options.extract_embedded),
            FileType::Zip
            | FileType::Docx
            | FileType::Xlsx
            | FileType::Pptx
            | FileType::Doc
            | FileType::Xls
            | FileType::Ppt
            | FileType::Numbers
            | FileType::Pages
            | FileType::Key => formats::zip::read_zip(data, self.options.extract_embedded),
            FileType::Rtf => formats::rtf::read_rtf(data),
            FileType::InDesign => formats::indesign::read_indesign(data),
            FileType::Pcap => formats::pcap::read_pcap(data),
            FileType::Pcapng => formats::pcap::read_pcapng(data),
            // Canon VRD / DR4
            FileType::Vrd => formats::canon_vrd::read_vrd(data).or_else(|_| Ok(Vec::new())),
            FileType::Dr4 => formats::canon_vrd::read_dr4(data).or_else(|_| Ok(Vec::new())),
            // Metadata / Other
            FileType::Xmp => formats::xmp_file::read_xmp(data),
            FileType::Svg => formats::svg::read_svg(data),
            FileType::Html => {
                // SVG files that weren't detected by magic (e.g., via extension fallback)
                let is_svg = data.windows(4).take(512).any(|w| w == b"<svg");
                if is_svg {
                    formats::svg::read_svg(data)
                } else {
                    formats::html::read_html(data)
                }
            }
            FileType::Exe => formats::exe::read_exe(data),
            FileType::Font => {
                // AFM: Adobe Font Metrics text file
                if data.starts_with(b"StartFontMetrics") {
                    return formats::font::read_afm(data);
                }
                // PFA: PostScript Type 1 ASCII font
                if data.starts_with(b"%!PS-AdobeFont") || data.starts_with(b"%!FontType1") {
                    return formats::font::read_pfa(data).or_else(|_| Ok(Vec::new()));
                }
                // PFB: PostScript Type 1 Binary font
                if data.len() >= 2 && data[0] == 0x80 && (data[1] == 0x01 || data[1] == 0x02) {
                    return formats::font::read_pfb(data).or_else(|_| Ok(Vec::new()));
                }
                formats::font::read_font(data)
            }
            // Audio with ID3
            FileType::WavPack | FileType::Dsf => formats::id3::read_mp3(data),
            FileType::Ape => formats::ape::read_ape(data),
            FileType::Mpc => formats::ape::read_mpc(data),
            FileType::Aac => formats::aac::read_aac(data),
            FileType::RealAudio => {
                formats::real_audio::read_real_audio(data).or_else(|_| Ok(Vec::new()))
            }
            FileType::RealMedia => {
                formats::real_media::read_real_media(data).or_else(|_| Ok(Vec::new()))
            }
            // Misc formats
            FileType::Czi => formats::czi::read_czi(data).or_else(|_| Ok(Vec::new())),
            FileType::PhotoCd => formats::photo_cd::read_photo_cd(data).or_else(|_| Ok(Vec::new())),
            FileType::Dicom => formats::dicom::read_dicom(data),
            FileType::Fits => formats::fits::read_fits(data),
            FileType::Fit => formats::fit::read_fit_with_ee(data, self.options.extract_embedded),
            FileType::Flv => formats::flv::read_flv(data),
            FileType::Mxf => formats::mxf::read_mxf(data, self.options.extract_embedded)
                .or_else(|_| Ok(Vec::new())),
            FileType::Swf => formats::swf::read_swf(data),
            FileType::Hdr => formats::hdr::read_hdr(data),
            FileType::DjVu => formats::djvu::read_djvu(data),
            FileType::Xcf => formats::gimp::read_xcf(data),
            FileType::Mie => formats::mie::read_mie(data),
            FileType::Lfp => formats::lytro::read_lfp(data),
            // FileType::Miff dispatched via string extension below
            FileType::Fpf => formats::flir_fpf::read_fpf(data),
            FileType::Flif => formats::flif::read_flif(data),
            FileType::Bpg => formats::bpg::read_bpg(data),
            FileType::Pcx => formats::pcx::read_pcx(data),
            FileType::Pict => formats::pict::read_pict(data),
            FileType::Mpeg => formats::mpeg::read_mpeg(data),
            FileType::M2ts => formats::m2ts::read_m2ts(data, self.options.extract_embedded),
            FileType::Gzip => formats::gzip::read_gzip(data),
            FileType::Rar => formats::rar::read_rar(data),
            FileType::SevenZ => formats::sevenz::read_7z(data),
            FileType::Dss => formats::dss::read_dss(data),
            FileType::Moi => formats::moi::read_moi(data),
            FileType::MacOs => formats::macos::read_macos(data),
            FileType::Json => formats::json_format::read_json(data),
            // New formats
            FileType::Pgf => formats::pgf::read_pgf(data),
            FileType::Xisf => formats::xisf::read_xisf(data),
            FileType::Torrent => formats::torrent::read_torrent(data),
            FileType::Mobi => formats::palm::read_palm(data),
            FileType::Psp => formats::psp::read_psp(data),
            FileType::SonyPmp => formats::sony_pmp::read_sony_pmp(data),
            FileType::Audible => formats::audible::read_audible(data),
            FileType::Exr => formats::openexr::read_openexr(data),
            // New formats
            FileType::Plist => {
                if data.starts_with(b"bplist") {
                    formats::plist::read_binary_plist_tags(data)
                } else {
                    formats::plist::read_xml_plist(data)
                }
            }
            FileType::Aae => {
                if data.starts_with(b"bplist") {
                    formats::plist::read_binary_plist_tags(data)
                } else {
                    formats::plist::read_aae_plist(data)
                }
            }
            FileType::KyoceraRaw => formats::kyocera_raw::read_kyocera_raw(data),
            FileType::PortableFloatMap => formats::pfm::read_pfm(data),
            FileType::Ods
            | FileType::Odt
            | FileType::Odp
            | FileType::Odg
            | FileType::Odf
            | FileType::Odb
            | FileType::Odi
            | FileType::Odc => formats::zip::read_zip(data, self.options.extract_embedded),
            FileType::Lif => formats::lif::read_lif(data),
            FileType::Rwz => formats::rawzor::read_rawzor(data),
            FileType::Jxr => formats::jxr::read_jxr(data),
            FileType::Miff => formats::miff::read_miff(data).or_else(|_| Ok(Vec::new())),
            FileType::Tnef => formats::tnef::read_tnef(data).or_else(|_| Ok(Vec::new())),
            FileType::Wpg => formats::wpg::read_wpg(data).or_else(|_| Ok(Vec::new())),
            FileType::Dv => {
                formats::dv::read_dv(data, data.len() as u64).or_else(|_| Ok(Vec::new()))
            }
            FileType::Itc => formats::itc::read_itc(data).or_else(|_| Ok(Vec::new())),
            FileType::Iso => formats::iso::read_iso(data).or_else(|_| Ok(Vec::new())),
            FileType::Afm => formats::font::read_afm(data).or_else(|_| Ok(Vec::new())),
            FileType::Pfa => formats::font::read_pfa(data).or_else(|_| Ok(Vec::new())),
            FileType::Pfb => formats::font::read_pfb(data).or_else(|_| Ok(Vec::new())),
            FileType::Dfont => formats::font::read_font(data).or_else(|_| Ok(Vec::new())),
            FileType::Xml | FileType::Inx => {
                formats::xmp_file::read_xmp(data).or_else(|_| Ok(Vec::new()))
            }
            FileType::Eps => {
                formats::postscript::read_postscript(data, self.options.extract_embedded)
            }
            _ => Err(Error::UnsupportedFileType(format!("{}", file_type))),
        }
    }

    /// Fallback: try to read file based on extension for formats without magic detection.
    fn process_by_extension(&self, data: &[u8], path: &Path) -> Result<Vec<Tag>> {
        let ext = path
            .extension()
            .and_then(|e| e.to_str())
            .unwrap_or("")
            .to_ascii_lowercase();

        match ext.as_str() {
            "ppm" | "pgm" | "pbm" => formats::ppm::read_ppm(data),
            "pfm" => {
                // PFM can be Portable Float Map or Printer Font Metrics
                if data.len() >= 3 && data[0] == b'P' && (data[1] == b'f' || data[1] == b'F') {
                    formats::ppm::read_ppm(data)
                } else {
                    Ok(Vec::new()) // Printer Font Metrics
                }
            }
            "json" => formats::json_format::read_json(data),
            "svg" => formats::svg::read_svg(data),
            "ram" => formats::ram::read_ram(data).or_else(|_| Ok(Vec::new())),
            "txt" | "log" | "igc" => Ok(compute_text_tags(data, false)),
            "csv" => Ok(compute_text_tags(data, true)),
            "url" => formats::lnk::read_url(data).or_else(|_| Ok(Vec::new())),
            "lnk" => formats::lnk::read_lnk(data).or_else(|_| Ok(Vec::new())),
            "gpx" | "kml" | "xml" | "inx" => formats::xmp_file::read_xmp(data),
            "plist" => {
                if data.starts_with(b"bplist") {
                    formats::plist::read_binary_plist_tags(data).or_else(|_| Ok(Vec::new()))
                } else {
                    formats::plist::read_xml_plist(data).or_else(|_| Ok(Vec::new()))
                }
            }
            "aae" => {
                if data.starts_with(b"bplist") {
                    formats::plist::read_binary_plist_tags(data).or_else(|_| Ok(Vec::new()))
                } else {
                    formats::plist::read_aae_plist(data).or_else(|_| Ok(Vec::new()))
                }
            }
            "vcf" | "ics" | "vcard" => {
                let s = crate::encoding::decode_utf8_or_latin1(&data[..data.len().min(100)]);
                if s.contains("BEGIN:VCALENDAR") {
                    formats::vcard::read_ics(data).or_else(|_| Ok(Vec::new()))
                } else {
                    formats::vcard::read_vcf(data).or_else(|_| Ok(Vec::new()))
                }
            }
            "xcf" => Ok(Vec::new()), // GIMP
            "vrd" => formats::canon_vrd::read_vrd(data).or_else(|_| Ok(Vec::new())),
            "dr4" => formats::canon_vrd::read_dr4(data).or_else(|_| Ok(Vec::new())),
            "indd" | "indt" => Ok(Vec::new()), // InDesign
            "x3f" => formats::sigma_raw::read_x3f(data).or_else(|_| Ok(Vec::new())),
            "mie" => Ok(Vec::new()), // MIE
            "exr" => Ok(Vec::new()), // OpenEXR
            "wpg" => formats::wpg::read_wpg(data).or_else(|_| Ok(Vec::new())),
            "moi" => formats::moi::read_moi(data).or_else(|_| Ok(Vec::new())),
            "macos" => formats::macos::read_macos(data).or_else(|_| Ok(Vec::new())),
            "dpx" => formats::dpx::read_dpx(data).or_else(|_| Ok(Vec::new())),
            "r3d" => formats::red::read_r3d(data).or_else(|_| Ok(Vec::new())),
            "tnef" => formats::tnef::read_tnef(data).or_else(|_| Ok(Vec::new())),
            "ppt" | "fpx" => formats::flashpix::read_fpx(data).or_else(|_| Ok(Vec::new())),
            "fpf" => formats::flir_fpf::read_fpf(data).or_else(|_| Ok(Vec::new())),
            "itc" => formats::itc::read_itc(data).or_else(|_| Ok(Vec::new())),
            "mpg" | "mpeg" | "m1v" | "m2v" | "mpv" => {
                formats::mpeg::read_mpeg(data).or_else(|_| Ok(Vec::new()))
            }
            "dv" => formats::dv::read_dv(data, data.len() as u64).or_else(|_| Ok(Vec::new())),
            "czi" => formats::czi::read_czi(data).or_else(|_| Ok(Vec::new())),
            "miff" => formats::miff::read_miff(data).or_else(|_| Ok(Vec::new())),
            "lfp" | "mrc" | "dss" | "mobi" | "psp" | "pgf" | "raw" | "pmp" | "torrent" | "xisf"
            | "mxf" | "dfont" => Ok(Vec::new()),
            "iso" => formats::iso::read_iso(data).or_else(|_| Ok(Vec::new())),
            "afm" => formats::font::read_afm(data).or_else(|_| Ok(Vec::new())),
            "pfa" => formats::font::read_pfa(data).or_else(|_| Ok(Vec::new())),
            "pfb" => formats::font::read_pfb(data).or_else(|_| Ok(Vec::new())),
            _ => Err(Error::UnsupportedFileType(ext)),
        }
    }
}

impl Default for ExifTool {
    fn default() -> Self {
        Self::new()
    }
}

/// Detect OpenDocument file type by reading the `mimetype` entry from a ZIP.
/// Returns None if not an OpenDocument file.
/// Refine an EXE file's (FileType, MIMEType, FileTypeExtension) from its magic, mirroring
/// ExifTool's EXE SetFileType. MIME is always application/octet-stream for these.
fn exe_subtype(d: &[u8]) -> Option<(&'static str, &'static str, &'static str)> {
    const MIME: &str = "application/octet-stream";
    if d.len() < 8 {
        return None;
    }
    // ELF: 0x7F 'E' 'L' 'F'; data[5] endianness (1=LE,2=BE); e_type at offset 16 (2 bytes)
    if &d[0..4] == b"\x7fELF" && d.len() >= 18 {
        let le = d[5] == 1;
        let e_type = if le {
            u16::from_le_bytes([d[16], d[17]])
        } else {
            u16::from_be_bytes([d[16], d[17]])
        };
        return Some(match e_type {
            1 => ("ELF relocatable", MIME, "o"),
            2 => ("ELF executable", MIME, ""),
            3 => ("ELF shared library", MIME, "so"),
            4 => ("ELF core file", MIME, ""),
            _ => ("ELF", MIME, ""),
        });
    }
    // Mach-O thin binary: magic FEEDFACE/FEEDFACF (BE) or CEFAEDFE/CFFAEDFE (LE).
    let magic_be = u32::from_be_bytes([d[0], d[1], d[2], d[3]]);
    let macho = matches!(magic_be, 0xFEEDFACE | 0xFEEDFACF | 0xCEFAEDFE | 0xCFFAEDFE);
    if macho && d.len() >= 16 {
        let le = matches!(magic_be, 0xCEFAEDFE | 0xCFFAEDFE);
        let filetype = if le {
            u32::from_le_bytes([d[12], d[13], d[14], d[15]])
        } else {
            u32::from_be_bytes([d[12], d[13], d[14], d[15]])
        };
        return Some(match filetype {
            1 => ("Mach-O object file", MIME, "o"),
            6 => ("Mach-O dynamic link library", MIME, "dylib"),
            8 => ("Mach-O dynamic bound bundle", MIME, "dylib"),
            9 => ("Mach-O dynamic link library stub", MIME, "dylib"),
            _ => ("Mach-O executable", MIME, ""),
        });
    }
    // Mach-O fat binary: CAFEBABE / BEBAFECA
    if matches!(magic_be, 0xCAFEBABE | 0xBEBAFECA) {
        return Some(("Mach-O fat binary executable", MIME, ""));
    }
    // ar archive ("!<arch>\n"): static library (Mach-O if it contains Mach-O members).
    if d.starts_with(b"!<arch>\n") {
        let is_macho = d.windows(4).take(4096).any(|w| {
            let m = u32::from_be_bytes([w[0], w[1], w[2], w[3]]);
            matches!(
                m,
                0xFEEDFACE | 0xFEEDFACF | 0xCEFAEDFE | 0xCFFAEDFE | 0xCAFEBABE
            )
        });
        return Some(if is_macho {
            ("Mach-O static library", MIME, "a")
        } else {
            ("Static library", MIME, "a")
        });
    }
    // PE (Windows): "MZ" then PE header; machine field selects Win32/Win64.
    if &d[0..2] == b"MZ" && d.len() >= 0x40 {
        let pe_off = u32::from_le_bytes([d[0x3c], d[0x3d], d[0x3e], d[0x3f]]) as usize;
        if pe_off + 6 <= d.len() && &d[pe_off..pe_off + 4] == b"PE\0\0" {
            let machine = u16::from_le_bytes([d[pe_off + 4], d[pe_off + 5]]);
            return Some(match machine {
                0x8664 | 0xAA64 => ("Win64 EXE", MIME, "exe"),
                _ => ("Win32 EXE", MIME, "exe"),
            });
        }
    }
    None
}

/// Whether a FileType is a TIFF-based RAW variant (shares TIFF magic, refined by extension).
fn is_tiff_based(ft: FileType) -> bool {
    matches!(
        ft,
        FileType::Dng
            | FileType::Cr2
            | FileType::Nef
            | FileType::Arw
            | FileType::Sr2
            | FileType::Orf
            | FileType::Pef
            | FileType::Erf
            | FileType::Rwl
            | FileType::Mef
            | FileType::Srw
            | FileType::Gpr
            | FileType::Arq
            | FileType::ThreeFR
            | FileType::Dcr
            | FileType::Rw2
            | FileType::Srf
            | FileType::Iiq
            | FileType::Btf
    )
}

/// Refine an OLE2 compound document (DOC/XLS/PPT) by scanning the directory for
/// well-known UTF-16LE stream names. Returns None (→ keep DOC) when none match.
fn detect_ole2_type(data: &[u8]) -> Option<FileType> {
    fn has_utf16(data: &[u8], name: &str) -> bool {
        let needle: Vec<u8> = name.encode_utf16().flat_map(|u| u.to_le_bytes()).collect();
        data.windows(needle.len()).any(|w| w == needle.as_slice())
    }
    if has_utf16(data, "PowerPoint Document") {
        Some(FileType::Ppt)
    } else if has_utf16(data, "Workbook") || has_utf16(data, "Book") {
        Some(FileType::Xls)
    } else {
        None
    }
}

/// Detect an iWork (KEY/PAGES/NUMBERS) ZIP. ExifTool recognises these by the
/// presence of an iWork marker member, then maps the file type from the
/// extension (ZIP.pm `%iWorkType` / Process_iWork).
fn detect_iwork_type(data: &[u8], path: &Path) -> Option<FileType> {
    const MARKERS: &[&[u8]] = &[
        b"index.xml",
        b"index.apxl",
        b"QuickLook/Thumbnail.jpg",
        b"Index/Document.iwa",
        b"Index/Slide.iwa",
        b"Index/Tables/DataList.iwa",
    ];
    let has_marker = MARKERS
        .iter()
        .any(|m| data.windows(m.len()).any(|w| w == *m));
    if !has_marker {
        return None;
    }
    let ext = path
        .extension()
        .and_then(|e| e.to_str())
        .unwrap_or("")
        .to_ascii_lowercase();
    match ext.as_str() {
        "numbers" | "nmbtemplate" => Some(FileType::Numbers),
        "pages" => Some(FileType::Pages),
        "key" | "kth" => Some(FileType::Key),
        _ => None,
    }
}

/// Content-dependent FileType code / MIME refinements (ExifTool SetFileType with a
/// content test). Returns (code, mime); the extension keeps its default.
fn refine_filetype_by_content(file_type: FileType, data: &[u8]) -> Option<(String, String)> {
    match file_type {
        // Printer Font Metrics (font, starts 0x00 0x01/0x02) vs Portable Float Map (image, "PF").
        FileType::PortableFloatMap if data.len() >= 2 && data[0] == 0x00 && data[1] <= 0x02 => {
            Some(("PFM".into(), "application/x-font-type1".into()))
        }
        // XML property list → application/xml (binary plist keeps application/x-plist).
        FileType::Plist if !data.starts_with(b"bplist") => {
            Some(("PLIST".into(), "application/xml".into()))
        }
        // Naked JPEG XL codestream (FF 0A) vs the ISOBMFF container.
        FileType::Jxl if data.starts_with(&[0xFF, 0x0A]) => {
            Some(("JXL Codestream".into(), file_type.mime_type().to_string()))
        }
        // Extended WebP: VP8X chunk at offset 12.
        FileType::WebP if data.len() >= 16 && &data[12..16] == b"VP8X" => {
            Some(("Extended WEBP".into(), file_type.mime_type().to_string()))
        }
        // Multi-page DjVu: "DJVM" form type at offset 12.
        FileType::DjVu if data.len() >= 16 && &data[12..16] == b"DJVM" => Some((
            "DJVU (multi-page)".into(),
            file_type.mime_type().to_string(),
        )),
        _ => None,
    }
}

fn detect_opendocument_type(data: &[u8]) -> Option<FileType> {
    // OpenDocument ZIPs have "mimetype" as the FIRST local file entry (uncompressed)
    if data.len() < 30 || data[0..4] != [0x50, 0x4B, 0x03, 0x04] {
        return None;
    }
    let compression = u16::from_le_bytes([data[8], data[9]]);
    let compressed_size = u32::from_le_bytes([data[18], data[19], data[20], data[21]]) as usize;
    let name_len = u16::from_le_bytes([data[26], data[27]]) as usize;
    let extra_len = u16::from_le_bytes([data[28], data[29]]) as usize;
    let name_start = 30;
    if name_start + name_len > data.len() {
        return None;
    }
    let filename = std::str::from_utf8(&data[name_start..name_start + name_len]).unwrap_or("");
    if filename != "mimetype" || compression != 0 {
        return None;
    }
    let content_start = name_start + name_len + extra_len;
    let content_end = (content_start + compressed_size).min(data.len());
    if content_start >= content_end {
        return None;
    }
    let mime = std::str::from_utf8(&data[content_start..content_end])
        .unwrap_or("")
        .trim();
    match mime {
        "application/vnd.oasis.opendocument.spreadsheet" => Some(FileType::Ods),
        "application/vnd.oasis.opendocument.text" => Some(FileType::Odt),
        "application/vnd.oasis.opendocument.presentation" => Some(FileType::Odp),
        "application/vnd.oasis.opendocument.graphics" => Some(FileType::Odg),
        "application/vnd.oasis.opendocument.formula" => Some(FileType::Odf),
        "application/vnd.oasis.opendocument.database" => Some(FileType::Odb),
        "application/vnd.oasis.opendocument.image" => Some(FileType::Odi),
        "application/vnd.oasis.opendocument.chart" => Some(FileType::Odc),
        _ => None,
    }
}

/// Detect the file type of a file at the given path.
pub fn get_file_type<P: AsRef<Path>>(path: P) -> Result<FileType> {
    let path = path.as_ref();
    let mut file = fs::File::open(path).map_err(Error::Io)?;
    let mut header = [0u8; 256];
    use std::io::Read;
    let n = file.read(&mut header).map_err(Error::Io)?;

    if let Some(ft) = file_type::detect_from_magic(&header[..n]) {
        return Ok(ft);
    }

    if let Some(ext) = path.extension().and_then(|e| e.to_str()) {
        if let Some(ft) = file_type::detect_from_extension(ext) {
            return Ok(ft);
        }
    }

    Err(Error::UnsupportedFileType("unknown".into()))
}

/// Classification of EXIF tags into IFD groups.
enum ExifIfdGroup {
    Ifd0,
    ExifIfd,
    Gps,
}

/// Determine which IFD a tag belongs to based on its ID.
fn classify_exif_tag(tag_id: u16) -> ExifIfdGroup {
    match tag_id {
        // ExifIFD tags
        0x829A..=0x829D | 0x8822..=0x8827 | 0x8830 | 0x9000..=0x9292 | 0xA000..=0xA435 => {
            ExifIfdGroup::ExifIfd
        }
        // GPS tags
        0x0000..=0x001F if tag_id <= 0x001F => ExifIfdGroup::Gps,
        // Everything else → IFD0
        _ => ExifIfdGroup::Ifd0,
    }
}

/// Extract existing EXIF entries from a JPEG file's APP1 segment.
fn extract_existing_exif_entries(
    jpeg_data: &[u8],
    target_bo: ByteOrderMark,
) -> Vec<exif_writer::IfdEntry> {
    let mut entries = Vec::new();

    // Find EXIF APP1 segment
    let mut pos = 2; // Skip SOI
    while pos + 4 <= jpeg_data.len() {
        if jpeg_data[pos] != 0xFF {
            pos += 1;
            continue;
        }
        let marker = jpeg_data[pos + 1];
        pos += 2;

        if marker == 0xDA || marker == 0xD9 {
            break; // SOS or EOI
        }
        if marker == 0xFF || marker == 0x00 || marker == 0xD8 || (0xD0..=0xD7).contains(&marker) {
            continue;
        }

        if pos + 2 > jpeg_data.len() {
            break;
        }
        let seg_len = u16::from_be_bytes([jpeg_data[pos], jpeg_data[pos + 1]]) as usize;
        if seg_len < 2 || pos + seg_len > jpeg_data.len() {
            break;
        }

        let seg_data = &jpeg_data[pos + 2..pos + seg_len];

        // EXIF APP1
        if marker == 0xE1 && seg_data.len() > 14 && seg_data.starts_with(b"Exif\0\0") {
            let tiff_data = &seg_data[6..];
            extract_ifd_entries(tiff_data, target_bo, &mut entries);
            break;
        }

        pos += seg_len;
    }

    entries
}

/// Extract IFD entries from TIFF data, re-encoding values in the target byte order.
fn extract_ifd_entries(
    tiff_data: &[u8],
    target_bo: ByteOrderMark,
    entries: &mut Vec<exif_writer::IfdEntry>,
) {
    use crate::metadata::exif::parse_tiff_header;

    let header = match parse_tiff_header(tiff_data) {
        Ok(h) => h,
        Err(_) => return,
    };

    let src_bo = header.byte_order;

    // Read IFD0
    read_ifd_for_merge(
        tiff_data,
        header.ifd0_offset as usize,
        src_bo,
        target_bo,
        entries,
    );

    // Find ExifIFD and GPS pointers
    let ifd0_offset = header.ifd0_offset as usize;
    if ifd0_offset + 2 > tiff_data.len() {
        return;
    }
    let count = read_u16_bo(tiff_data, ifd0_offset, src_bo) as usize;
    for i in 0..count {
        let eoff = ifd0_offset + 2 + i * 12;
        if eoff + 12 > tiff_data.len() {
            break;
        }
        let tag = read_u16_bo(tiff_data, eoff, src_bo);
        let value_off = read_u32_bo(tiff_data, eoff + 8, src_bo) as usize;

        match tag {
            0x8769 => read_ifd_for_merge(tiff_data, value_off, src_bo, target_bo, entries),
            0x8825 => read_ifd_for_merge(tiff_data, value_off, src_bo, target_bo, entries),
            _ => {}
        }
    }
}

/// Read a single IFD and extract entries for merge.
fn read_ifd_for_merge(
    data: &[u8],
    offset: usize,
    src_bo: ByteOrderMark,
    target_bo: ByteOrderMark,
    entries: &mut Vec<exif_writer::IfdEntry>,
) {
    if offset + 2 > data.len() {
        return;
    }
    let count = read_u16_bo(data, offset, src_bo) as usize;

    for i in 0..count {
        let eoff = offset + 2 + i * 12;
        if eoff + 12 > data.len() {
            break;
        }

        let tag = read_u16_bo(data, eoff, src_bo);
        let dtype = read_u16_bo(data, eoff + 2, src_bo);
        let count_val = read_u32_bo(data, eoff + 4, src_bo);

        // Skip sub-IFD pointers and MakerNote
        if tag == 0x8769 || tag == 0x8825 || tag == 0xA005 || tag == 0x927C {
            continue;
        }

        let type_size = match dtype {
            1 | 2 | 6 | 7 => 1usize,
            3 | 8 => 2,
            4 | 9 | 11 | 13 => 4,
            5 | 10 | 12 => 8,
            _ => continue,
        };

        let total_size = type_size * count_val as usize;
        let raw_data = if total_size <= 4 {
            data[eoff + 8..eoff + 12].to_vec()
        } else {
            let voff = read_u32_bo(data, eoff + 8, src_bo) as usize;
            if voff + total_size > data.len() {
                continue;
            }
            data[voff..voff + total_size].to_vec()
        };

        // Re-encode multi-byte values if byte orders differ
        let final_data = if src_bo != target_bo && type_size > 1 {
            reencode_bytes(&raw_data, dtype, count_val as usize, src_bo, target_bo)
        } else {
            raw_data[..total_size].to_vec()
        };

        let format = match dtype {
            1 => exif_writer::ExifFormat::Byte,
            2 => exif_writer::ExifFormat::Ascii,
            3 => exif_writer::ExifFormat::Short,
            4 => exif_writer::ExifFormat::Long,
            5 => exif_writer::ExifFormat::Rational,
            6 => exif_writer::ExifFormat::SByte,
            7 => exif_writer::ExifFormat::Undefined,
            8 => exif_writer::ExifFormat::SShort,
            9 => exif_writer::ExifFormat::SLong,
            10 => exif_writer::ExifFormat::SRational,
            11 => exif_writer::ExifFormat::Float,
            12 => exif_writer::ExifFormat::Double,
            _ => continue,
        };

        entries.push(exif_writer::IfdEntry {
            tag,
            format,
            data: final_data,
        });
    }
}

/// Re-encode multi-byte values when converting between byte orders.
fn reencode_bytes(
    data: &[u8],
    dtype: u16,
    count: usize,
    src_bo: ByteOrderMark,
    dst_bo: ByteOrderMark,
) -> Vec<u8> {
    let mut out = Vec::with_capacity(data.len());
    match dtype {
        3 | 8 => {
            // 16-bit
            for i in 0..count {
                let v = read_u16_bo(data, i * 2, src_bo);
                match dst_bo {
                    ByteOrderMark::LittleEndian => out.extend_from_slice(&v.to_le_bytes()),
                    ByteOrderMark::BigEndian => out.extend_from_slice(&v.to_be_bytes()),
                }
            }
        }
        4 | 9 | 11 | 13 => {
            // 32-bit
            for i in 0..count {
                let v = read_u32_bo(data, i * 4, src_bo);
                match dst_bo {
                    ByteOrderMark::LittleEndian => out.extend_from_slice(&v.to_le_bytes()),
                    ByteOrderMark::BigEndian => out.extend_from_slice(&v.to_be_bytes()),
                }
            }
        }
        5 | 10 => {
            // Rational (two 32-bit)
            for i in 0..count {
                let n = read_u32_bo(data, i * 8, src_bo);
                let d = read_u32_bo(data, i * 8 + 4, src_bo);
                match dst_bo {
                    ByteOrderMark::LittleEndian => {
                        out.extend_from_slice(&n.to_le_bytes());
                        out.extend_from_slice(&d.to_le_bytes());
                    }
                    ByteOrderMark::BigEndian => {
                        out.extend_from_slice(&n.to_be_bytes());
                        out.extend_from_slice(&d.to_be_bytes());
                    }
                }
            }
        }
        12 => {
            // 64-bit double
            for i in 0..count {
                let mut bytes = [0u8; 8];
                bytes.copy_from_slice(&data[i * 8..i * 8 + 8]);
                if src_bo != dst_bo {
                    bytes.reverse();
                }
                out.extend_from_slice(&bytes);
            }
        }
        _ => out.extend_from_slice(data),
    }
    out
}

fn read_u16_bo(data: &[u8], offset: usize, bo: ByteOrderMark) -> u16 {
    if offset + 2 > data.len() {
        return 0;
    }
    match bo {
        ByteOrderMark::LittleEndian => u16::from_le_bytes([data[offset], data[offset + 1]]),
        ByteOrderMark::BigEndian => u16::from_be_bytes([data[offset], data[offset + 1]]),
    }
}

fn read_u32_bo(data: &[u8], offset: usize, bo: ByteOrderMark) -> u32 {
    if offset + 4 > data.len() {
        return 0;
    }
    match bo {
        ByteOrderMark::LittleEndian => u32::from_le_bytes([
            data[offset],
            data[offset + 1],
            data[offset + 2],
            data[offset + 3],
        ]),
        ByteOrderMark::BigEndian => u32::from_be_bytes([
            data[offset],
            data[offset + 1],
            data[offset + 2],
            data[offset + 3],
        ]),
    }
}

/// Map tag name to numeric EXIF tag ID.
fn tag_name_to_id(name: &str) -> Option<u16> {
    encode_exif_tag(name, "", "", ByteOrderMark::BigEndian).map(|(id, _, _)| id)
}

/// Convert a tag value to a safe filename.
fn value_to_filename(value: &str) -> String {
    value
        .chars()
        .map(|c| match c {
            '/' | '\\' | ':' | '*' | '?' | '"' | '<' | '>' | '|' => '_',
            c if c.is_control() => '_',
            c => c,
        })
        .collect::<String>()
        .trim()
        .to_string()
}

/// Parse a date shift string like "+1:0:0" (add 1 hour) or "-0:30:0" (subtract 30 min).
/// Returns (sign, hours, minutes, seconds).
pub fn parse_date_shift(shift: &str) -> Option<(i32, u32, u32, u32)> {
    let (sign, rest) = if let Some(stripped) = shift.strip_prefix('-') {
        (-1, stripped)
    } else if let Some(stripped) = shift.strip_prefix('+') {
        (1, stripped)
    } else {
        (1, shift)
    };

    let parts: Vec<&str> = rest.split(':').collect();
    match parts.len() {
        1 => {
            let h: u32 = parts[0].parse().ok()?;
            Some((sign, h, 0, 0))
        }
        2 => {
            let h: u32 = parts[0].parse().ok()?;
            let m: u32 = parts[1].parse().ok()?;
            Some((sign, h, m, 0))
        }
        3 => {
            let h: u32 = parts[0].parse().ok()?;
            let m: u32 = parts[1].parse().ok()?;
            let s: u32 = parts[2].parse().ok()?;
            Some((sign, h, m, s))
        }
        _ => None,
    }
}

/// Shift a datetime string by the given amount.
/// Input format: "YYYY:MM:DD HH:MM:SS"
pub fn shift_datetime(datetime: &str, shift: &str) -> Option<String> {
    let (sign, hours, minutes, seconds) = parse_date_shift(shift)?;

    // Parse date/time
    if datetime.len() < 19 {
        return None;
    }
    let year: i32 = datetime[0..4].parse().ok()?;
    let month: u32 = datetime[5..7].parse().ok()?;
    let day: u32 = datetime[8..10].parse().ok()?;
    let hour: u32 = datetime[11..13].parse().ok()?;
    let min: u32 = datetime[14..16].parse().ok()?;
    let sec: u32 = datetime[17..19].parse().ok()?;

    // Convert to total seconds, shift, convert back
    let total_secs = (hour * 3600 + min * 60 + sec) as i64
        + sign as i64 * (hours * 3600 + minutes * 60 + seconds) as i64;

    let days_shift = if total_secs < 0 {
        -1 - (-total_secs - 1) / 86400
    } else {
        total_secs / 86400
    };

    let time_secs = ((total_secs % 86400) + 86400) % 86400;
    let new_hour = (time_secs / 3600) as u32;
    let new_min = ((time_secs % 3600) / 60) as u32;
    let new_sec = (time_secs % 60) as u32;

    // Simple day shifting (doesn't handle month/year rollover perfectly for large shifts)
    let mut new_day = day as i32 + days_shift as i32;
    let mut new_month = month;
    let mut new_year = year;

    let days_in_month = |m: u32, y: i32| -> i32 {
        match m {
            1 | 3 | 5 | 7 | 8 | 10 | 12 => 31,
            4 | 6 | 9 | 11 => 30,
            2 => {
                if (y % 4 == 0 && y % 100 != 0) || y % 400 == 0 {
                    29
                } else {
                    28
                }
            }
            _ => 30,
        }
    };

    while new_day > days_in_month(new_month, new_year) {
        new_day -= days_in_month(new_month, new_year);
        new_month += 1;
        if new_month > 12 {
            new_month = 1;
            new_year += 1;
        }
    }
    while new_day < 1 {
        new_month = if new_month == 1 { 12 } else { new_month - 1 };
        if new_month == 12 {
            new_year -= 1;
        }
        new_day += days_in_month(new_month, new_year);
    }

    Some(format!(
        "{:04}:{:02}:{:02} {:02}:{:02}:{:02}",
        new_year, new_month, new_day, new_hour, new_min, new_sec
    ))
}

/// Group assignment of the file-level pseudo-tags, ported from
/// `%Image::ExifTool::System` and the `%allGroupsExifTool` entries of
/// `%Image::ExifTool::Extra` in `ExifTool.pm`.
///
/// ExifTool resolves these tags through a single table, so their groups do not
/// depend on which parser produced them: `Warning` is reported in the `ExifTool`
/// group whether it was raised by the PNG, QuickTime or MRC reader. Only family 1
/// splits the System tags out of `File` — their family 0 stays `File`, matching
/// `exiftool -G0`. The ExifTool pseudo-tags sit in `ExifTool` for all three
/// families.
///
/// Entries are `(name, family0, family1, family2)`.
const FILE_LEVEL_GROUPS: &[(&str, &str, &str, &str)] = &[
    // ExifTool computes CurrentIPTCDigest with `FoundTag`, so it lands in the
    // Extra table (`GROUPS => { 0 => 'File', 1 => 'File', 2 => 'Image' }`,
    // ExifTool.pm line 1285/1771), never a format reader's own group.
    ("CurrentIPTCDigest", "File", "File", "Image"),
    ("Directory", "File", "System", "Other"),
    ("Error", "ExifTool", "ExifTool", "ExifTool"),
    ("ExifToolVersion", "ExifTool", "ExifTool", "ExifTool"),
    ("FileAccessDate", "File", "System", "Time"),
    ("FileCreateDate", "File", "System", "Time"),
    ("FileInodeChangeDate", "File", "System", "Time"),
    ("FileModifyDate", "File", "System", "Time"),
    ("FileName", "File", "System", "Other"),
    ("FilePermissions", "File", "System", "Other"),
    ("FileSize", "File", "System", "Other"),
    ("Warning", "ExifTool", "ExifTool", "ExifTool"),
];

/// The `(family0, family1, family2)` groups [`FILE_LEVEL_GROUPS`] assigns to
/// `name`, or `None` if `name` is not a file-level pseudo-tag.
fn file_level_group(name: &str) -> Option<(&'static str, &'static str, &'static str)> {
    FILE_LEVEL_GROUPS
        .iter()
        .find(|(n, ..)| *n == name)
        .map(|&(_, f0, f1, f2)| (f0, f1, f2))
}

// Only used by the `#[cfg(unix)]` File:System FilePermissions pseudo-tag above;
// not compiled on Windows (which lacks Unix mode bits).
//
// Port of ExifTool's FilePermissions PrintConv: a leading file-type character
// (`-` for a regular file, `d`, `l`, …) followed by nine r/w/x flags for
// owner/group/other, e.g. mode 0o100664 → "-rw-rw-r--".
#[cfg(unix)]
fn format_file_permissions(mode: u32) -> String {
    let type_char = match mode & 0o170000 {
        0o010000 => 'p', // FIFO
        0o020000 => 'c', // character special
        0o040000 => 'd', // directory
        0o060000 => 'b', // block special
        0o120000 => 'l', // symlink
        0o140000 => 's', // socket
        _ => '-',
    };
    let mut s = String::with_capacity(10);
    s.push(type_char);
    let mut mask = 0o400u32;
    while mask > 0 {
        for ch in ['r', 'w', 'x'] {
            s.push(if mode & mask != 0 { ch } else { '-' });
            mask >>= 1;
        }
    }
    s
}

/// File contents exposed as a byte slice, backed either by a memory map or — when
/// mapping is unavailable (empty file, unsupported FS, mapping error) — an owned
/// buffer. Both `Deref` to `[u8]` so callers are agnostic to the backing store.
enum FileData {
    Mapped(memmap2::Mmap),
    Owned(Vec<u8>),
}

impl std::ops::Deref for FileData {
    type Target = [u8];
    fn deref(&self) -> &[u8] {
        match self {
            FileData::Mapped(m) => m,
            FileData::Owned(v) => v,
        }
    }
}

/// Map a file read-only for parsing. Zero-length files (which cannot be mapped)
/// and any mapping failure fall back to a plain `fs::read`.
fn map_file_for_read(path: &Path) -> Result<FileData> {
    let file = fs::File::open(path).map_err(Error::Io)?;
    let len = file.metadata().map_err(Error::Io)?.len();
    if len == 0 {
        return Ok(FileData::Owned(Vec::new()));
    }
    // SAFETY: the mapping is only ever read, never written, and the `Mmap` is
    // dropped before `extract_info` returns. If another process truncates the
    // file mid-parse the kernel may raise SIGBUS — the same exposure ExifTool's
    // own random-access reads have; acceptable for a read-only metadata tool.
    match unsafe { memmap2::Mmap::map(&file) } {
        Ok(m) => Ok(FileData::Mapped(m)),
        Err(_) => Ok(FileData::Owned(fs::read(path).map_err(Error::Io)?)),
    }
}

/// Port of ExifTool ConvertFileSize (decimal units): %.1f below 10× a unit, %.0f above.
fn format_file_size(bytes: u64) -> String {
    let v = bytes as f64;
    if bytes < 2000 {
        format!("{} bytes", bytes)
    } else if bytes < 10_000 {
        format!("{:.1} kB", v / 1000.0)
    } else if bytes < 2_000_000 {
        format!("{:.0} kB", v / 1000.0)
    } else if bytes < 10_000_000 {
        format!("{:.1} MB", v / 1_000_000.0)
    } else if bytes < 2_000_000_000 {
        format!("{:.0} MB", v / 1_000_000.0)
    } else if bytes < 10_000_000_000 {
        format!("{:.1} GB", v / 1_000_000_000.0)
    } else {
        format!("{:.0} GB", v / 1_000_000_000.0)
    }
}

/// Check if a tag name is typically XMP.
fn is_xmp_tag(tag: &str) -> bool {
    matches!(
        tag.to_lowercase().as_str(),
        "title"
            | "description"
            | "subject"
            | "creator"
            | "rights"
            | "keywords"
            | "rating"
            | "label"
            | "hierarchicalsubject"
    )
}

/// Encode an EXIF tag value to binary.
/// Returns (tag_id, format, encoded_data) or None if tag is unknown.
fn encode_exif_tag(
    tag_name: &str,
    value: &str,
    _group: &str,
    bo: ByteOrderMark,
) -> Option<(u16, exif_writer::ExifFormat, Vec<u8>)> {
    let tag_lower = tag_name.to_lowercase();

    // Map common tag names to EXIF tag IDs and formats
    let (tag_id, format): (u16, exif_writer::ExifFormat) = match tag_lower.as_str() {
        // IFD0 string tags
        "imagedescription" => (0x010E, exif_writer::ExifFormat::Ascii),
        "make" => (0x010F, exif_writer::ExifFormat::Ascii),
        "model" => (0x0110, exif_writer::ExifFormat::Ascii),
        "software" => (0x0131, exif_writer::ExifFormat::Ascii),
        "modifydate" | "datetime" => (0x0132, exif_writer::ExifFormat::Ascii),
        "artist" => (0x013B, exif_writer::ExifFormat::Ascii),
        "copyright" => (0x8298, exif_writer::ExifFormat::Ascii),
        // IFD0 numeric tags
        "orientation" => (0x0112, exif_writer::ExifFormat::Short),
        "xresolution" => (0x011A, exif_writer::ExifFormat::Rational),
        "yresolution" => (0x011B, exif_writer::ExifFormat::Rational),
        "resolutionunit" => (0x0128, exif_writer::ExifFormat::Short),
        // ExifIFD tags
        "datetimeoriginal" => (0x9003, exif_writer::ExifFormat::Ascii),
        "createdate" | "datetimedigitized" => (0x9004, exif_writer::ExifFormat::Ascii),
        "usercomment" => (0x9286, exif_writer::ExifFormat::Undefined),
        "imageuniqueid" => (0xA420, exif_writer::ExifFormat::Ascii),
        "ownername" | "cameraownername" => (0xA430, exif_writer::ExifFormat::Ascii),
        "serialnumber" | "bodyserialnumber" => (0xA431, exif_writer::ExifFormat::Ascii),
        "lensmake" => (0xA433, exif_writer::ExifFormat::Ascii),
        "lensmodel" => (0xA434, exif_writer::ExifFormat::Ascii),
        "lensserialnumber" => (0xA435, exif_writer::ExifFormat::Ascii),
        _ => return None,
    };

    let encoded = match format {
        exif_writer::ExifFormat::Ascii => exif_writer::encode_ascii(value),
        exif_writer::ExifFormat::Short => {
            let v: u16 = value.parse().ok()?;
            exif_writer::encode_u16(v, bo)
        }
        exif_writer::ExifFormat::Long => {
            let v: u32 = value.parse().ok()?;
            exif_writer::encode_u32(v, bo)
        }
        exif_writer::ExifFormat::Rational => {
            // Parse "N/D" or just "N"
            if let Some(slash) = value.find('/') {
                let num: u32 = value[..slash].trim().parse().ok()?;
                let den: u32 = value[slash + 1..].trim().parse().ok()?;
                exif_writer::encode_urational(num, den, bo)
            } else if let Ok(v) = value.parse::<f64>() {
                // Convert float to rational
                let den = 10000u32;
                let num = (v * den as f64).round() as u32;
                exif_writer::encode_urational(num, den, bo)
            } else {
                return None;
            }
        }
        exif_writer::ExifFormat::Undefined => {
            // UserComment: 8 bytes charset + data
            let mut data = vec![0x41, 0x53, 0x43, 0x49, 0x49, 0x00, 0x00, 0x00]; // "ASCII\0\0\0"
            data.extend_from_slice(value.as_bytes());
            data
        }
        _ => return None,
    };

    Some((tag_id, format, encoded))
}

/// Compute text file tags (from Perl Text.pm).
fn compute_text_tags(data: &[u8], is_csv: bool) -> Vec<Tag> {
    let mut tags = Vec::new();
    let mk = |name: &str, val: String| Tag {
        id: crate::tag::TagId::Text(name.into()),
        name: name.into(),
        description: name.into(),
        group: crate::tag::TagGroup {
            family0: "File".into(),
            family1: "File".into(),
            family2: "Other".into(),
            family3: "Main".into(),
        },
        raw_value: Value::String(val.clone()),
        print_value: val,
        priority: 0,
    };

    // Detect encoding and BOM
    let is_ascii = data.iter().all(|&b| b < 128);
    let has_utf8_bom = data.starts_with(&[0xEF, 0xBB, 0xBF]);
    let has_utf16le_bom =
        data.starts_with(&[0xFF, 0xFE]) && !data.starts_with(&[0xFF, 0xFE, 0x00, 0x00]);
    let has_utf16be_bom = data.starts_with(&[0xFE, 0xFF]);
    let has_utf32le_bom = data.starts_with(&[0xFF, 0xFE, 0x00, 0x00]);
    let has_utf32be_bom = data.starts_with(&[0x00, 0x00, 0xFE, 0xFF]);

    // Detect if file has weird non-text control characters (like multi-byte unicode without BOM)
    let has_weird_ctrl = data.iter().any(|&b| {
        (b <= 0x06) || (0x0e..=0x1a).contains(&b) || (0x1c..=0x1f).contains(&b) || b == 0x7f
    });

    let (encoding, is_bom, is_utf16) = if has_utf32le_bom {
        ("utf-32le", true, false)
    } else if has_utf32be_bom {
        ("utf-32be", true, false)
    } else if has_utf16le_bom {
        ("utf-16le", true, true)
    } else if has_utf16be_bom {
        ("utf-16be", true, true)
    } else if has_weird_ctrl {
        // Not a text file (has binary-like control chars but no recognized multi-byte marker)
        return tags;
    } else if is_ascii {
        ("us-ascii", false, false)
    } else {
        // Check UTF-8
        let is_valid_utf8 = std::str::from_utf8(data).is_ok();
        if is_valid_utf8 {
            if has_utf8_bom {
                ("utf-8", true, false)
            } else {
                // Check if it has high bytes suggesting iso-8859-1 vs utf-8
                // Perl's IsUTF8: returns >0 if valid UTF-8 with multi-byte, 0 if ASCII, <0 if invalid
                // For simplicity: valid UTF-8 without BOM = utf-8
                ("utf-8", false, false)
            }
        } else if !data.iter().any(|&b| (0x80..=0x9f).contains(&b)) {
            ("iso-8859-1", false, false)
        } else {
            ("unknown-8bit", false, false)
        }
    };

    tags.push(mk("MIMEEncoding", encoding.into()));

    if is_bom {
        tags.push(mk("ByteOrderMark", "Yes".into()));
    }

    // Count newlines and detect type
    let has_cr = data.contains(&b'\r');
    let has_lf = data.contains(&b'\n');
    let newline_type = if has_cr && has_lf {
        "Windows CRLF"
    } else if has_lf {
        "Unix LF"
    } else if has_cr {
        "Macintosh CR"
    } else {
        "(none)"
    };
    tags.push(mk("Newlines", newline_type.into()));

    if is_csv {
        // CSV analysis: detect delimiter, quoting, column count, row count
        let text = crate::encoding::decode_utf8_or_latin1(data);
        let mut delim = "";
        let mut quot = "";
        let mut ncols = 1usize;
        let mut nrows = 0usize;

        for line in text.lines() {
            if nrows == 0 {
                // Detect delimiter from first line
                let comma_count = line.matches(',').count();
                let semi_count = line.matches(';').count();
                let tab_count = line.matches('\t').count();
                if comma_count > semi_count && comma_count > tab_count {
                    delim = ",";
                    ncols = comma_count + 1;
                } else if semi_count > tab_count {
                    delim = ";";
                    ncols = semi_count + 1;
                } else if tab_count > 0 {
                    delim = "\t";
                    ncols = tab_count + 1;
                } else {
                    delim = "";
                    ncols = 1;
                }
                // Detect quoting
                if line.contains('"') {
                    quot = "\"";
                } else if line.contains('\'') {
                    quot = "'";
                }
            }
            nrows += 1;
            if nrows >= 1000 {
                break;
            }
        }

        let delim_display = match delim {
            "," => "Comma",
            ";" => "Semicolon",
            "\t" => "Tab",
            _ => "(none)",
        };
        let quot_display = match quot {
            "\"" => "Double quotes",
            "'" => "Single quotes",
            _ => "(none)",
        };

        tags.push(mk("Delimiter", delim_display.into()));
        tags.push(mk("Quoting", quot_display.into()));
        tags.push(mk("ColumnCount", ncols.to_string()));
        if nrows > 0 {
            tags.push(mk("RowCount", nrows.to_string()));
        }
    } else if !is_utf16 {
        // Line count and word count for plain text files (not UTF-16/32)
        // ExifTool counts each ReadLine, so trailing content without a final newline
        // still counts as a line.
        let nl_count = data.iter().filter(|&&b| b == b'\n').count();
        let line_count = if !data.is_empty() && data.last() != Some(&b'\n') {
            nl_count + 1
        } else {
            nl_count
        };
        tags.push(mk("LineCount", line_count.to_string()));

        let text = crate::encoding::decode_utf8_or_latin1(data);
        let word_count = text.split_whitespace().count();
        tags.push(mk("WordCount", word_count.to_string()));
    }

    tags
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn new_has_default_options() {
        let et = ExifTool::new();
        assert!(!et.options().duplicates);
        assert!(et.options().print_conv);
        assert_eq!(et.options().fast_scan, 0);
        assert!(et.options().requested_tags.is_empty());
        assert_eq!(et.options().extract_embedded, 0);
        assert_eq!(et.options().show_unknown, 0);
        assert!(!et.options().process_compressed);
        assert!(!et.options().use_mwg);
    }

    #[test]
    fn tag_matches_request_group_qualified() {
        let tag = Tag {
            id: crate::tag::TagId::Text("By-line".into()),
            name: "By-line".into(),
            description: "By-line".into(),
            group: crate::tag::TagGroup {
                family0: "IPTC".into(),
                family1: "IPTC".into(),
                family2: "Author".into(),
                family3: "Main".into(),
            },
            raw_value: Value::String("Martín".into()),
            print_value: "Martín".into(),
            priority: 1,
        };
        // Bare name (case-insensitive).
        assert!(ExifTool::tag_matches_request(&tag, "By-line"));
        assert!(ExifTool::tag_matches_request(&tag, "by-line"));
        // Group-qualified against families 0 and 2.
        assert!(ExifTool::tag_matches_request(&tag, "IPTC:By-line"));
        assert!(ExifTool::tag_matches_request(&tag, "Author:By-line"));
        // Wildcards.
        assert!(ExifTool::tag_matches_request(&tag, "IPTC:*"));
        assert!(ExifTool::tag_matches_request(&tag, "*"));
        // Wrong group / wrong name → no match.
        assert!(!ExifTool::tag_matches_request(&tag, "EXIF:By-line"));
        assert!(!ExifTool::tag_matches_request(&tag, "IPTC:Make"));
        assert!(!ExifTool::tag_matches_request(&tag, "Headline"));
    }

    #[test]
    fn with_options_preserves_custom() {
        let opts = Options {
            duplicates: true,
            print_conv: false,
            fast_scan: 2,
            requested_tags: vec!["Artist".to_string()],
            extract_embedded: 1,
            show_unknown: 1,
            process_compressed: true,
            use_mwg: true,
            geolocation: true,
        };
        let et = ExifTool::with_options(opts.clone());
        assert!(et.options().duplicates);
        assert!(!et.options().print_conv);
        assert_eq!(et.options().fast_scan, 2);
        assert_eq!(et.options().requested_tags, vec!["Artist".to_string()]);
        assert_eq!(et.options().extract_embedded, 1);
        assert_eq!(et.options().show_unknown, 1);
        assert!(et.options().process_compressed);
        assert!(et.options().use_mwg);
    }

    #[test]
    fn set_new_value_simple_tag() {
        let mut et = ExifTool::new();
        et.set_new_value("Artist", Some("John"));
        assert_eq!(et.new_values.len(), 1);
        assert_eq!(et.new_values[0].tag, "Artist");
        assert_eq!(et.new_values[0].group, None);
        assert_eq!(et.new_values[0].value, Some("John".to_string()));
    }

    #[test]
    fn set_new_value_with_group_prefix() {
        let mut et = ExifTool::new();
        et.set_new_value("XMP:Title", Some("Test"));
        assert_eq!(et.new_values.len(), 1);
        assert_eq!(et.new_values[0].tag, "Title");
        assert_eq!(et.new_values[0].group, Some("XMP".to_string()));
        assert_eq!(et.new_values[0].value, Some("Test".to_string()));
    }

    #[test]
    fn set_new_value_delete() {
        let mut et = ExifTool::new();
        et.set_new_value("Comment", None);
        assert_eq!(et.new_values.len(), 1);
        assert_eq!(et.new_values[0].tag, "Comment");
        assert_eq!(et.new_values[0].value, None);
    }

    #[test]
    fn clear_new_values_empties_queue() {
        let mut et = ExifTool::new();
        et.set_new_value("Artist", Some("A"));
        et.set_new_value("Copyright", Some("B"));
        assert_eq!(et.new_values.len(), 2);
        et.clear_new_values();
        assert!(et.new_values.is_empty());
    }

    #[test]
    fn set_new_value_multiple() {
        let mut et = ExifTool::new();
        et.set_new_value("Artist", Some("John"));
        et.set_new_value("IPTC:Keywords", Some("test"));
        et.set_new_value("XMP:Subject", None);
        assert_eq!(et.new_values.len(), 3);
        assert_eq!(et.new_values[1].group, Some("IPTC".to_string()));
        assert_eq!(et.new_values[1].tag, "Keywords");
        assert_eq!(et.new_values[2].value, None);
    }

    #[test]
    fn options_mut_modifies() {
        let mut et = ExifTool::new();
        et.options_mut().duplicates = true;
        et.options_mut().fast_scan = 3;
        assert!(et.options().duplicates);
        assert_eq!(et.options().fast_scan, 3);
    }

    #[test]
    fn default_options() {
        let opts = Options::default();
        assert!(!opts.duplicates);
        assert!(opts.print_conv);
        assert_eq!(opts.fast_scan, 0);
    }
}