diff --git a/README.md b/README.md
index 03b5971..1bed58e 100644
--- a/README.md
+++ b/README.md
@@ -404,6 +404,23 @@ comments,world,epub,publish,server}` implement them; `src/auth.rs` owns the rati
 token formula used by both the EPUB writer and the server; `src/types.rs` is the
 contract between stages.
 
+### Auditing images against real articles
+
+Image handling fails in ways no synthetic fixture predicts, because every
+publisher invents its own lazy-loading scheme. `examples/image_audit.rs` replays
+the extraction and image pipeline over the articles of issues already published
+and counts what actually reaches the page:
+
+```sh
+cargo run --release --example image_audit -- \
+    --cache /tmp/pagecache ~/bookorbit/books/daily-epub/*.epub
+```
+
+It prints per-article `refs / embedded / shown / placeholders`, a tally of loss
+reasons, and totals. Pages are cached on first run, so a change can be measured
+against byte-identical input; `--dump <title substring>` lists the URLs one
+article resolved to. It needs the network and is not part of `cargo test`.
+
 ---
 
 ## Known limitations
@@ -425,6 +442,10 @@ From spec §7, plus what implementation turned up:
   falls back to matching the feed title.
 - **Images are downloaded once per edition** (the two editions need different
   resolutions and colour profiles), so an image-heavy issue makes two passes.
+- **An image that cannot be embedded is dropped, not announced,** unless its alt
+  text is a real description — decorative rules, spacers and dead links would
+  otherwise litter the page with `[image: …]` lines. Verify image handling with
+  `examples/image_audit.rs` after touching extraction.
 - **`dc:date` rides inside a `dcterms:date` metadata fragment** because
   `epub-builder` neither exposes `dc:date` nor accepts a non-`chrono` date. The
   OPF output is correct; the mechanism is a workaround.
diff --git a/examples/image_audit.rs b/examples/image_audit.rs
new file mode 100644
index 0000000..aa73217
--- /dev/null
+++ b/examples/image_audit.rs
@@ -0,0 +1,446 @@
+//! Real-world image audit: replay the extraction + image pipeline over the
+//! articles of already-published issues and report what reaches the page.
+//!
+//! Unit tests can only prove the code does what we think on markup we wrote.
+//! This runs the same code over the pages the issues were actually built from,
+//! so a regression in lazy-image handling, URL matching or re-encoding shows up
+//! as a number rather than a hunch.
+//!
+//! ```text
+//! cargo run --release --example image_audit -- \
+//!     --cache /tmp/pagecache ~/bookorbit/books/daily-epub/*.epub
+//! ```
+//!
+//! Pages are cached on disk after the first run, so before/after comparisons
+//! see byte-identical input. Articles are re-fetched and re-extracted through
+//! readability even when the original issue took its body from Miniflux, so the
+//! counts describe the fetch path, not that specific issue's history.
+
+use std::collections::BTreeMap;
+use std::io::Read;
+use std::path::{Path, PathBuf};
+
+use daily_epub::epub::{build, images};
+use daily_epub::extract::{self, Extractor};
+use daily_epub::types::{Article, EntryId, ExtractMethod, ImageAsset, Pick, SourceKind, SourceRef};
+
+/// One article we are auditing, pulled back out of a published EPUB.
+#[derive(Debug, Clone)]
+struct Target {
+    issue: String,
+    entry_id: EntryId,
+    title: String,
+    url: String,
+}
+
+/// What the pipeline did with one article's images.
+#[derive(Debug, Default, Clone)]
+struct Outcome {
+    /// `<img>` elements in the extracted body.
+    refs: usize,
+    /// Images that downloaded and re-encoded successfully.
+    embedded: usize,
+    /// `<img src="images/…">` in the rendered chapter.
+    rendered: usize,
+    /// `[image: …]` paragraphs in the rendered chapter.
+    placeholders: usize,
+    /// Reasons individual images did not make it, most specific first.
+    losses: Vec<String>,
+}
+
+#[tokio::main]
+async fn main() {
+    tracing_subscriber::fmt()
+        .with_writer(std::io::stderr)
+        .with_env_filter(
+            tracing_subscriber::EnvFilter::try_from_default_env()
+                .unwrap_or_else(|_| "error".into()),
+        )
+        .init();
+
+    let mut cache = PathBuf::from("/tmp/daily-epub-audit-cache");
+    let mut dump: Option<String> = None;
+    let mut epubs: Vec<PathBuf> = Vec::new();
+    let mut args = std::env::args().skip(1);
+    while let Some(arg) = args.next() {
+        match arg.as_str() {
+            "--cache" => cache = PathBuf::from(args.next().expect("--cache needs a path")),
+            "--dump" => dump = Some(args.next().expect("--dump needs a title substring")),
+            other => epubs.push(PathBuf::from(other)),
+        }
+    }
+    if epubs.is_empty() {
+        eprintln!("usage: image_audit [--cache DIR] <issue.epub>...");
+        std::process::exit(2);
+    }
+    std::fs::create_dir_all(&cache).expect("cache dir");
+
+    // One entry per article: passing `*.epub` picks up both editions of every
+    // issue, and the same article twice would collide on its asset ids.
+    let mut seen: std::collections::HashSet<EntryId> = std::collections::HashSet::new();
+    let targets: Vec<Target> = epubs
+        .iter()
+        .flat_map(|p| targets_from_epub(p))
+        .filter(|t| seen.insert(t.entry_id))
+        .collect();
+    eprintln!(
+        "auditing {} articles from {} files",
+        targets.len(),
+        epubs.len()
+    );
+
+    let http = daily_epub::http::build_client(std::time::Duration::from_secs(20)).unwrap();
+    let extractor = Extractor::new(http.clone(), vec![]);
+
+    // Extract every article first, then run one issue-wide image pass, exactly
+    // as `epub::build_edition` does.
+    let mut picks: Vec<Pick> = Vec::new();
+    let mut skipped: Vec<(Target, String)> = Vec::new();
+    for target in &targets {
+        match body_for(&cache, &http, &extractor, target).await {
+            Ok(html) => picks.push(pick_for(target, html)),
+            Err(e) => skipped.push((target.clone(), e)),
+        }
+    }
+
+    let assets =
+        images::collect_for_issue(&http, &picks, daily_epub::types::Edition::Standard).await;
+    eprintln!("embedded {} images", assets.len());
+
+    let by_entry: BTreeMap<EntryId, Vec<&ImageAsset>> = assets.iter().fold(
+        BTreeMap::new(),
+        |mut acc: BTreeMap<EntryId, Vec<&ImageAsset>>, a| {
+            if let Some(id) = entry_of(&a.id) {
+                acc.entry(id).or_default().push(a);
+            }
+            acc
+        },
+    );
+
+    let mut totals = Outcome::default();
+    let mut rows: Vec<(Target, Outcome)> = Vec::new();
+    for (target, pick) in targets_of(&targets, &picks) {
+        let refs = images::extract_img_refs(&pick.article.content_html);
+        let embedded = by_entry.get(&target.entry_id).map_or(0, |v| v.len());
+        let body = build::prepare_body(&pick.article.content_html, &assets);
+        let rendered = body.matches("<img src=\"images/").count();
+        let placeholders = body.matches("image-placeholder").count();
+
+        let embedded_urls: Vec<&str> = by_entry
+            .get(&target.entry_id)
+            .map(|v| v.iter().map(|a| a.source_url.as_str()).collect())
+            .unwrap_or_default();
+        let losses = refs
+            .iter()
+            .filter(|r| !embedded_urls.contains(&r.src.as_str()))
+            .map(|r| classify(&r.src))
+            .collect::<Vec<_>>();
+
+        if dump.as_ref().is_some_and(|d| target.title.contains(d)) {
+            println!("\n== {} [{}]", target.title, target.url);
+            for r in &refs {
+                let state = if embedded_urls.contains(&r.src.as_str()) {
+                    "ok  "
+                } else {
+                    "MISS"
+                };
+                println!("  {state} {}", r.src);
+            }
+            println!("  --- rendered body images:");
+            for cap in body.split("<img src=\"").skip(1) {
+                println!("  {}", cap.split('"').next().unwrap_or_default());
+            }
+        }
+
+        let outcome = Outcome {
+            refs: refs.len(),
+            embedded,
+            rendered,
+            placeholders,
+            losses,
+        };
+        totals.refs += outcome.refs;
+        totals.embedded += outcome.embedded;
+        totals.rendered += outcome.rendered;
+        totals.placeholders += outcome.placeholders;
+        totals.losses.extend(outcome.losses.iter().cloned());
+        rows.push((target.clone(), outcome));
+    }
+
+    println!(
+        "\n{:<12} {:>5} {:>5} {:>5} {:>5}  article",
+        "issue", "refs", "emb", "shown", "ph"
+    );
+    println!("{}", "-".repeat(100));
+    for (t, o) in &rows {
+        let flag = if o.placeholders > 0 || o.rendered < o.refs {
+            "!"
+        } else {
+            " "
+        };
+        println!(
+            "{:<12} {:>5} {:>5} {:>5} {:>5} {} {} [{}]",
+            t.issue,
+            o.refs,
+            o.embedded,
+            o.rendered,
+            o.placeholders,
+            flag,
+            truncate(&t.title, 44),
+            host(&t.url),
+        );
+        for loss in &o.losses {
+            println!("{:>36}   lost: {loss}", "");
+        }
+    }
+
+    println!("\n== totals");
+    println!("  articles           {}", rows.len());
+    println!("  <img> in bodies    {}", totals.refs);
+    println!("  embedded           {}", totals.embedded);
+    println!("  shown in chapters  {}", totals.rendered);
+    println!("  placeholders       {}", totals.placeholders);
+    println!(
+        "  orphaned assets    {}",
+        totals.embedded.saturating_sub(totals.rendered)
+    );
+    let mut kinds: BTreeMap<String, usize> = BTreeMap::new();
+    for loss in &totals.losses {
+        *kinds.entry(loss.clone()).or_default() += 1;
+    }
+    println!("\n== loss reasons");
+    for (kind, n) in &kinds {
+        println!("  {n:>4}  {kind}");
+    }
+    if !skipped.is_empty() {
+        println!("\n== unfetchable ({})", skipped.len());
+        for (t, e) in &skipped {
+            println!("  {} — {e}", truncate(&t.title, 50));
+        }
+    }
+}
+
+/// Pair each target with the pick built from it (targets that failed to fetch
+/// have no pick and are skipped).
+fn targets_of<'a>(targets: &'a [Target], picks: &'a [Pick]) -> Vec<(&'a Target, &'a Pick)> {
+    picks
+        .iter()
+        .filter_map(|p| {
+            targets
+                .iter()
+                .find(|t| t.entry_id == p.article.best_entry_id)
+                .map(|t| (t, p))
+        })
+        .collect()
+}
+
+fn entry_of(asset_id: &str) -> Option<EntryId> {
+    asset_id
+        .strip_prefix("img-")?
+        .split('-')
+        .next()?
+        .parse()
+        .ok()
+}
+
+/// Why one `<img>` never became an embedded asset — a guess from its URL, good
+/// enough to group the failures.
+fn classify(src: &str) -> String {
+    let lower = src.to_ascii_lowercase();
+    if src.contains('{') || src.contains('}') {
+        format!("unresolved URL template — {}", truncate(src, 90))
+    } else if src.contains(' ') || src.contains("%20") {
+        format!("srcset blob in src — {}", truncate(src, 90))
+    } else if lower.contains(".svg") {
+        format!("svg — {}", truncate(src, 90))
+    } else if src.starts_with("data:") {
+        "data: URI (lazy placeholder)".to_string()
+    } else {
+        format!("download/decode failed — {}", truncate(src, 90))
+    }
+}
+
+/// Fetch + extract one article, using the on-disk page cache.
+async fn body_for(
+    cache: &Path,
+    _http: &reqwest::Client,
+    extractor: &Extractor,
+    target: &Target,
+) -> Result<String, String> {
+    let key = cache.join(format!("{:x}.html", seahash(&target.url)));
+    let url_key = key.with_extension("url");
+    if !key.exists() {
+        // `fetch_readable` does the fetching we want but returns readability's
+        // output; cache the raw page instead so extraction changes are visible.
+        let (raw, final_url) = raw_fetch(&target.url).await?;
+        std::fs::write(&key, raw).map_err(|e| e.to_string())?;
+        std::fs::write(&url_key, final_url).map_err(|e| e.to_string())?;
+    }
+    let bytes = std::fs::read(&key).map_err(|e| e.to_string())?;
+    if bytes.is_empty() {
+        return Err("empty page".into());
+    }
+    // Relative URLs belong to the page we landed on, not the one we asked for.
+    let base = std::fs::read_to_string(&url_key).unwrap_or_else(|_| target.url.clone());
+    let html = String::from_utf8_lossy(&bytes).into_owned();
+    let _ = extractor;
+    let readable = extract::readability(&html, &base).map_err(|e| e.to_string())?;
+    Ok(extract::sanitize_with_base(
+        &extract::normalize_img_tags(&readable),
+        &base,
+    ))
+}
+
+/// A plain page fetch with the extractor's desktop UA and gzip handling,
+/// returning the body and the URL the fetch landed on.
+async fn raw_fetch(url: &str) -> Result<(Vec<u8>, String), String> {
+    let client = reqwest::Client::builder()
+        .user_agent(extract::DESKTOP_UA)
+        .gzip(true)
+        .timeout(std::time::Duration::from_secs(30))
+        .build()
+        .map_err(|e| e.to_string())?;
+    let resp = client
+        .get(url)
+        .header(
+            reqwest::header::ACCEPT,
+            "text/html,application/xhtml+xml,application/xml;q=0.9,*/*;q=0.8",
+        )
+        .send()
+        .await
+        .map_err(|e| e.to_string())?;
+    if !resp.status().is_success() {
+        return Err(format!("HTTP {}", resp.status()));
+    }
+    let final_url = resp.url().to_string();
+    resp.bytes()
+        .await
+        .map(|b| (b.to_vec(), final_url))
+        .map_err(|e| e.to_string())
+}
+
+fn pick_for(target: &Target, content_html: String) -> Pick {
+    let word_count = extract::word_count(&content_html);
+    let image_urls = extract::collect_image_urls(&content_html, &target.url);
+    Pick {
+        article: Article {
+            id: target.entry_id,
+            canonical_url: target.url.clone(),
+            title: target.title.clone(),
+            best_entry_id: target.entry_id,
+            content_html,
+            word_count,
+            excerpt_only: false,
+            image_count: image_urls.len() as i64,
+            image_urls,
+            sources: vec![SourceRef {
+                entry_id: target.entry_id,
+                feed_id: 1,
+                feed_title: "Feed".into(),
+                category: None,
+                kind: SourceKind::Feed,
+            }],
+            first_seen: "2026-08-15T05:30:00Z".parse().unwrap(),
+            url: target.url.clone(),
+            author: None,
+            feed_id: 1,
+            feed_title: "Feed".into(),
+            category: None,
+            published_at: None,
+            comments_url: None,
+            social: vec![],
+            extract_method: ExtractMethod::Readability,
+        },
+        section: "Audit".into(),
+        position: 0,
+        is_lead: false,
+        summary: None,
+        llm: None,
+        discussion: None,
+    }
+}
+
+/// Pull `(entry id, title, url)` out of every article chapter in an EPUB.
+fn targets_from_epub(path: &Path) -> Vec<Target> {
+    let issue = path
+        .file_stem()
+        .map(|s| s.to_string_lossy().replace("The Daily EPUB - ", ""))
+        .unwrap_or_default();
+    let file = std::fs::File::open(path).expect("open epub");
+    let mut zip = zip::ZipArchive::new(file).expect("read epub");
+    let names: Vec<String> = (0..zip.len())
+        .filter_map(|i| zip.by_index(i).ok().map(|f| f.name().to_string()))
+        .filter(|n| n.contains("art-") && n.ends_with(".xhtml"))
+        .collect();
+
+    let mut out = Vec::new();
+    for name in names {
+        let mut xhtml = String::new();
+        if zip
+            .by_name(&name)
+            .and_then(|mut f| f.read_to_string(&mut xhtml).map_err(Into::into))
+            .is_err()
+        {
+            continue;
+        }
+        let Some(entry_id) = name
+            .rsplit('/')
+            .next()
+            .and_then(|f| f.strip_prefix("art-"))
+            .and_then(|f| f.strip_suffix(".xhtml"))
+            .and_then(|f| f.parse::<EntryId>().ok())
+        else {
+            continue;
+        };
+        let Some(url) = between(&xhtml, r#"class="read-online"><a href=""#, '"') else {
+            continue;
+        };
+        let title = between(&xhtml, "<title>", '<').unwrap_or_else(|| "?".into());
+        out.push(Target {
+            issue: issue.clone(),
+            entry_id,
+            title: unescape(&title),
+            url: unescape(&url),
+        });
+    }
+    out
+}
+
+fn between(haystack: &str, prefix: &str, end: char) -> Option<String> {
+    let start = haystack.find(prefix)? + prefix.len();
+    let rest = &haystack[start..];
+    let stop = rest.find(end)?;
+    Some(rest[..stop].to_string())
+}
+
+fn unescape(s: &str) -> String {
+    s.replace("&amp;", "&")
+        .replace("&#39;", "'")
+        .replace("&quot;", "\"")
+        .replace("&lt;", "<")
+        .replace("&gt;", ">")
+}
+
+fn host(url: &str) -> String {
+    url::Url::parse(url)
+        .ok()
+        .and_then(|u| u.host_str().map(str::to_string))
+        .unwrap_or_default()
+}
+
+fn truncate(s: &str, n: usize) -> String {
+    if s.chars().count() <= n {
+        return s.to_string();
+    }
+    s.chars().take(n.saturating_sub(1)).collect::<String>() + "…"
+}
+
+/// Tiny stable hash for cache filenames.
+fn seahash(s: &str) -> u64 {
+    let mut h: u64 = 0xcbf2_9ce4_8422_2325;
+    for b in s.as_bytes() {
+        h ^= u64::from(*b);
+        h = h.wrapping_mul(0x0000_0100_0000_01b3);
+    }
+    h
+}
diff --git a/src/epub/build.rs b/src/epub/build.rs
index 9793808..2bb1e82 100644
--- a/src/epub/build.rs
+++ b/src/epub/build.rs
@@ -451,7 +451,7 @@ fn encode_cover(pixmap: tiny_skia::Pixmap, edition: Edition) -> Result<Vec<u8>,
 // ---------------------------------------------------------------------------
 
 /// Prepare article markup for XHTML: rewrite images, sanitize, self-close voids.
-fn prepare_body(html: &str, images_: &[ImageAsset]) -> String {
+pub fn prepare_body(html: &str, images_: &[ImageAsset]) -> String {
     // Sanitize first: image rewriting emits our own trusted markup (including the
     // `image-placeholder` class, which ammonia would otherwise strip).
     let cleaned = ammonia::clean(html);
@@ -1013,7 +1013,7 @@ pub mod fixtures {
             title: title.to_string(),
             best_entry_id: entry_id,
             content_html: format!(
-                "<p>Body of <em>{title}</em> with an image.</p><img src=\"https://img.example/{entry_id}.png\" alt=\"A chart\"><p>More words &amp; things.</p>"
+                "<p>Body of <em>{title}</em> with an image.</p><img src=\"https://img.example/{entry_id}.png\" alt=\"A chart of the daily figures\"><p>More words &amp; things.</p>"
             ),
             word_count: 1200,
             excerpt_only: false,
@@ -1307,7 +1307,11 @@ mod tests {
         assert!(chapter.xhtml.contains("Read online"));
         assert!(chapter.xhtml.contains("href=\"disc-1001.xhtml\""));
         // The un-downloaded image degrades to a placeholder.
-        assert!(chapter.xhtml.contains("[image: A chart]"));
+        assert!(
+            chapter
+                .xhtml
+                .contains("[image: A chart of the daily figures]")
+        );
         assert_xml_ok(&chapter.xhtml);
     }
 
diff --git a/src/epub/images.rs b/src/epub/images.rs
index a002155..e6a4c11 100644
--- a/src/epub/images.rs
+++ b/src/epub/images.rs
@@ -22,8 +22,8 @@ pub const CONCURRENCY: usize = 8;
 pub const ISSUE_ASSET_BUDGET_BYTES: usize = 25 * 1024 * 1024;
 /// Images smaller than this in either dimension are decorative — skipped (§3.10).
 pub const MIN_DIMENSION_PX: u32 = 24;
-/// Images referenced per article are already capped at 12 by extraction (§3.3).
-pub const MAX_IMAGES_PER_ARTICLE: usize = 12;
+/// Width/height an SVG is rasterized to when it declares no intrinsic size.
+const SVG_FALLBACK_SIZE: u32 = 1000;
 
 /// HTML void elements: XHTML requires them self-closed (§3.10 "valid XHTML").
 pub const VOID_ELEMENTS: &[&str] = &[
@@ -136,6 +136,8 @@ pub fn extract_img_refs(html: &str) -> Vec<ImgRef> {
 pub async fn download(http: &reqwest::Client, url: &str) -> Option<Vec<u8>> {
     let resp = http
         .get(url)
+        // Some CDNs answer `Accept: */*` with an HTML interstitial (§3.10).
+        .header(reqwest::header::ACCEPT, "image/*,*/*;q=0.8")
         .timeout(Duration::from_secs(DOWNLOAD_TIMEOUT_SECS))
         .send()
         .await
@@ -175,8 +177,14 @@ pub async fn download(http: &reqwest::Client, url: &str) -> Option<Vec<u8>> {
 /// Decode, resize/grayscale, flatten transparency to white and re-encode (§3.10).
 ///
 /// Line art with transparency is kept as PNG after flattening; everything else
-/// becomes JPEG. Returns `None` for undecodable sources (SVG/WebP without support).
+/// becomes JPEG. SVG is rasterized first — charts and diagrams are frequently
+/// vector-only, and dropping them loses the point of the article. Returns `None`
+/// for sources no decoder handles.
 pub fn reencode(bytes: &[u8], profile: ImageProfile) -> Option<(Vec<u8>, &'static str)> {
+    if looks_like_svg(bytes) {
+        let raster = rasterize_svg(bytes, profile)?;
+        return reencode(&raster, profile);
+    }
     let format = image::guess_format(bytes).ok();
     let decoded = image::load_from_memory(bytes)
         .map_err(|e| tracing::debug!("undecodable image: {e}"))
@@ -239,6 +247,74 @@ pub fn reencode(bytes: &[u8], profile: ImageProfile) -> Option<(Vec<u8>, &'stati
     Some((out.into_inner(), "image/jpeg"))
 }
 
+/// True when `bytes` are an SVG document (possibly behind an XML prolog or BOM).
+fn looks_like_svg(bytes: &[u8]) -> bool {
+    let head = &bytes[..bytes.len().min(1024)];
+    let text = String::from_utf8_lossy(head);
+    let text = text.trim_start_matches('\u{feff}').trim_start();
+    text.starts_with("<svg")
+        || (text.starts_with("<?xml") || text.starts_with("<!DOCTYPE svg")) && text.contains("<svg")
+}
+
+/// Rasterize an SVG to a PNG at the profile's target width (§3.10).
+///
+/// The profile's own resize pass then handles the height cap, so this only has
+/// to land in the right ballpark.
+fn rasterize_svg(bytes: &[u8], profile: ImageProfile) -> Option<Vec<u8>> {
+    let mut options = resvg::usvg::Options::default();
+    options.fontdb_mut().load_system_fonts();
+    let tree = resvg::usvg::Tree::from_data(bytes, &options)
+        .map_err(|e| tracing::debug!("svg did not parse: {e}"))
+        .ok()?;
+
+    // An `<svg>` that parses but draws nothing is not an image, it is a stray
+    // tag: rasterizing it would embed a blank rectangle.
+    if tree.root().children().is_empty() {
+        tracing::debug!("svg has nothing to draw");
+        return None;
+    }
+    let size = tree.size();
+    let (sw, sh) = (size.width(), size.height());
+    if !(sw.is_finite() && sh.is_finite()) || sw <= 0.0 || sh <= 0.0 {
+        return None;
+    }
+    // Judge "decorative" by the declared size, before scaling: a 16×16 icon is
+    // an icon however large we choose to draw it.
+    if sw < MIN_DIMENSION_PX as f32 || sh < MIN_DIMENSION_PX as f32 {
+        tracing::debug!(sw, sh, "skipping decorative svg");
+        return None;
+    }
+    // Vector art has no native resolution, so render straight at the edition's
+    // target width — upscaling a rasterized copy afterwards would only blur it.
+    let target_w = profile
+        .max_width
+        .max(SVG_FALLBACK_SIZE.min(profile.max_width));
+    let scale = (target_w as f32 / sw).min(profile.max_height as f32 / sh);
+    let scale = if scale.is_finite() && scale > 0.0 {
+        scale
+    } else {
+        1.0
+    };
+    let (w, h) = (
+        (sw * scale).round().max(1.0) as u32,
+        (sh * scale).round().max(1.0) as u32,
+    );
+    let mut pixmap = tiny_skia::Pixmap::new(w, h)?;
+    // E-ink has no transparency; render onto white so alpha never becomes black.
+    pixmap.fill(tiny_skia::Color::WHITE);
+    resvg::render(
+        &tree,
+        tiny_skia::Transform::from_scale(scale, scale),
+        &mut pixmap.as_mut(),
+    );
+    let rgba = image::RgbaImage::from_raw(w, h, pixmap.take_demultiplied())?;
+    let mut png = Cursor::new(Vec::new());
+    DynamicImage::ImageRgba8(rgba)
+        .write_to(&mut png, ImageFormat::Png)
+        .ok()?;
+    Some(png.into_inner())
+}
+
 /// Composite over an opaque white page — e-ink has no transparency (§3.10).
 fn flatten_to_white(img: &DynamicImage) -> DynamicImage {
     let rgba = img.to_rgba8();
@@ -302,7 +378,6 @@ fn pending_for_pick(pick: &Pick) -> Vec<PendingImage> {
     }
     refs.into_iter()
         .filter(|r| r.src.starts_with("http://") || r.src.starts_with("https://"))
-        .take(MAX_IMAGES_PER_ARTICLE)
         .enumerate()
         .map(|(i, r)| PendingImage {
             id: format!("img-{entry_id}-{i}"),
@@ -409,8 +484,12 @@ pub(crate) fn tag_name(inner: &str) -> String {
         .to_ascii_lowercase()
 }
 
-/// Parse `name="value"` pairs out of a tag body.
-fn parse_attrs(inner: &str) -> Vec<(String, String)> {
+/// Parse `name="value"` pairs out of a tag body, with entity-decoded values.
+///
+/// Decoding matters: this scanner reads markup that ammonia has serialized, and
+/// ammonia writes `&` in a URL as `&amp;`. A raw comparison against a URL that
+/// came out of a real HTML parser would never match (§3.10).
+pub(crate) fn parse_attrs(inner: &str) -> Vec<(String, String)> {
     let mut attrs = Vec::new();
     let bytes: Vec<char> = inner.chars().collect();
     let mut i = 0;
@@ -457,11 +536,60 @@ fn parse_attrs(inner: &str) -> Vec<(String, String)> {
                 }
             }
         }
-        attrs.push((name, value));
+        attrs.push((name, decode_entities(&value)));
     }
     attrs
 }
 
+/// Decode the handful of entities an HTML serializer emits inside attributes.
+///
+/// Numeric forms are included because feeds and WordPress write `&#038;` for
+/// `&`; anything else is left alone rather than guessed at.
+pub(crate) fn decode_entities(s: &str) -> String {
+    if !s.contains('&') {
+        return s.to_string();
+    }
+    let mut out = String::with_capacity(s.len());
+    let mut rest = s;
+    while let Some(i) = rest.find('&') {
+        out.push_str(&rest[..i]);
+        let tail = &rest[i..];
+        let Some(end) = tail[..tail.len().min(12)].find(';') else {
+            out.push('&');
+            rest = &tail[1..];
+            continue;
+        };
+        let entity = &tail[1..end];
+        let decoded = match entity {
+            "amp" => Some('&'),
+            "lt" => Some('<'),
+            "gt" => Some('>'),
+            "quot" => Some('"'),
+            "apos" | "#39" => Some('\''),
+            "nbsp" => Some('\u{a0}'),
+            _ => entity
+                .strip_prefix('#')
+                .and_then(|n| match n.strip_prefix(['x', 'X']) {
+                    Some(hex) => u32::from_str_radix(hex, 16).ok(),
+                    None => n.parse::<u32>().ok(),
+                })
+                .and_then(char::from_u32),
+        };
+        match decoded {
+            Some(c) => {
+                out.push(c);
+                rest = &tail[end + 1..];
+            }
+            None => {
+                out.push('&');
+                rest = &tail[1..];
+            }
+        }
+    }
+    out.push_str(rest);
+    out
+}
+
 /// Escape a string for use inside a double-quoted XML attribute.
 fn attr_escape(s: &str) -> String {
     let mut out = String::with_capacity(s.len());
@@ -491,6 +619,40 @@ pub fn text_escape(s: &str) -> String {
     out
 }
 
+/// Whether an image we could not embed is worth telling the reader about.
+///
+/// Only descriptive alt text qualifies. A filename, a bare label like `red line`
+/// on a divider rule, or no alt at all carries nothing the reader loses by not
+/// seeing the picture — announcing those turns every decorative graphic and
+/// dead link into a line of clutter, which is how the placeholders got out of
+/// hand in the first place.
+fn alt_is_worth_announcing(alt: &str) -> bool {
+    const MIN_DESCRIPTIVE_WORDS: usize = 4;
+    !alt.is_empty()
+        && !is_filename_alt(alt)
+        && alt.split_whitespace().count() >= MIN_DESCRIPTIVE_WORDS
+}
+
+/// True for alt text that is really just the uploaded filename — `IMG_0808.JPG`,
+/// `cut pieces v01.JPG`, `chart-final-2.png`.
+fn is_filename_alt(alt: &str) -> bool {
+    let alt = alt.trim();
+    if alt.contains(' ') && alt.split_whitespace().count() > 4 {
+        return false;
+    }
+    let Some((stem, ext)) = alt.rsplit_once('.') else {
+        return false;
+    };
+    let ext = ext.to_ascii_lowercase();
+    matches!(
+        ext.as_str(),
+        "jpg" | "jpeg" | "png" | "gif" | "webp" | "svg" | "avif" | "bmp" | "heic"
+    ) && !stem.is_empty()
+        && stem
+            .chars()
+            .all(|c| c.is_alphanumeric() || matches!(c, ' ' | '_' | '-' | '.'))
+}
+
 /// Rewrite `<img src>` to the embedded hrefs, replacing misses with the
 /// `[image: alt]` placeholder paragraph (§3.10).
 pub fn rewrite_img_srcs(html: &str, assets: &[ImageAsset]) -> String {
@@ -531,15 +693,15 @@ pub fn rewrite_img_srcs(html: &str, assets: &[ImageAsset]) -> String {
                         attr_escape(alt)
                     ));
                 }
+                // An image we could not embed is only worth announcing when its
+                // alt text tells the reader something; otherwise the `<img>`
+                // just goes away. That covers the decorative graphics the
+                // re-encoder deliberately skips as well as genuine misses (§3.10).
+                None if !alt_is_worth_announcing(&alt) => {}
                 None => {
-                    let label = if alt.is_empty() {
-                        "image unavailable"
-                    } else {
-                        &alt
-                    };
                     out.push_str(&format!(
                         "<p class=\"image-placeholder\">[image: {}]</p>",
-                        text_escape(label)
+                        text_escape(&alt)
                     ));
                 }
             }
@@ -618,20 +780,76 @@ mod tests {
     #[test]
     fn rewrites_hits_and_placeholders_misses() {
         let assets = vec![asset("https://e.g/a.png", "images/img-1-0.jpg")];
-        let html = r#"<p>x</p><img src="https://e.g/a.png" alt="Alt &amp; more"><img src="https://e.g/gone.png" alt="Missing">"#;
+        let html = r#"<p>x</p><img src="https://e.g/a.png" alt="Alt &amp; more"><img src="https://e.g/gone.png" alt="A chart of missing things">"#;
         let out = rewrite_img_srcs(html, &assets);
-        assert!(out.contains(r#"<img src="images/img-1-0.jpg" alt="Alt &amp;amp; more"/>"#));
-        assert!(out.contains(r#"<p class="image-placeholder">[image: Missing]</p>"#));
+        // The alt round-trips through one level of escaping, not two.
+        assert!(out.contains(r#"<img src="images/img-1-0.jpg" alt="Alt &amp; more"/>"#));
+        assert!(
+            out.contains(r#"<p class="image-placeholder">[image: A chart of missing things]</p>"#)
+        );
         assert!(!out.contains("gone.png"));
     }
 
+    /// The whole point of the fix: ammonia writes `&amp;` into the markup, and
+    /// the asset was keyed on the URL a real parser produced.
     #[test]
-    fn placeholder_falls_back_when_alt_is_missing() {
-        let out = rewrite_img_srcs(r#"<img src="https://e.g/x.png">"#, &[]);
+    fn entity_encoded_urls_still_match_their_asset() {
+        let assets = vec![asset(
+            "https://e.g/a.jpg?id=1&width=980",
+            "images/img-1-0.jpg",
+        )];
+        let html = r#"<img src="https://e.g/a.jpg?id=1&amp;width=980" alt="Chart"/>"#;
+        assert!(rewrite_img_srcs(html, &assets).contains(r#"src="images/img-1-0.jpg""#));
+        // The numeric spelling WordPress emits works too.
+        let html = r#"<img src="https://e.g/a.jpg?id=1&#038;width=980" alt="Chart"/>"#;
+        assert!(rewrite_img_srcs(html, &assets).contains(r#"src="images/img-1-0.jpg""#));
+    }
+
+    #[test]
+    fn unembeddable_images_only_speak_up_when_the_alt_says_something() {
+        // No alt at all: the image simply disappears.
         assert_eq!(
-            out,
-            r#"<p class="image-placeholder">[image: image unavailable]</p>"#
+            rewrite_img_srcs(r#"<img src="https://e.g/x.png">"#, &[]),
+            ""
         );
+        // A filename is not a description.
+        assert_eq!(
+            rewrite_img_srcs(r#"<img src="https://e.g/x.png" alt="IMG_0808.JPG">"#, &[]),
+            ""
+        );
+        // Neither is the label on a decorative divider rule.
+        assert_eq!(
+            rewrite_img_srcs(r#"<img src="https://e.g/rule.png" alt="red line">"#, &[]),
+            ""
+        );
+        // A real description is worth keeping.
+        assert!(
+            rewrite_img_srcs(
+                r#"<img src="https://e.g/x.png" alt="A man in a hard hat stands over a well hole">"#,
+                &[]
+            )
+            .contains("[image: A man in a hard hat stands over a well hole]")
+        );
+    }
+
+    #[test]
+    fn filename_alt_detection() {
+        for yes in [
+            "IMG_0808.JPG",
+            "cut pieces v01.JPG",
+            "chart-final-2.png",
+            "diagram.svg",
+        ] {
+            assert!(is_filename_alt(yes), "{yes} should read as a filename");
+        }
+        for no in [
+            "A hydrogen well head",
+            "",
+            "Fig. 3",
+            "The lion-man of Hohlenstein-Stadel, carved from mammoth ivory.",
+        ] {
+            assert!(!is_filename_alt(no), "{no} should read as a description");
+        }
     }
 
     #[test]
@@ -685,7 +903,37 @@ mod tests {
             .write_to(&mut png, ImageFormat::Png)
             .unwrap();
         assert!(reencode(&png.into_inner(), ImageProfile::STANDARD).is_none());
+        // A bare `<svg>` tag with nothing to draw is markup, not a picture.
         assert!(reencode(b"<svg>not an image</svg>", ImageProfile::STANDARD).is_none());
+        assert!(reencode(b"not an image at all", ImageProfile::STANDARD).is_none());
+    }
+
+    #[test]
+    fn svg_charts_are_rasterized_rather_than_dropped() {
+        let svg = br##"<svg xmlns="http://www.w3.org/2000/svg" width="400" height="300">
+            <rect x="10" y="10" width="380" height="280" fill="#3355bb"/>
+            <circle cx="200" cy="150" r="60" fill="#ffcc00"/>
+        </svg>"##;
+        let (bytes, mime) = reencode(svg, ImageProfile::STANDARD).expect("svg rasterizes");
+        assert_eq!(mime, "image/png", "flat colour art stays lossless");
+        let decoded = image::load_from_memory(&bytes).expect("decodable output");
+        // Drawn at the edition's target width, not at the SVG's nominal size.
+        assert_eq!(decoded.dimensions(), (1200, 900));
+        assert!(!decoded.color().has_alpha(), "rendered onto white");
+
+        // An XML prolog and a leading BOM must not hide the format.
+        let with_prolog = format!(
+            "\u{feff}<?xml version=\"1.0\"?>{}",
+            String::from_utf8_lossy(svg)
+        );
+        let (x4, _) = reencode(with_prolog.as_bytes(), ImageProfile::X4).expect("x4 rasterizes");
+        let x4 = image::load_from_memory(&x4).unwrap();
+        assert!(x4.width() <= 480 && x4.height() <= 800);
+
+        // A 16×16 icon is decorative however large we could draw it.
+        let icon = br#"<svg xmlns="http://www.w3.org/2000/svg" width="16" height="16">
+            <rect width="16" height="16"/></svg>"#;
+        assert!(reencode(icon, ImageProfile::STANDARD).is_none());
     }
 
     #[test]
diff --git a/src/extract.rs b/src/extract.rs
index 5dfde4b..894398a 100644
--- a/src/extract.rs
+++ b/src/extract.rs
@@ -11,14 +11,13 @@ use futures::StreamExt;
 use scraper::{Html, Node, Selector};
 use url::Url;
 
+use crate::epub::images::{tag_end, tag_name};
 use crate::types::{Article, ExtractMethod, Extracted};
 
 /// Word count at or above which Miniflux content is treated as full text (§3.3).
 pub const FULL_TEXT_MIN_WORDS: i64 = 250;
 /// Maximum bytes downloaded when fetching an article page (§3.3).
 pub const MAX_FETCH_BYTES: usize = 3 * 1024 * 1024;
-/// Maximum images collected per article (§3.3).
-pub const MAX_IMAGES_PER_ARTICLE: usize = 12;
 /// Note appended to bodies we could only excerpt (§3.3).
 pub const EXCERPT_NOTE: &str = "(excerpt only — read online)";
 
@@ -130,20 +129,37 @@ impl Extractor {
         let _guard = span.enter();
 
         // 1. Miniflux content, when it already looks like full text.
-        let feed_html = sanitize_with_base(&article.content_html, &article.url);
+        let feed_html =
+            sanitize_with_base(&normalize_img_tags(&article.content_html), &article.url);
         let feed_words = word_count(&feed_html);
         if feed_words >= FULL_TEXT_MIN_WORDS {
-            return self.finish(article, feed_html, feed_words, ExtractMethod::Miniflux);
+            return self.finish(
+                article,
+                feed_html,
+                feed_words,
+                ExtractMethod::Miniflux,
+                &article.url,
+            );
         }
 
         // 2. Fetch the page and run readability over it.
         if self.can_fetch() {
             match self.fetch_readable(&article.url).await {
-                Ok(html) => {
-                    let clean = sanitize_with_base(&html, &article.url);
+                Ok(page) => {
+                    // Relative URLs in the markup belong to the page we ended up
+                    // on, not the one we asked for: shortener and syndication
+                    // links land on another host entirely.
+                    let clean =
+                        sanitize_with_base(&normalize_img_tags(&page.html), &page.final_url);
                     let words = word_count(&clean);
                     if words > feed_words && words > 0 {
-                        return self.finish(article, clean, words, ExtractMethod::Readability);
+                        return self.finish(
+                            article,
+                            clean,
+                            words,
+                            ExtractMethod::Readability,
+                            &page.final_url,
+                        );
                     }
                     tracing::debug!(words, feed_words, "readability was not an improvement");
                 }
@@ -160,20 +176,24 @@ impl Extractor {
             format!("{feed_html}<p>{EXCERPT_NOTE}</p>")
         };
         let words = word_count(&body);
-        let mut out = self.finish(article, body, words, ExtractMethod::Excerpt);
+        let mut out = self.finish(article, body, words, ExtractMethod::Excerpt, &article.url);
         out.excerpt_only = true;
         out
     }
 
     /// Assemble the [`Extracted`] value once a body has been chosen.
+    ///
+    /// `base_url` is what the body's relative URLs resolve against — the page we
+    /// landed on for a fetched article, the article URL otherwise.
     fn finish(
         &self,
         article: &Article,
         content_html: String,
         words: i64,
         method: ExtractMethod,
+        base_url: &str,
     ) -> Extracted {
-        let image_urls = collect_image_urls(&content_html, &article.url);
+        let image_urls = collect_image_urls(&content_html, base_url);
         let excerpt_only = method == ExtractMethod::Excerpt
             || looks_paywalled(&article.url, words, &self.paywall_domains);
         Extracted {
@@ -220,7 +240,10 @@ impl Extractor {
 
     /// Fetch `url` (10s timeout, desktop UA, [`MAX_FETCH_BYTES`] cap) and run
     /// `dom_smoothie` readability over it (§3.3).
-    pub async fn fetch_readable(&self, url: &str) -> Result<String, ExtractError> {
+    ///
+    /// Returns the URL the fetch actually landed on alongside the markup, so
+    /// callers resolve relative links against the right origin.
+    pub async fn fetch_readable(&self, url: &str) -> Result<Page, ExtractError> {
         let Some(http) = &self.http else {
             return Err(ExtractError::FetchDisabled);
         };
@@ -237,6 +260,7 @@ impl Extractor {
         if !response.status().is_success() {
             return Err(ExtractError::Status(response.status().as_u16()));
         }
+        let final_url = response.url().to_string();
         if let Some(ct) = response
             .headers()
             .get(reqwest::header::CONTENT_TYPE)
@@ -256,17 +280,36 @@ impl Extractor {
             body.extend_from_slice(&chunk);
         }
         let html = String::from_utf8_lossy(&body).into_owned();
-        readability(&html, url)
+        Ok(Page {
+            html: readability(&html, &final_url)?,
+            final_url,
+        })
     }
 }
 
+/// An article page after fetching and readability (§3.3).
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct Page {
+    /// Readability's main-content markup.
+    pub html: String,
+    /// Where the fetch ended up, after any redirects — the base for relative URLs.
+    pub final_url: String,
+}
+
 /// Run `dom_smoothie` over a fetched page and return its main-content HTML (§3.3).
+///
+/// The page is normalized first ([`prepare_for_readability`]) so that readability
+/// sees plain, already-resolved `<img src>` elements. Left to itself it damages
+/// them in two ways: it deletes whole subtrees (lightbox `<button>` wrappers take
+/// their image with them) and its lazy-image heuristic overwrites a perfectly
+/// good `src` with whatever other attribute happens to contain `.jpg`.
 pub fn readability(html: &str, url: &str) -> Result<String, ExtractError> {
+    let html = prepare_for_readability(html);
     let config = dom_smoothie::Config {
         max_elements_to_parse: 60_000,
         ..Default::default()
     };
-    let mut readability = dom_smoothie::Readability::new(html, Some(url), Some(config))
+    let mut readability = dom_smoothie::Readability::new(html.as_str(), Some(url), Some(config))
         .map_err(|_| ExtractError::NoContent)?;
     let parsed = readability.parse().map_err(|_| ExtractError::NoContent)?;
     let content = parsed.content.to_string();
@@ -300,6 +343,303 @@ fn merge_paywall_domains(configured: Vec<String>) -> Vec<String> {
     domains
 }
 
+// ---------------------------------------------------------------------------
+// Image normalization (§3.3)
+// ---------------------------------------------------------------------------
+
+/// Elements that readability deletes outright, and which a page may nevertheless
+/// have wrapped around an image (lightbox triggers, mostly).
+const IMAGE_WRAPPER_TAGS: &[&str] = &["button", "form", "fieldset", "object"];
+
+/// Attributes lazy-loading libraries use for the real image URL.
+///
+/// These outrank `src`, because a page only sets them when `src` is a stand-in:
+/// a transparent GIF, a blurred thumbnail, an inline SVG spacer. Attributes that
+/// merely *look* image-ish (`data-template`, `data-attrs`, `data-orig-file`) are
+/// deliberately absent — those hold templates and metadata, and preferring them
+/// is exactly the mistake readability's own heuristic makes.
+const LAZY_SRC_ATTRS: &[&str] = &[
+    "data-src",
+    "data-lazy-src",
+    "data-original",
+    "data-runner-src",
+    "data-full-src",
+    "data-hi-res-src",
+    "data-image-src",
+];
+
+/// Widest `srcset` candidate we will pick; above this we are downloading pixels
+/// the re-encoder immediately throws away.
+const MAX_SRCSET_WIDTH: u32 = 2000;
+
+/// Make a fetched page safe to hand to readability (§3.3).
+///
+/// Two passes, both about images: unwrap the elements that would take an image
+/// with them when readability deletes them, then reduce every `<img>` to a plain
+/// `src`/`alt`/`title` triple. The second pass is what stops readability's own
+/// lazy-image heuristic from replacing a working `src` — with no `srcset`,
+/// `loading` or `data-*` attributes left on the element, it has nothing to
+/// substitute and leaves the image alone.
+pub fn prepare_for_readability(html: &str) -> String {
+    normalize_img_tags(&unwrap_image_wrappers(html))
+}
+
+/// Replace image-only `<button>`/`<form>`/`<fieldset>`/`<object>` wrappers with
+/// their contents (§3.3).
+///
+/// A `<button>` holding nothing but an image is a lightbox trigger, not a
+/// control: the image is the content. Wrappers that also carry text are left
+/// alone, because those really are interface.
+pub fn unwrap_image_wrappers(html: &str) -> String {
+    let mut out = String::with_capacity(html.len());
+    let mut cursor = 0usize;
+    while let Some(rel) = html[cursor..].find('<') {
+        let start = cursor + rel;
+        out.push_str(&html[cursor..start]);
+        let Some(end) = tag_end(html, start) else {
+            out.push_str(&html[start..]);
+            return out;
+        };
+        let raw = &html[start..end];
+        let inner = raw.trim_start_matches('<').trim_end_matches('>');
+        let name = tag_name(inner);
+
+        if IMAGE_WRAPPER_TAGS.contains(&name.as_str())
+            && !inner.trim_end().ends_with('/')
+            && let Some((content, after)) = element_content(html, end, &name)
+            && content.contains("<img")
+            && html_to_text(content).trim().is_empty()
+        {
+            out.push_str(&unwrap_image_wrappers(content));
+            cursor = after;
+            continue;
+        }
+
+        out.push_str(raw);
+        cursor = end;
+    }
+    out.push_str(&html[cursor..]);
+    out
+}
+
+/// The content of an element whose open tag ended at `body_start`, plus the
+/// offset just past its close tag. `None` when the element is never closed.
+fn element_content<'a>(html: &'a str, body_start: usize, name: &str) -> Option<(&'a str, usize)> {
+    let open = format!("<{name}");
+    let close = format!("</{name}");
+    let mut depth = 1usize;
+    let mut cursor = body_start;
+    while let Some(rel) = html[cursor..].find('<') {
+        let start = cursor + rel;
+        let end = tag_end(html, start)?;
+        let tag = &html[start..end];
+        let lower = tag.to_ascii_lowercase();
+        if lower.starts_with(&close) {
+            depth -= 1;
+            if depth == 0 {
+                return Some((&html[body_start..start], end));
+            }
+        } else if lower.starts_with(&open)
+            && !lower[open.len()..].starts_with(|c: char| c.is_alphanumeric() || c == '-')
+            && !tag.trim_end().ends_with("/>")
+        {
+            depth += 1;
+        }
+        cursor = end;
+    }
+    None
+}
+
+/// Reduce every `<img>` to `<img src alt title>` with a usable URL (§3.3).
+///
+/// The `src` a page ships is not automatically the one to use: it can be a lazy
+/// placeholder (`data:image/svg+xml,…`), an unresolved template
+/// (`…/resize/{width}/…`), a JSON blob a framework parked there, or an entire
+/// `srcset` string. Candidates are tried in order and the first plausible one
+/// wins; an image with no plausible candidate is dropped, because a broken
+/// `<img>` only becomes clutter further down the pipeline.
+pub fn normalize_img_tags(html: &str) -> String {
+    let mut out = String::with_capacity(html.len());
+    let mut cursor = 0usize;
+    // `<picture>` puts the real candidates on sibling `<source>` elements.
+    let mut picture_srcset: Option<String> = None;
+
+    while let Some(rel) = html[cursor..].find('<') {
+        let start = cursor + rel;
+        out.push_str(&html[cursor..start]);
+        let Some(end) = tag_end(html, start) else {
+            out.push_str(&html[start..]);
+            return out;
+        };
+        let raw = &html[start..end];
+        let inner = raw
+            .trim_start_matches('<')
+            .trim_end_matches('>')
+            .trim_end_matches('/');
+        let name = tag_name(inner);
+
+        match name.as_str() {
+            "picture" => {
+                picture_srcset = None;
+                out.push_str(raw);
+            }
+            "source" => {
+                let attrs = crate::epub::images::parse_attrs(inner);
+                if picture_srcset.is_none()
+                    && let Some(set) =
+                        attr(&attrs, "srcset").or_else(|| attr(&attrs, "data-srcset"))
+                {
+                    picture_srcset = Some(set.to_string());
+                }
+                out.push_str(raw);
+            }
+            "img" => {
+                let attrs = crate::epub::images::parse_attrs(inner);
+                if let Some(src) = best_img_src(&attrs, picture_srcset.as_deref()) {
+                    out.push_str("<img src=\"");
+                    out.push_str(&escape_attr(&src));
+                    out.push('"');
+                    for key in ["alt", "title"] {
+                        if let Some(v) = attr(&attrs, key) {
+                            out.push(' ');
+                            out.push_str(key);
+                            out.push_str("=\"");
+                            out.push_str(&escape_attr(v));
+                            out.push('"');
+                        }
+                    }
+                    out.push_str("/>");
+                } else {
+                    tracing::debug!(tag = %&raw[..raw.len().min(120)], "dropping unusable img");
+                }
+            }
+            _ => out.push_str(raw),
+        }
+        if name == "picture" && inner.starts_with('/') {
+            picture_srcset = None;
+        }
+        cursor = end;
+    }
+    out.push_str(&html[cursor..]);
+    out
+}
+
+fn attr<'a>(attrs: &'a [(String, String)], name: &str) -> Option<&'a str> {
+    attrs
+        .iter()
+        .find(|(k, _)| k == name)
+        .map(|(_, v)| v.trim())
+        .filter(|v| !v.is_empty())
+}
+
+/// Pick the best URL for one `<img>` from everything the element carries.
+fn best_img_src(attrs: &[(String, String)], picture_srcset: Option<&str>) -> Option<String> {
+    for key in LAZY_SRC_ATTRS {
+        if let Some(v) = attr(attrs, key).filter(|s| plausible_url(s)) {
+            return Some(v.to_string());
+        }
+    }
+    if let Some(src) = attr(attrs, "src").filter(|s| plausible_url(s)) {
+        return Some(src.to_string());
+    }
+    if let Some(from_set) = attr(attrs, "srcset").and_then(best_from_srcset) {
+        return Some(from_set);
+    }
+    if let Some(from_set) = attr(attrs, "data-srcset").and_then(best_from_srcset) {
+        return Some(from_set);
+    }
+    picture_srcset.and_then(best_from_srcset)
+}
+
+/// The widest candidate in a `srcset` that is still worth downloading.
+///
+/// Parsed by whitespace rather than by comma: the URLs of several image CDNs
+/// contain commas of their own, and splitting on those shreds them.
+fn best_from_srcset(srcset: &str) -> Option<String> {
+    let mut best: Option<(u32, String)> = None;
+    let mut smallest: Option<(u32, String)> = None;
+    let mut pending: Option<String> = None;
+
+    let mut consider = |url: String, width: u32| {
+        if width <= MAX_SRCSET_WIDTH && best.as_ref().is_none_or(|(w, _)| width > *w) {
+            best = Some((width, url.clone()));
+        }
+        if smallest.as_ref().is_none_or(|(w, _)| width < *w) {
+            smallest = Some((width, url));
+        }
+    };
+
+    for token in srcset.split_whitespace() {
+        let token = token.trim_end_matches(',');
+        if token.is_empty() {
+            continue;
+        }
+        match parse_descriptor(token) {
+            Some(width) => {
+                if let Some(url) = pending.take() {
+                    consider(url, width);
+                }
+            }
+            None => {
+                // A URL with no descriptor of its own still counts, at width 1x.
+                if let Some(url) = pending.replace(token.to_string()) {
+                    consider(url, 1);
+                }
+            }
+        }
+    }
+    if let Some(url) = pending.take() {
+        consider(url, 1);
+    }
+
+    best.or(smallest)
+        .map(|(_, url)| url)
+        .filter(|u| plausible_url(u))
+}
+
+/// `800w` → 800, `2x` → a synthetic width so density candidates sort sensibly.
+fn parse_descriptor(token: &str) -> Option<u32> {
+    let (value, unit) = token.split_at(token.len().checked_sub(1)?);
+    match unit {
+        "w" => value.parse::<u32>().ok(),
+        "x" => value
+            .parse::<f32>()
+            .ok()
+            .map(|d| (d * 1000.0).round().clamp(1.0, f32::from(u16::MAX)) as u32),
+        _ => None,
+    }
+}
+
+/// Whether a string can serve as an image URL at all.
+///
+/// This is deliberately about shape, not about the host: whitespace, braces and
+/// quotes mean we are looking at a `srcset` blob, an unfilled URL template or a
+/// serialized object, none of which will ever resolve.
+fn plausible_url(candidate: &str) -> bool {
+    let candidate = candidate.trim();
+    if candidate.is_empty() || candidate.len() > 2048 {
+        return false;
+    }
+    if candidate.starts_with("data:") || candidate.starts_with("about:") {
+        return false;
+    }
+    if candidate
+        .chars()
+        .any(|c| c.is_whitespace() || matches!(c, '{' | '}' | '"' | '\'' | '<' | '>' | '\\'))
+    {
+        return false;
+    }
+    // `%20` is a space that survived encoding — same blob, different spelling.
+    !candidate.contains("%20")
+}
+
+fn escape_attr(s: &str) -> String {
+    s.replace('&', "&amp;")
+        .replace('<', "&lt;")
+        .replace('>', "&gt;")
+        .replace('"', "&quot;")
+}
+
 // ---------------------------------------------------------------------------
 // Sanitization (§3.3)
 // ---------------------------------------------------------------------------
@@ -404,8 +744,10 @@ pub fn word_count(html: &str) -> i64 {
         .count() as i64
 }
 
-/// Absolute image URLs referenced by `html`, resolved against `base_url`,
-/// capped at [`MAX_IMAGES_PER_ARTICLE`] (§3.3).
+/// Absolute image URLs referenced by `html`, resolved against `base_url` (§3.3).
+///
+/// Every image an article carries is kept: a photo essay with thirty pictures is
+/// a photo essay, and the issue-wide byte budget is the real backstop.
 pub fn collect_image_urls(html: &str, base_url: &str) -> Vec<String> {
     let Ok(selector) = Selector::parse("img") else {
         return Vec::new();
@@ -433,9 +775,6 @@ pub fn collect_image_urls(html: &str, base_url: &str) -> Vec<String> {
         if seen.insert(url.clone()) {
             out.push(url);
         }
-        if out.len() >= MAX_IMAGES_PER_ARTICLE {
-            break;
-        }
     }
     out
 }
@@ -598,14 +937,15 @@ mod tests {
     }
 
     #[test]
-    fn image_collection_resolves_and_caps() {
+    fn image_collection_resolves_and_keeps_every_image() {
         let mut html = String::from(r#"<img src="/a.png"><img src="https://cdn.dev/b.png">"#);
         html.push_str(r#"<img data-src="c.png"><img src="/a.png"><img src="data:image/png;x">"#);
         for i in 0..20 {
             html.push_str(&format!(r#"<img src="/n{i}.png">"#));
         }
         let urls = collect_image_urls(&html, "https://blog.dev/posts/one");
-        assert_eq!(urls.len(), MAX_IMAGES_PER_ARTICLE);
+        // Two named images, the data-src one, and all twenty of the rest: no cap.
+        assert_eq!(urls.len(), 23);
         assert_eq!(urls[0], "https://blog.dev/a.png");
         assert_eq!(urls[1], "https://cdn.dev/b.png");
         assert_eq!(urls[2], "https://blog.dev/posts/c.png");
@@ -615,6 +955,221 @@ mod tests {
         assert!(collect_image_urls("<p>none</p>", "https://blog.dev").is_empty());
     }
 
+    // -----------------------------------------------------------------------
+    // Image normalization — each case is a page shape seen in a real issue.
+    // -----------------------------------------------------------------------
+
+    fn src_of(html: &str) -> Vec<String> {
+        Html::parse_fragment(html)
+            .select(&Selector::parse("img").unwrap())
+            .filter_map(|e| e.value().attr("src").map(str::to_string))
+            .collect()
+    }
+
+    #[test]
+    fn lazy_placeholder_src_gives_way_to_the_real_url() {
+        // IEEE Spectrum: an inline SVG spacer with the URL parked on data-runner-src.
+        let html = r#"<img alt="A well head" lazy-loadable="true"
+            src="data:image/svg+xml,%3Csvg%20xmlns=%27http://www.w3.org/2000/svg%27%3E%3C/svg%3E"
+            data-runner-src="https://spectrum.ieee.org/media-library/well.jpg?id=675&amp;width=980"/>"#;
+        assert_eq!(
+            src_of(&normalize_img_tags(html)),
+            ["https://spectrum.ieee.org/media-library/well.jpg?id=675&width=980"]
+        );
+    }
+
+    #[test]
+    fn unresolved_url_templates_fall_through_to_a_real_candidate() {
+        // NPR: readability copies data-template over a perfectly good src.
+        let html = r#"<img alt="Meghan Cliffel"
+            src="https://npr.brightspotcdn.com/resize/{width}/quality/{quality}/x.jpg"
+            srcset="https://npr.brightspotcdn.com/resize/1100/quality/50/x.jpg 1100w"/>"#;
+        assert_eq!(
+            src_of(&normalize_img_tags(html)),
+            ["https://npr.brightspotcdn.com/resize/1100/quality/50/x.jpg"]
+        );
+        // With nothing usable anywhere, the image goes rather than becoming a
+        // request for a picture that says "Image".
+        let only_template =
+            r#"<img alt="x" src="https://cdn.dev/resize/{width}/quality/{quality}/x.jpg"/>"#;
+        assert!(src_of(&normalize_img_tags(only_template)).is_empty());
+    }
+
+    #[test]
+    fn a_srcset_blob_parked_in_src_is_rejected_and_reparsed() {
+        // dfarq: readability copies the entire srcset string into src.
+        let blob = "https://i0.wp.com/x.jpg?resize=300%2C158&ssl=1 300w, \
+                    https://i0.wp.com/x.jpg?resize=1024%2C540&ssl=1 1024w, \
+                    https://i0.wp.com/x.jpg?w=3000&ssl=1 3000w";
+        let html = format!(r#"<img alt="printer" src="{blob}" srcset="{blob}"/>"#);
+        // The widest candidate under the download ceiling wins; 3000w does not.
+        assert_eq!(
+            src_of(&normalize_img_tags(&html)),
+            ["https://i0.wp.com/x.jpg?resize=1024%2C540&ssl=1"]
+        );
+    }
+
+    #[test]
+    fn a_json_blob_parked_in_src_is_rejected() {
+        // Substack: readability copies data-attrs (JSON) over src.
+        let html = r#"<img alt="" src="{&quot;src&quot;:&quot;https://s3.dev/a.jpeg&quot;,&quot;width&quot;:1000}"
+            srcset="https://substackcdn.com/image/fetch/$s_!y1,w_1456,c_limit/a.jpeg 1456w"/>"#;
+        assert_eq!(
+            src_of(&normalize_img_tags(html)),
+            ["https://substackcdn.com/image/fetch/$s_!y1,w_1456,c_limit/a.jpeg"]
+        );
+    }
+
+    #[test]
+    fn picture_sources_back_up_an_empty_img() {
+        let html = r#"<picture>
+            <source srcset="https://cdn.dev/a.avif 800w" type="image/avif"/>
+            <source srcset="https://cdn.dev/a.webp 800w" type="image/webp"/>
+            <img alt="A photo"/>
+        </picture>"#;
+        assert_eq!(
+            src_of(&normalize_img_tags(html)),
+            ["https://cdn.dev/a.avif"]
+        );
+    }
+
+    #[test]
+    fn a_good_src_is_never_traded_for_a_metadata_attribute() {
+        // `data-template`, `data-attrs` and friends hold templates and JSON, not
+        // URLs — trading a working src for one of those is the original sin.
+        let html = r#"<img alt="A photo" loading="lazy" class="lazyload"
+            src="https://cdn.dev/real.jpg"
+            data-template="https://cdn.dev/{width}/real.jpg"
+            data-orig-file="https://cdn.dev/orig.jpg"/>"#;
+        assert_eq!(
+            src_of(&normalize_img_tags(html)),
+            ["https://cdn.dev/real.jpg"]
+        );
+    }
+
+    #[test]
+    fn a_lazy_loader_placeholder_loses_to_its_data_src() {
+        // iRunFar (a3-lazy-load): src is a shared 1×1 spacer that would download
+        // and re-encode perfectly happily, and be the wrong picture.
+        let html = r#"<img class="lazy lazy-hidden" alt="Brooks Cascadia 20"
+            src="//www.irunfar.com/wp-content/plugins/a3-lazy-load/assets/images/lazy_placeholder.gif"
+            data-src="https://s3.amazonaws.com/www.irunfar.com/uploads/Brooks-Cascadia-20.jpg"/>"#;
+        assert_eq!(
+            src_of(&normalize_img_tags(html)),
+            ["https://s3.amazonaws.com/www.irunfar.com/uploads/Brooks-Cascadia-20.jpg"]
+        );
+    }
+
+    #[test]
+    fn normalization_keeps_alt_and_title_and_drops_the_rest() {
+        let html = r#"<img src="/a.png" alt="An &amp; alt" title="T" class="x" width="900"
+            onerror="evil()" srcset="/b.png 2x"/>"#;
+        let out = normalize_img_tags(html);
+        assert!(out.contains(r#"alt="An &amp; alt""#), "{out}");
+        assert!(out.contains(r#"title="T""#), "{out}");
+        for gone in ["class=", "width=", "onerror", "srcset="] {
+            assert!(
+                !out.contains(gone),
+                "expected {gone} to be dropped from {out}"
+            );
+        }
+        // Relative URLs survive: sanitize_with_base absolutizes them later.
+        assert_eq!(src_of(&out), ["/a.png"]);
+    }
+
+    #[test]
+    fn lightbox_buttons_no_longer_take_their_image_with_them() {
+        // Nautilus: readability deletes <button> and everything inside it.
+        let html = r#"<figure class="wp-block-image">
+            <button type="button" aria-label="Enlarge image">
+                <img class="wp-image-1" src="https://cdn.dev/flower.png?w=710" alt=""/>
+            </button>
+            <figcaption>BEAUTIFUL DANGER: a belladonna flower.</figcaption>
+        </figure>"#;
+        let out = unwrap_image_wrappers(html);
+        assert!(!out.contains("<button"), "{out}");
+        assert!(!out.contains("</button>"), "{out}");
+        assert!(out.contains("flower.png"), "{out}");
+        assert!(out.contains("BEAUTIFUL DANGER"), "caption survives: {out}");
+    }
+
+    #[test]
+    fn buttons_that_are_really_buttons_are_left_alone() {
+        let html = r#"<button class="subscribe">Subscribe <img src="/icon.png" alt=""/></button>"#;
+        assert_eq!(unwrap_image_wrappers(html), html);
+        // And an image-free control is untouched too.
+        let plain = r#"<button>Share</button>"#;
+        assert_eq!(unwrap_image_wrappers(plain), plain);
+    }
+
+    #[test]
+    fn nested_and_unclosed_wrappers_do_not_derail_the_scan() {
+        let nested = r#"<button><button><img src="/a.png"/></button></button><p>after</p>"#;
+        let out = unwrap_image_wrappers(nested);
+        assert!(!out.contains("button"), "{out}");
+        assert!(out.contains("/a.png") && out.contains("after"), "{out}");
+        // An open tag that never closes is passed through rather than eating
+        // the rest of the document.
+        let unclosed = r#"<button><img src="/a.png"/><p>rest</p>"#;
+        assert!(unwrap_image_wrappers(unclosed).contains("rest"));
+    }
+
+    #[test]
+    fn srcset_descriptors_pick_the_widest_usable_candidate() {
+        assert_eq!(
+            best_from_srcset("/a.png 150w, /b.png 800w, /c.png 4000w").as_deref(),
+            Some("/b.png")
+        );
+        // Density descriptors work as an ordering too.
+        assert_eq!(
+            best_from_srcset("/a.png 1x, /b.png 2x").as_deref(),
+            Some("/b.png")
+        );
+        // A bare URL with no descriptor is still a candidate.
+        assert_eq!(best_from_srcset("/only.png").as_deref(), Some("/only.png"));
+        // Every candidate too wide: take the narrowest rather than nothing.
+        assert_eq!(
+            best_from_srcset("/big.png 3000w, /huge.png 5000w").as_deref(),
+            Some("/big.png")
+        );
+        assert_eq!(best_from_srcset("   ").as_deref(), None);
+    }
+
+    #[test]
+    fn plausibility_is_about_shape_not_host() {
+        for good in [
+            "https://cdn.dev/a.jpg?id=1&width=980",
+            "/_next/image?url=https%3A%2F%2Fx.dev%2Fa.png&w=3840&q=75",
+            "https://substackcdn.com/image/fetch/$s_!y1,w_1456/a.jpeg",
+        ] {
+            assert!(plausible_url(good), "{good} should be usable");
+        }
+        for bad in [
+            "",
+            "data:image/svg+xml,%3Csvg%3E%3C/svg%3E",
+            "https://cdn.dev/resize/{width}/a.jpg",
+            r#"{"src":"https://cdn.dev/a.jpg"}"#,
+            "https://cdn.dev/a.jpg 300w, https://cdn.dev/b.jpg 600w",
+            "https://cdn.dev/a.jpg%20300w",
+        ] {
+            assert!(!plausible_url(bad), "{bad} should be rejected");
+        }
+    }
+
+    #[tokio::test]
+    async fn feed_content_is_normalized_too() {
+        // Feeds carry the same lazy markup pages do.
+        let extractor = Extractor::offline(vec![]);
+        let body = format!(
+            r#"<p>{}</p><img src="data:image/gif;base64,zz" data-src="https://cdn.dev/real.jpg"/>"#,
+            "lorem ".repeat(400)
+        );
+        let article = article("https://blog.dev/p", &body);
+        let out = extractor.extract(&article).await;
+        assert_eq!(out.method, ExtractMethod::Miniflux);
+        assert_eq!(out.image_urls, ["https://cdn.dev/real.jpg"]);
+    }
+
     #[test]
     fn paywall_heuristic() {
         let domains = merge_paywall_domains(vec!["paywalled.dev".into()]);
@@ -685,6 +1240,41 @@ mod tests {
         assert!(articles[1].excerpt_only);
     }
 
+    /// A redirected article's relative image URLs belong to where it landed.
+    ///
+    /// `postgr.es/p/9sl` redirects to `boringsql.com`; resolving its `/images/…`
+    /// against the shortener produced two 404s and two lost charts.
+    #[tokio::test]
+    async fn relative_urls_resolve_against_the_url_we_landed_on() {
+        use axum::response::{Html, Redirect};
+        use axum::routing::get;
+
+        let page = format!(
+            "<html><body><article><h1>Post</h1>{}\
+             <img src=\"images/chart.svg\" alt=\"A chart\"/></article></body></html>",
+            "<p>Body copy that readability will happily keep. </p>".repeat(40)
+        );
+        let app = axum::Router::new()
+            .route("/p/9sl", get(|| async { Redirect::to("/posts/real/") }))
+            .route("/posts/real/", get(move || async move { Html(page) }));
+        let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
+        let addr = listener.local_addr().unwrap();
+        let server = tokio::spawn(async move { axum::serve(listener, app).await });
+
+        let http = crate::http::build_client(Duration::from_secs(5)).unwrap();
+        let extractor = Extractor::new(http, vec![]);
+        let article = article(&format!("http://{addr}/p/9sl"), "<p>stub</p>");
+        let out = extractor.extract(&article).await;
+        server.abort();
+
+        assert_eq!(out.method, ExtractMethod::Readability);
+        assert_eq!(
+            out.image_urls,
+            [format!("http://{addr}/posts/real/images/chart.svg")],
+            "the shortener path must not be the base"
+        );
+    }
+
     #[tokio::test]
     async fn offline_extractor_never_fetches() {
         let extractor = Extractor::offline(vec![]);