bugfix: SSRF allowlist, image size cap, robust session detect (0.34.1)

Four crawler defences in one PR (all four threats the review flagged in §3 of REVIEW.md): - New crawler::safety module with is_safe_url + accumulate_capped + fetch_bytes_capped. Rejects non-http(s) schemes, RFC1918 / loopback / link-local / CGNAT / ULA / IPv6-link-local hosts, and any host not on the operator's allowlist (defaults to CRAWLER_START_URL host + CRAWLER_CDN_HOST + CRAWLER_DOWNLOAD_ALLOWLIST extras). - Streaming size cap (CRAWLER_MAX_IMAGE_BYTES, default 32 MiB) so a 10 GiB \"image\" can't fill memory before disk. - looks_like_image() reject path: non-image bytes fail the chapter or cover instead of being stored as .bin and served as application/octet-stream. - session::classify_chapter_probe: three-way classifier replaces the binary #avatar_menu check at content.rs:115. A transient hiccup (broken-page body, or logged-in-but-no-reader) now retries with backoff instead of falsely freezing every worker on session_expired. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-28 20:00:22 +02:00
14 changed files with 816 additions and 179 deletions
--- a/.env.example
+++ b/.env.example
@@ -44,6 +44,14 @@ MAX_REQUEST_BYTES=209715200
 # Default 20 MiB.
 MAX_FILE_BYTES=20971520

+# ----- Crawler download safety -----
+# Hosts the crawler is allowed to fetch images/covers from, in addition
+# to CRAWLER_START_URL's host and CRAWLER_CDN_HOST. Comma-separated.
+# Defends against SSRF via scraped <img src="http://10.0.0.1/...">.
+CRAWLER_DOWNLOAD_ALLOWLIST=
+# Hard cap on a single image body. Default 32 MiB.
+CRAWLER_MAX_IMAGE_BYTES=33554432
+
 # ----- Frontend -----
 # The frontend container runs SvelteKit's Node adapter on :3000 and
 # proxies /api/* to BACKEND_URL via src/hooks.server.ts. In compose the
@@ -51,8 +59,3 @@ MAX_FILE_BYTES=20971520
 # internal docker network. Override only if you're running the
 # frontend container against a backend somewhere else.
 BACKEND_URL=http://backend:8080
-# Per-request wall-clock cap for the /api/* reverse proxy (milliseconds).
-# Default 300000 (5 min) covers a typical 200 MiB chapter upload over
-# 25 Mbps; raise for users on slower upstream links or lower if a
-# tighter front proxy already bounds the request lifetime.
-BACKEND_PROXY_TIMEOUT_MS=300000
--- a/backend/Cargo.lock
+++ b/backend/Cargo.lock
@@ -1470,7 +1470,7 @@ checksum = "c41e0c4fef86961ac6d6f8a82609f55f31b05e4fce149ac5710e439df7619ba4"

 [[package]]
 name = "mangalord"
-version = "0.34.0"
+version = "0.34.1"
 dependencies = [
 "anyhow",
 "argon2",
@@ -2324,6 +2324,7 @@ dependencies = [
 "cookie",
 "cookie_store",
 "futures-core",
+ "futures-util",
 "http",
 "http-body",
 "http-body-util",
@@ -2343,12 +2344,14 @@ dependencies = [
 "sync_wrapper",
 "tokio",
 "tokio-rustls",
+ "tokio-util",
 "tower",
 "tower-http",
 "tower-service",
 "url",
 "wasm-bindgen",
 "wasm-bindgen-futures",
+ "wasm-streams",
 "web-sys",
 "webpki-roots",
 ]
@@ -3527,6 +3530,19 @@ dependencies = [
 "wasmparser",
 ]

+[[package]]
+name = "wasm-streams"
+version = "0.4.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "15053d8d85c7eccdbefef60f06769760a563c7f0a9d6902a13d35c7800b0ad65"
+dependencies = [
+ "futures-util",
+ "js-sys",
+ "wasm-bindgen",
+ "wasm-bindgen-futures",
+ "web-sys",
+]
+
 [[package]]
 name = "wasmparser"
 version = "0.244.0"
--- a/backend/Cargo.toml
+++ b/backend/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "mangalord"
-version = "0.34.0"
+version = "0.34.1"
 edition = "2021"
 default-run = "mangalord"

@@ -46,7 +46,7 @@ futures-util = "0.3"
 bytes = "1"
 chromiumoxide = { version = "0.7", features = ["tokio-runtime", "_fetcher-rusttls-tokio"], default-features = false }
 scraper = "0.20"
-reqwest = { version = "0.12", default-features = false, features = ["rustls-tls", "socks", "cookies"] }
+reqwest = { version = "0.12", default-features = false, features = ["rustls-tls", "socks", "cookies", "stream"] }

 [dev-dependencies]
 tempfile = "3"
--- a/backend/src/app.rs
+++ b/backend/src/app.rs
@@ -19,6 +19,7 @@ use crate::crawler::daemon::{self, ChapterDispatcher, DaemonConfig, MetadataPass
 use crate::crawler::jobs::JobPayload;
 use crate::crawler::pipeline::{self, MetadataStats};
 use crate::crawler::rate_limit::HostRateLimiters;
+use crate::crawler::safety::DownloadAllowlist;
 use crate::crawler::session;
 use crate::crawler::source::{target as target_source, DiscoverMode};
 use crate::repo;
@@ -153,6 +154,8 @@ async fn spawn_crawler_daemon(
            start_url: url.clone(),
            mode_pref: cfg.mode,
            incremental_stop_after: cfg.incremental_stop_after,
+            download_allowlist: cfg.download_allowlist.clone(),
+            max_image_bytes: cfg.max_image_bytes,
        });
        m
    });
@@ -163,6 +166,8 @@ async fn spawn_crawler_daemon(
        storage: Arc::clone(&storage),
        http,
        rate: Arc::clone(&rate),
+        download_allowlist: cfg.download_allowlist.clone(),
+        max_image_bytes: cfg.max_image_bytes,
    });

    // Shared cancellation: daemon shutdown cancels the BrowserManager's
@@ -216,6 +221,8 @@ struct RealMetadataPass {
    start_url: String,
    mode_pref: CrawlerModePref,
    incremental_stop_after: usize,
+    download_allowlist: DownloadAllowlist,
+    max_image_bytes: usize,
 }

 #[async_trait]
@@ -238,6 +245,8 @@ impl MetadataPass for RealMetadataPass {
            0,
            false,
            mode,
+            &self.download_allowlist,
+            self.max_image_bytes,
        )
        .await
    }
@@ -293,6 +302,8 @@ struct RealChapterDispatcher {
    storage: Arc<dyn Storage>,
    http: reqwest::Client,
    rate: Arc<HostRateLimiters>,
+    download_allowlist: DownloadAllowlist,
+    max_image_bytes: usize,
 }

 #[async_trait]
@@ -331,6 +342,8 @@ impl ChapterDispatcher for RealChapterDispatcher {
                    manga_id,
                    &source_url,
                    false,
+                    &self.download_allowlist,
+                    self.max_image_bytes,
                )
                .await?;
                drop(lease);
--- a/backend/src/bin/crawler.rs
+++ b/backend/src/bin/crawler.rs
@@ -229,6 +229,33 @@ async fn run(
    }
    let rate = Arc::new(rate);

+    // SSRF defence: only download from the catalog host + CDN host
+    // (plus optional CRAWLER_DOWNLOAD_ALLOWLIST extras), and cap
+    // single-image downloads at CRAWLER_MAX_IMAGE_BYTES bytes.
+    let mut allowlist =
+        mangalord::crawler::safety::DownloadAllowlist::new();
+    if let Ok(parsed) = reqwest::Url::parse(start_url) {
+        if let Some(h) = parsed.host_str() {
+            allowlist = allowlist.allow(h);
+        }
+    }
+    if let Some(host) = cdn_host {
+        allowlist = allowlist.allow(host);
+    }
+    if let Ok(extras) = std::env::var("CRAWLER_DOWNLOAD_ALLOWLIST") {
+        for piece in extras.split(',') {
+            let trimmed = piece.trim();
+            if !trimmed.is_empty() {
+                allowlist = allowlist.allow(trimmed);
+            }
+        }
+    }
+    let max_image_bytes: usize = std::env::var("CRAWLER_MAX_IMAGE_BYTES")
+        .ok()
+        .and_then(|s| s.parse().ok())
+        .unwrap_or(mangalord::crawler::safety::DEFAULT_MAX_IMAGE_BYTES);
+    let allowlist = Arc::new(allowlist);
+
    let stats = pipeline::run_metadata_pass(
        manager.as_ref(),
        db,
@@ -239,6 +266,8 @@ async fn run(
        limit,
        skip_chapters,
        mode,
+        allowlist.as_ref(),
+        max_image_bytes,
    )
    .await?;
    tracing::info!(?stats, "metadata pass complete");
@@ -253,6 +282,8 @@ async fn run(
            "target",
            chapter_workers,
            force_refetch_chapters,
+            Arc::clone(&allowlist),
+            max_image_bytes,
        )
        .await?;
    }
@@ -276,6 +307,8 @@ async fn sync_bookmarked_chapter_content(
    source_id: &str,
    workers: usize,
    force_refetch: bool,
+    allowlist: Arc<mangalord::crawler::safety::DownloadAllowlist>,
+    max_image_bytes: usize,
 ) -> anyhow::Result<()> {
    let pending: Vec<(Uuid, Uuid, String)> = sqlx::query_as(
        r#"
@@ -312,6 +345,7 @@ async fn sync_bookmarked_chapter_content(
            let storage = Arc::clone(&storage);
            let rate = Arc::clone(&rate);
            let manager = Arc::clone(&manager);
+            let allowlist = Arc::clone(&allowlist);
            let stats = &stats;
            async move {
                if session_expired.load(std::sync::atomic::Ordering::Relaxed) {
@@ -336,6 +370,8 @@ async fn sync_bookmarked_chapter_content(
                    manga_id,
                    &source_url,
                    force_refetch,
+                    allowlist.as_ref(),
+                    max_image_bytes,
                )
                .await;
                drop(lease);
--- a/backend/src/config.rs
+++ b/backend/src/config.rs
@@ -5,6 +5,7 @@ use chrono::NaiveTime;
 use chrono_tz::Tz;

 use crate::crawler::browser::LaunchOptions;
+use crate::crawler::safety::{DownloadAllowlist, DEFAULT_MAX_IMAGE_BYTES};
 use crate::crawler::source::DiscoverMode;

 /// What `CRAWLER_MODE` was set to. `Auto` is the daemon's default —
@@ -93,6 +94,13 @@ pub struct CrawlerConfig {
    /// `stop_after_unchanged` threshold supplied to Incremental in both
    /// `Auto` (post-seed) and `Explicit(Incremental)` modes.
    pub incremental_stop_after: usize,
+    /// Hosts the crawler is allowed to download images / covers from.
+    /// Always seeded with the host of `start_url` and (when set) the
+    /// configured `cdn_host`. Additional hosts can be added via
+    /// `CRAWLER_DOWNLOAD_ALLOWLIST` (comma-separated).
+    pub download_allowlist: DownloadAllowlist,
+    /// Hard upper bound on a single image download. Defaults to 32 MiB.
+    pub max_image_bytes: usize,
 }

 impl Default for CrawlerConfig {
@@ -115,6 +123,8 @@ impl Default for CrawlerConfig {
            browser: LaunchOptions::headless(),
            mode: CrawlerModePref::Auto,
            incremental_stop_after: 20,
+            download_allowlist: DownloadAllowlist::new(),
+            max_image_bytes: DEFAULT_MAX_IMAGE_BYTES,
        }
    }
 }
@@ -172,6 +182,14 @@ impl CrawlerConfig {
        let incremental_stop_after =
            env_u64("CRAWLER_INCREMENTAL_STOP_AFTER", 20).max(1) as usize;
        let mode = parse_mode_env(incremental_stop_after)?;
+        let start_url = std::env::var("CRAWLER_START_URL")
+            .ok()
+            .filter(|s| !s.trim().is_empty());
+        let cdn_host = std::env::var("CRAWLER_CDN_HOST")
+            .ok()
+            .filter(|s| !s.trim().is_empty());
+        let download_allowlist =
+            build_download_allowlist(start_url.as_deref(), cdn_host.as_deref());
        Ok(Self {
            daemon_enabled: env_bool("CRAWLER_DAEMON", true),
            daily_at,
@@ -179,13 +197,9 @@ impl CrawlerConfig {
            idle_timeout: Duration::from_secs(env_u64("CRAWLER_IDLE_TIMEOUT_S", 600)),
            chapter_workers: env_u64("CRAWLER_CHAPTER_WORKERS", 1).max(1) as usize,
            retention_days: env_u64("CRAWLER_JOB_RETENTION_DAYS", 7) as u32,
-            start_url: std::env::var("CRAWLER_START_URL")
-                .ok()
-                .filter(|s| !s.trim().is_empty()),
+            start_url,
            rate_ms: env_u64("CRAWLER_RATE_MS", 1000),
-            cdn_host: std::env::var("CRAWLER_CDN_HOST")
-                .ok()
-                .filter(|s| !s.trim().is_empty()),
+            cdn_host,
            cdn_rate_ms: env_u64("CRAWLER_CDN_RATE_MS", env_u64("CRAWLER_RATE_MS", 1000)),
            phpsessid: std::env::var("CRAWLER_PHPSESSID")
                .ok()
@@ -202,10 +216,45 @@ impl CrawlerConfig {
            browser: LaunchOptions::from_env(),
            mode,
            incremental_stop_after,
+            download_allowlist,
+            max_image_bytes: env_usize("CRAWLER_MAX_IMAGE_BYTES", DEFAULT_MAX_IMAGE_BYTES),
        })
    }
 }

+/// Build the download allowlist from env. Always includes
+/// `CRAWLER_START_URL`'s host (so the crawler can fetch covers from
+/// the catalog itself) and `CRAWLER_CDN_HOST` when set. Additional
+/// hosts can be supplied via `CRAWLER_DOWNLOAD_ALLOWLIST` (comma-
+/// separated). Empty by default — meaning the crawler refuses to
+/// download anything when no source is configured, which is the safe
+/// fail-closed posture.
+fn build_download_allowlist(
+    start_url: Option<&str>,
+    cdn_host: Option<&str>,
+) -> DownloadAllowlist {
+    let mut allow = DownloadAllowlist::new();
+    if let Some(url) = start_url {
+        if let Ok(parsed) = reqwest::Url::parse(url) {
+            if let Some(h) = parsed.host_str() {
+                allow = allow.allow(h);
+            }
+        }
+    }
+    if let Some(host) = cdn_host {
+        allow = allow.allow(host);
+    }
+    if let Ok(extras) = std::env::var("CRAWLER_DOWNLOAD_ALLOWLIST") {
+        for piece in extras.split(',') {
+            let trimmed = piece.trim();
+            if !trimmed.is_empty() {
+                allow = allow.allow(trimmed);
+            }
+        }
+    }
+    allow
+}
+
 /// Parse `CRAWLER_MODE`. Empty/unset → `Auto`. Recognized values are
 /// `auto`, `backfill`, and `incremental` (case-insensitive). Anything
 /// else is a hard error so a typo can't silently fall through to the
--- a/backend/src/crawler/content.rs
+++ b/backend/src/crawler/content.rs
@@ -18,7 +18,8 @@ use uuid::Uuid;

 use crate::crawler::detect::PageError;
 use crate::crawler::rate_limit::HostRateLimiters;
-use crate::crawler::session;
+use crate::crawler::safety::{fetch_bytes_capped, looks_like_image, DownloadAllowlist};
+use crate::crawler::session::{self, ChapterProbe};
 use crate::storage::Storage;

 /// Parse the chapter page DOM and return the page images in `pageN`
@@ -88,6 +89,8 @@ pub async fn sync_chapter_content(
    manga_id: Uuid,
    source_url: &str,
    force_refetch: bool,
+    allowlist: &DownloadAllowlist,
+    max_image_bytes: usize,
 ) -> anyhow::Result<SyncOutcome> {
    // Skip if already fetched, unless caller explicitly forces.
    if !force_refetch {
@@ -110,16 +113,28 @@ pub async fn sync_chapter_content(
        .with_context(|| format!("open chapter page {source_url}"))?;
    page.wait_for_navigation().await.context("wait for chapter nav")?;

-    // Session probe: avatar present == still logged in. Missing means
-    // PHPSESSID expired; bail the entire crawler run.
-    if page.find_element("#avatar_menu").await.is_err() {
-        page.close().await.ok();
-        return Ok(SyncOutcome::SessionExpired);
-    }
-
    let html = page.content().await.context("read chapter html")?;
    page.close().await.ok();

+    // Three-way session classification: distinguishes a transient
+    // hiccup (broken-page body or logged-in-but-no-reader) from a
+    // genuine PHPSESSID expiry (no reader and no avatar widget). The
+    // earlier binary `#avatar_menu` check conflated both and froze
+    // every worker on a layout shift.
+    match session::classify_chapter_probe(&html) {
+        ChapterProbe::Unauthenticated => return Ok(SyncOutcome::SessionExpired),
+        ChapterProbe::Transient => {
+            // Surface as a typed Err so the dispatcher path runs
+            // ack_failed with exponential backoff (rather than the
+            // session-expired sticky flag).
+            anyhow::bail!(
+                "chapter page at {source_url} returned a transient response \
+                 (broken-page body or reader didn't render); will retry"
+            );
+        }
+        ChapterProbe::Ok => {}
+    }
+
    let images = parse_chapter_pages(&html)
        .with_context(|| format!("parse chapter pages at {source_url}"))?;
    if images.is_empty() {
@@ -138,18 +153,29 @@ pub async fn sync_chapter_content(
            format!("join image URL {} onto {source_url}", img.url)
        })?;
        rate.wait_for(url.as_str()).await?;
-        let resp = http
-            .get(url.clone())
-            // Source CDNs commonly check Referer. Set it to the
-            // chapter page — matches what the browser would send.
-            .header(reqwest::header::REFERER, source_url)
-            .send()
-            .await
-            .with_context(|| format!("GET {url}"))?
-            .error_for_status()
-            .with_context(|| format!("non-2xx for {url}"))?;
-        let bytes = resp.bytes().await.context("read image body")?.to_vec();
-        let ext = infer::get(&bytes).map(|k| k.extension()).unwrap_or("bin");
+        let bytes = fetch_bytes_capped(
+            http,
+            url.as_str(),
+            Some(source_url),
+            allowlist,
+            max_image_bytes,
+        )
+        .await?
+        .to_vec();
+        // Reject any non-image response: the only valid output of an
+        // image URL is an image. `infer` returns None on truncated
+        // bytes too, which also wants to be a failure not a silent
+        // `.bin` extension.
+        if !looks_like_image(&bytes) {
+            anyhow::bail!(
+                "image URL {url} returned non-image bytes \
+                 (first 16: {:?}); refusing to store as binary blob",
+                &bytes.get(..16.min(bytes.len()))
+            );
+        }
+        let ext = infer::get(&bytes)
+            .map(|k| k.extension())
+            .expect("looks_like_image asserted infer succeeded");
        fetched.push((img.page_number, bytes, ext));
    }

@@ -194,8 +220,9 @@ pub async fn sync_chapter_content(
    Ok(SyncOutcome::Fetched { pages: fetched.len() })
 }

-// Suppress unused-import warning for `session` until the bin/crawler
-// wiring lands in this branch and uses it through this module.
+// Suppress unused-import warning for `session::registrable_domain`
+// until the bin/crawler wiring lands in this branch and uses it
+// through this module.
 #[allow(dead_code)]
 fn _keep_session_in_scope() {
    let _ = session::registrable_domain;
--- a/backend/src/crawler/mod.rs
+++ b/backend/src/crawler/mod.rs
@@ -22,5 +22,6 @@ pub mod diff;
 pub mod jobs;
 pub mod pipeline;
 pub mod rate_limit;
+pub mod safety;
 pub mod session;
 pub mod source;
--- a/backend/src/crawler/pipeline.rs
+++ b/backend/src/crawler/pipeline.rs
@@ -9,6 +9,7 @@ use uuid::Uuid;
 use crate::crawler::browser_manager::BrowserManager;
 use crate::crawler::jobs::{self, EnqueueResult, JobPayload};
 use crate::crawler::rate_limit::HostRateLimiters;
+use crate::crawler::safety::{fetch_bytes_capped, looks_like_image, DownloadAllowlist};
 use crate::crawler::source::target::TargetSource;
 use crate::crawler::source::{DiscoverMode, FetchContext, Source};
 use crate::repo;
@@ -62,6 +63,8 @@ pub async fn run_metadata_pass(
    limit: usize,
    skip_chapters: bool,
    mode: DiscoverMode,
+    allowlist: &DownloadAllowlist,
+    max_image_bytes: usize,
 ) -> anyhow::Result<MetadataStats> {
    let lease = browser_manager
        .acquire()
@@ -181,6 +184,8 @@ pub async fn run_metadata_pass(
                        &r.url,
                        upsert.manga_id,
                        cover_url,
+                        allowlist,
+                        max_image_bytes,
                    )
                    .await
                    {
@@ -382,6 +387,7 @@ pub struct EnqueueSummary {
 /// pipeline because the CLI still calls it from its inline chapter-content
 /// loop; once the worker pool fully replaces that path we can fold this
 /// into `pipeline` proper.
+#[allow(clippy::too_many_arguments)]
 async fn download_and_store_cover(
    db: &PgPool,
    storage: &dyn Storage,
@@ -390,6 +396,8 @@ async fn download_and_store_cover(
    manga_url: &str,
    manga_id: Uuid,
    cover_url: &str,
+    allowlist: &DownloadAllowlist,
+    max_image_bytes: usize,
 ) -> anyhow::Result<()> {
    let absolute = reqwest::Url::parse(manga_url)
        .context("parse manga URL")?
@@ -397,17 +405,22 @@ async fn download_and_store_cover(
        .context("join cover URL onto manga URL")?;

    rate.wait_for(absolute.as_str()).await?;
-    let resp = http
-        .get(absolute.clone())
-        .header(reqwest::header::REFERER, manga_url)
-        .send()
-        .await
-        .with_context(|| format!("GET {absolute}"))?
-        .error_for_status()
-        .with_context(|| format!("non-2xx for {absolute}"))?;
-    let bytes = resp.bytes().await.context("read cover body")?;
-    let kind = infer::get(&bytes);
-    let ext = kind.map(|k| k.extension()).unwrap_or("bin");
+    let bytes = fetch_bytes_capped(
+        http,
+        absolute.as_str(),
+        Some(manga_url),
+        allowlist,
+        max_image_bytes,
+    )
+    .await?;
+    if !looks_like_image(&bytes) {
+        anyhow::bail!(
+            "cover URL {absolute} returned non-image bytes; refusing to store as binary blob"
+        );
+    }
+    let ext = infer::get(&bytes)
+        .map(|k| k.extension())
+        .expect("looks_like_image asserted infer succeeded");
    let key = format!("mangas/{manga_id}/cover.{ext}");

    storage
--- a/backend/src/crawler/safety.rs
+++ b/backend/src/crawler/safety.rs
@@ -0,0 +1,486 @@
+//! Defensive helpers for the image-download paths.
+//!
+//! Two threats this module addresses:
+//!
+//! - **SSRF**: a scraped chapter or manga page can embed an absolute
+//!   `<img src="http://10.0.0.1/...">`. The crawler runs inside the
+//!   backend container with intra-compose access to `postgres:5432`
+//!   and possibly other internal services; without a host check the
+//!   crawler would happily probe them. [`is_safe_url`] rejects
+//!   anything whose host isn't on the operator-configured allowlist,
+//!   plus any IP literal in RFC1918 / loopback / link-local / unique-
+//!   local space (including IPv4-mapped IPv6 like `::ffff:127.0.0.1`)
+//!   as a second defence for the case where an allowlisted hostname's
+//!   DNS happens to resolve to a literal private address.
+//!
+//!   **DNS rebinding is not covered.** A hostname like `cdn.allowed.com`
+//!   that *resolves* to `127.0.0.1` via hostile DNS bypasses the IP
+//!   check entirely — `is_safe_url` only inspects URL strings, not
+//!   resolved IPs. Mitigating that requires a custom reqwest resolver
+//!   that filters IPs after DNS, which would mean rebuilding reqwest's
+//!   connector. The allowlist + good operator DNS hygiene is the
+//!   realistic mitigation today.
+//!
+//! - **Unbounded download**: `Response::bytes().await` reads the full
+//!   body before returning. A malicious source serving a 10 GiB image
+//!   would fill memory and then disk. [`accumulate_capped`] streams
+//!   the body chunk-by-chunk into a [`bytes::BytesMut`] and bails as
+//!   soon as the running total exceeds the cap.
+//!
+//! Both helpers are pure-data: the SSRF check is keyed off a parsed
+//! URL string, and the byte accumulator is keyed off a generic stream.
+//! Easy to unit-test without a live network or browser.
+
+use std::net::IpAddr;
+
+use anyhow::{bail, Context};
+use bytes::BytesMut;
+use futures_util::StreamExt;
+use reqwest::Url;
+
+/// Default per-image download cap. A page image is generally <2 MiB;
+/// 32 MiB leaves headroom for high-resolution covers while still
+/// stopping a misbehaving CDN dead. Override via `CRAWLER_MAX_IMAGE_BYTES`.
+pub const DEFAULT_MAX_IMAGE_BYTES: usize = 32 * 1024 * 1024;
+
+/// Hosts that are always allowed in addition to the operator's
+/// configured allowlist. None by default — keeping the surface area
+/// minimal so the only way a URL gets through is if it matches an
+/// explicit catalog/CDN entry.
+#[derive(Clone, Debug, Default)]
+pub struct DownloadAllowlist {
+    hosts: Vec<String>,
+}
+
+impl DownloadAllowlist {
+    pub fn new() -> Self {
+        Self { hosts: Vec::new() }
+    }
+
+    /// Add a host (case-insensitive match). Sub-domains are *not*
+    /// implied: pass `cdn.example.com` and `example.com` separately
+    /// if both should be reachable.
+    pub fn allow(mut self, host: impl Into<String>) -> Self {
+        let h = host.into().to_ascii_lowercase();
+        if !h.is_empty() && !self.hosts.iter().any(|existing| existing == &h) {
+            self.hosts.push(h);
+        }
+        self
+    }
+
+    pub fn is_empty(&self) -> bool {
+        self.hosts.is_empty()
+    }
+
+    pub fn contains(&self, host: &str) -> bool {
+        let lower = host.to_ascii_lowercase();
+        self.hosts.iter().any(|h| h == &lower)
+    }
+}
+
+/// Verify a URL is safe for the crawler to fetch.
+///
+/// Rejects:
+/// - non-http(s) schemes (file://, gopher://, …),
+/// - any IP literal in private / loopback / link-local / unique-local
+///   space (defense in depth — a DNS allowlist alone wouldn't cover an
+///   attacker that places an entry like `cdn.evil` pointing at
+///   `192.168.1.1`),
+/// - the literal hostname `localhost`,
+/// - hosts that aren't on the supplied allowlist.
+///
+/// An empty allowlist rejects everything (the conservative default —
+/// callers must explicitly allow the catalog and CDN hosts).
+pub fn is_safe_url(raw_url: &str, allow: &DownloadAllowlist) -> Result<(), UrlSafetyError> {
+    let url = Url::parse(raw_url).map_err(|_| UrlSafetyError::Unparseable)?;
+    let scheme = url.scheme();
+    if scheme != "http" && scheme != "https" {
+        return Err(UrlSafetyError::BadScheme(scheme.to_string()));
+    }
+    let host = url.host_str().ok_or(UrlSafetyError::NoHost)?;
+    let lower_host = host.to_ascii_lowercase();
+    if lower_host == "localhost" {
+        return Err(UrlSafetyError::Loopback);
+    }
+    // Reject IP literals in private/loopback ranges regardless of the
+    // allowlist — if someone puts an IP literal on the allowlist they
+    // almost certainly didn't mean a private range.
+    // reqwest::Url normalises IPv6 literals as `[::1]` (brackets
+    // included) in `host_str()`. Strip the brackets before parsing.
+    let ip_candidate = lower_host
+        .strip_prefix('[')
+        .and_then(|s| s.strip_suffix(']'))
+        .unwrap_or(&lower_host);
+    if let Ok(ip) = ip_candidate.parse::<IpAddr>() {
+        if is_private_ip(&ip) {
+            return Err(UrlSafetyError::PrivateIp(ip));
+        }
+    }
+    if !allow.contains(&lower_host) {
+        return Err(UrlSafetyError::HostNotAllowed(lower_host));
+    }
+    Ok(())
+}
+
+fn is_private_ip(ip: &IpAddr) -> bool {
+    match ip {
+        IpAddr::V4(v4) => {
+            v4.is_loopback()
+                || v4.is_private()
+                || v4.is_link_local()
+                || v4.is_unspecified()
+                || v4.is_broadcast()
+                // CGNAT 100.64.0.0/10
+                || (v4.octets()[0] == 100 && (v4.octets()[1] & 0xC0) == 64)
+                // 169.254/16 link-local already covered, but 0.0.0.0/8 is special-use
+                || v4.octets()[0] == 0
+        }
+        IpAddr::V6(v6) => {
+            // IPv4-mapped IPv6 (::ffff:0:0/96): unwrap to the embedded
+            // IPv4 and recurse so `::ffff:127.0.0.1` is caught by the
+            // IPv4 loopback check rather than passing through.
+            // `Ipv6Addr::is_loopback()` only matches `::1` exactly.
+            if let Some(v4) = v6.to_ipv4_mapped() {
+                return is_private_ip(&IpAddr::V4(v4));
+            }
+            v6.is_loopback()
+                || v6.is_unspecified()
+                // fc00::/7 unique-local
+                || (v6.segments()[0] & 0xfe00) == 0xfc00
+                // fe80::/10 link-local
+                || (v6.segments()[0] & 0xffc0) == 0xfe80
+        }
+    }
+}
+
+#[derive(Debug, thiserror::Error, PartialEq, Eq)]
+pub enum UrlSafetyError {
+    #[error("URL is not parseable")]
+    Unparseable,
+    #[error("scheme {0:?} is not http or https")]
+    BadScheme(String),
+    #[error("URL is missing a host")]
+    NoHost,
+    #[error("host points at the loopback interface")]
+    Loopback,
+    #[error("host is a private/internal IP: {0}")]
+    PrivateIp(IpAddr),
+    #[error("host {0:?} is not on the crawler download allowlist")]
+    HostNotAllowed(String),
+}
+
+/// Drain a byte stream into a single buffer, bailing out as soon as
+/// the running total exceeds `max_bytes`. Generic over the stream so
+/// it's testable without a live HTTP response.
+pub async fn accumulate_capped<S, E>(stream: S, max_bytes: usize) -> anyhow::Result<bytes::Bytes>
+where
+    S: futures_core::Stream<Item = Result<bytes::Bytes, E>>,
+    E: std::error::Error + Send + Sync + 'static,
+{
+    let mut buf = BytesMut::new();
+    let mut stream = std::pin::pin!(stream);
+    while let Some(chunk) = stream.next().await {
+        let chunk = chunk.map_err(|e| anyhow::anyhow!("stream chunk: {e}"))?;
+        if buf.len().saturating_add(chunk.len()) > max_bytes {
+            bail!(
+                "response exceeds {max_bytes}-byte cap (received >{}+{})",
+                buf.len(),
+                chunk.len()
+            );
+        }
+        buf.extend_from_slice(&chunk);
+    }
+    Ok(buf.freeze())
+}
+
+/// Send `req` and stream the response into a length-limited buffer.
+/// Combines [`is_safe_url`] check + [`accumulate_capped`] so each
+/// call-site is one line.
+pub async fn fetch_bytes_capped(
+    http: &reqwest::Client,
+    url: &str,
+    referer: Option<&str>,
+    allow: &DownloadAllowlist,
+    max_bytes: usize,
+) -> anyhow::Result<bytes::Bytes> {
+    is_safe_url(url, allow).with_context(|| format!("reject unsafe URL {url}"))?;
+    let mut req = http.get(url);
+    if let Some(r) = referer {
+        req = req.header(reqwest::header::REFERER, r);
+    }
+    let resp = req
+        .send()
+        .await
+        .with_context(|| format!("GET {url}"))?
+        .error_for_status()
+        .with_context(|| format!("non-2xx for {url}"))?;
+    accumulate_capped(resp.bytes_stream(), max_bytes)
+        .await
+        .with_context(|| format!("download body for {url}"))
+}
+
+/// True when `bytes` sniffs as one of the *renderable* image formats
+/// the `/files/*key` endpoint can serve with a correct Content-Type:
+/// JPEG, PNG, WebP, GIF, AVIF. Matches the upload pipeline's
+/// whitelist in `upload::parse_image`.
+///
+/// `infer::MatcherType::Image` is intentionally NOT used — it also
+/// matches BMP, TIFF, HEIF, ICO, PSD, and JP2. Those would sniff as
+/// "image" here but [`api::files::content_type_for`] would fall back
+/// to `application/octet-stream`, prompting browsers to download
+/// instead of render. Keep the two layers aligned.
+pub fn looks_like_image(bytes: &[u8]) -> bool {
+    matches!(
+        infer::get(bytes).map(|k| k.mime_type()),
+        Some("image/jpeg" | "image/png" | "image/webp" | "image/gif" | "image/avif")
+    )
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use futures_util::stream;
+
+    fn allow_just(host: &str) -> DownloadAllowlist {
+        DownloadAllowlist::new().allow(host)
+    }
+
+    #[test]
+    fn safe_url_allows_listed_host() {
+        let allow = allow_just("cdn.example.com");
+        assert!(is_safe_url("https://cdn.example.com/img.jpg", &allow).is_ok());
+    }
+
+    #[test]
+    fn safe_url_blocks_unlisted_host() {
+        let allow = allow_just("cdn.example.com");
+        let err = is_safe_url("https://evil.example.org/img.jpg", &allow).unwrap_err();
+        assert!(matches!(err, UrlSafetyError::HostNotAllowed(h) if h == "evil.example.org"));
+    }
+
+    #[test]
+    fn safe_url_blocks_localhost_even_if_allowlisted() {
+        let allow = allow_just("localhost");
+        assert!(matches!(
+            is_safe_url("http://localhost:8080/", &allow).unwrap_err(),
+            UrlSafetyError::Loopback
+        ));
+    }
+
+    #[test]
+    fn safe_url_blocks_loopback_ipv4() {
+        let allow = allow_just("127.0.0.1");
+        assert!(matches!(
+            is_safe_url("http://127.0.0.1/", &allow).unwrap_err(),
+            UrlSafetyError::PrivateIp(_)
+        ));
+    }
+
+    #[test]
+    fn safe_url_blocks_rfc1918() {
+        let allow = allow_just("10.0.0.1");
+        for url in [
+            "http://10.0.0.1/",
+            "http://192.168.1.1/",
+            "http://172.16.0.5/",
+            "http://172.31.255.255/",
+        ] {
+            assert!(
+                matches!(
+                    is_safe_url(url, &allow).unwrap_err(),
+                    UrlSafetyError::PrivateIp(_)
+                ),
+                "should reject {url}"
+            );
+        }
+    }
+
+    #[test]
+    fn safe_url_blocks_link_local() {
+        let allow = allow_just("169.254.169.254");
+        // 169.254.169.254 is the AWS/GCP metadata service — the most
+        // dangerous SSRF target on a default cloud VM.
+        assert!(matches!(
+            is_safe_url("http://169.254.169.254/", &allow).unwrap_err(),
+            UrlSafetyError::PrivateIp(_)
+        ));
+    }
+
+    #[test]
+    fn safe_url_blocks_ipv6_loopback_and_ula() {
+        // Debug what host_str returns first — reqwest::Url normalises
+        // IPv6 literals as `[::1]` with brackets, which doesn't parse
+        // as `IpAddr` directly. The implementation strips them.
+        let allow = allow_just("[::1]");
+        let err = is_safe_url("http://[::1]/", &allow).unwrap_err();
+        assert!(
+            matches!(err, UrlSafetyError::PrivateIp(_)),
+            "expected PrivateIp, got {err:?}"
+        );
+        let allow = allow_just("[fd00::1]");
+        let err = is_safe_url("http://[fd00::1]/", &allow).unwrap_err();
+        assert!(
+            matches!(err, UrlSafetyError::PrivateIp(_)),
+            "expected PrivateIp, got {err:?}"
+        );
+    }
+
+    #[test]
+    fn safe_url_blocks_ipv4_mapped_ipv6_loopback() {
+        // `Ipv6Addr::is_loopback()` only matches `::1` exactly, so
+        // `::ffff:127.0.0.1` would slip through without the
+        // to_ipv4_mapped() unwrap in is_private_ip.
+        let allow = allow_just("[::ffff:127.0.0.1]");
+        let err = is_safe_url("http://[::ffff:127.0.0.1]/", &allow).unwrap_err();
+        assert!(
+            matches!(err, UrlSafetyError::PrivateIp(_)),
+            "expected PrivateIp, got {err:?}"
+        );
+    }
+
+    #[test]
+    fn safe_url_blocks_ipv4_mapped_ipv6_rfc1918() {
+        let allow = allow_just("[::ffff:10.0.0.1]");
+        let err = is_safe_url("http://[::ffff:10.0.0.1]/", &allow).unwrap_err();
+        assert!(matches!(err, UrlSafetyError::PrivateIp(_)));
+    }
+
+    #[test]
+    fn safe_url_blocks_non_http_schemes() {
+        let allow = allow_just("anywhere");
+        assert!(matches!(
+            is_safe_url("file:///etc/passwd", &allow).unwrap_err(),
+            UrlSafetyError::BadScheme(_)
+        ));
+        assert!(matches!(
+            is_safe_url("gopher://anywhere:70/", &allow).unwrap_err(),
+            UrlSafetyError::BadScheme(_)
+        ));
+    }
+
+    #[test]
+    fn safe_url_rejects_unparseable() {
+        let allow = allow_just("anywhere");
+        assert!(matches!(
+            is_safe_url("not a url", &allow).unwrap_err(),
+            UrlSafetyError::Unparseable
+        ));
+    }
+
+    #[test]
+    fn safe_url_empty_allowlist_rejects_everything() {
+        let allow = DownloadAllowlist::new();
+        let err = is_safe_url("https://cdn.example.com/img.jpg", &allow).unwrap_err();
+        assert!(matches!(err, UrlSafetyError::HostNotAllowed(_)));
+    }
+
+    #[test]
+    fn allowlist_matches_case_insensitively() {
+        let allow = DownloadAllowlist::new().allow("CDN.Example.COM");
+        assert!(is_safe_url("https://cdn.example.com/x.jpg", &allow).is_ok());
+        assert!(is_safe_url("https://CDN.EXAMPLE.com/x.jpg", &allow).is_ok());
+    }
+
+    #[tokio::test]
+    async fn accumulate_capped_returns_full_body_under_cap() {
+        let chunks: Vec<Result<bytes::Bytes, std::io::Error>> = vec![
+            Ok(bytes::Bytes::from_static(b"hello ")),
+            Ok(bytes::Bytes::from_static(b"world")),
+        ];
+        let s = stream::iter(chunks);
+        let out = accumulate_capped(s, 100).await.unwrap();
+        assert_eq!(out.as_ref(), b"hello world");
+    }
+
+    #[tokio::test]
+    async fn accumulate_capped_bails_past_cap() {
+        let chunks: Vec<Result<bytes::Bytes, std::io::Error>> = vec![
+            Ok(bytes::Bytes::from(vec![0u8; 50])),
+            Ok(bytes::Bytes::from(vec![0u8; 60])),
+        ];
+        let s = stream::iter(chunks);
+        let err = accumulate_capped(s, 100).await.unwrap_err();
+        assert!(err.to_string().contains("100-byte cap"));
+    }
+
+    #[tokio::test]
+    async fn accumulate_capped_surfaces_stream_errors() {
+        let chunks: Vec<Result<bytes::Bytes, std::io::Error>> = vec![
+            Ok(bytes::Bytes::from_static(b"ok")),
+            Err(std::io::Error::other("network blip")),
+        ];
+        let s = stream::iter(chunks);
+        let err = accumulate_capped(s, 100).await.unwrap_err();
+        assert!(err.to_string().contains("network blip"));
+    }
+
+    #[test]
+    fn looks_like_image_accepts_jpeg() {
+        // JPEG SOI + APP0 segment.
+        let jpeg = [0xff, 0xd8, 0xff, 0xe0, 0, 0x10, b'J', b'F', b'I', b'F'];
+        assert!(looks_like_image(&jpeg));
+    }
+
+    #[test]
+    fn looks_like_image_accepts_png() {
+        let png = [0x89, 0x50, 0x4e, 0x47, 0x0d, 0x0a, 0x1a, 0x0a, 0, 0, 0, 0];
+        assert!(looks_like_image(&png));
+    }
+
+    #[test]
+    fn looks_like_image_rejects_html_disguised_as_image() {
+        let html = b"<html><body>not an image</body></html>";
+        assert!(!looks_like_image(html));
+    }
+
+    #[test]
+    fn looks_like_image_rejects_empty() {
+        assert!(!looks_like_image(&[]));
+    }
+
+    #[test]
+    fn looks_like_image_rejects_renderable_but_unsupported_formats() {
+        // BMP, TIFF, ICO, PSD are `infer::MatcherType::Image` but the
+        // /files/*key handler doesn't have Content-Type mappings for
+        // them, so they'd be served as application/octet-stream and
+        // download instead of render. Reject at the crawler so we
+        // never land them in storage.
+        // BMP magic: "BM" + 4-byte size.
+        let bmp = [b'B', b'M', 0, 0, 0, 0, 0, 0, 0, 0, 0, 0];
+        assert!(!looks_like_image(&bmp), "BMP must be rejected (not renderable by /files)");
+
+        // TIFF little-endian magic: "II" + 42.
+        let tiff = [0x49, 0x49, 0x2a, 0x00, 0, 0, 0, 0];
+        assert!(!looks_like_image(&tiff), "TIFF must be rejected");
+
+        // ICO magic: 0x00,0x00,0x01,0x00.
+        let ico = [0x00, 0x00, 0x01, 0x00, 1, 0, 16, 16, 0, 0, 1, 0, 0x18, 0, 0x40, 0, 0, 0, 0x16, 0, 0, 0];
+        assert!(!looks_like_image(&ico), "ICO must be rejected");
+    }
+
+    #[test]
+    fn looks_like_image_accepts_webp_gif_avif() {
+        // Cover the three remaining whitelisted formats so a future
+        // tightening that drops one would fail noisily.
+        let webp = [
+            b'R', b'I', b'F', b'F',
+            0, 0, 0, 0,
+            b'W', b'E', b'B', b'P',
+            b'V', b'P', b'8', b' ',
+        ];
+        assert!(looks_like_image(&webp));
+
+        let gif = [b'G', b'I', b'F', b'8', b'7', b'a', 0, 0, 0, 0];
+        assert!(looks_like_image(&gif));
+
+        let avif = [
+            0x00, 0x00, 0x00, 0x18,
+            b'f', b't', b'y', b'p',
+            b'a', b'v', b'i', b'f',
+            0x00, 0x00, 0x00, 0x00,
+            b'm', b'i', b'f', b'1',
+            b'a', b'v', b'i', b'f',
+        ];
+        assert!(looks_like_image(&avif));
+    }
+}
--- a/backend/src/crawler/session.rs
+++ b/backend/src/crawler/session.rs
@@ -127,6 +127,54 @@ pub fn classify_probe(html: &str) -> SessionProbe {
    }
 }

+/// Three-way classification of a chapter page response.
+///
+/// Reader pages don't render `#logo`, so [`classify_probe`] can't be
+/// reused as-is. The chapter-specific marker is `a#pic_container`
+/// (asserted by the reader-page parser at `parse_chapter_pages`).
+///
+/// Order matters: broken-page body wins over selector matches, so a
+/// transient site-wide 5xx that happens to render the avatar widget
+/// elsewhere doesn't falsely reach `Ok`.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum ChapterProbe {
+    /// `a#pic_container` present — reader rendered. Whether
+    /// `#avatar_menu` is also there is informational; if the reader
+    /// loaded the session is by definition still good.
+    Ok,
+    /// Site rendered a "logged out" or "please log in" page (no
+    /// reader, no broken-page body, and no avatar widget either).
+    /// Distinguishes the genuine expired-session case from a
+    /// transient site hiccup.
+    Unauthenticated,
+    /// Broken-page body, or reader didn't render but the user is
+    /// still logged in (avatar widget present). Caller should retry
+    /// rather than blame the session.
+    Transient,
+}
+
+pub fn classify_chapter_probe(html: &str) -> ChapterProbe {
+    if is_broken_page_body(html) {
+        return ChapterProbe::Transient;
+    }
+    let doc = scraper::Html::parse_document(html);
+    let container = scraper::Selector::parse("a#pic_container").unwrap();
+    if doc.select(&container).next().is_some() {
+        return ChapterProbe::Ok;
+    }
+    let avatar = scraper::Selector::parse("#avatar_menu").unwrap();
+    if doc.select(&avatar).next().is_some() {
+        // Logged-in user, but the reader didn't render — most likely
+        // the layout shifted or the site is serving an interstitial.
+        ChapterProbe::Transient
+    } else {
+        // No reader, no avatar, no broken-body marker — site rendered
+        // the "please log in" page, which is the genuine session-
+        // expired signal on this route.
+        ChapterProbe::Unauthenticated
+    }
+}
+
 /// In-startup retry budget for the session probe. Small but non-zero —
 /// startup hitting a 5-second site hiccup shouldn't fail the operator
 /// with "PHPSESSID expired" when the session is actually fine.
@@ -273,6 +321,73 @@ mod tests {
        assert_eq!(classify_probe(""), SessionProbe::Transient);
    }

+    #[test]
+    fn classify_chapter_probe_ok_when_reader_rendered() {
+        let html = r#"
+            <html><body>
+              <a id="pic_container">
+                <img id="page1" src="https://cdn/1.jpg">
+              </a>
+            </body></html>
+        "#;
+        assert_eq!(classify_chapter_probe(html), ChapterProbe::Ok);
+    }
+
+    #[test]
+    fn classify_chapter_probe_unauthenticated_when_no_reader_and_no_avatar() {
+        // What a logged-out hit on a chapter URL renders: a normal
+        // site layout (header etc.) with a "please log in" body, but
+        // no reader and no avatar widget.
+        let html = r#"
+            <html><body>
+              <header><div id="logo">Catalog</div></header>
+              <main>Please log in to read this chapter.</main>
+            </body></html>
+        "#;
+        assert_eq!(
+            classify_chapter_probe(html),
+            ChapterProbe::Unauthenticated
+        );
+    }
+
+    #[test]
+    fn classify_chapter_probe_transient_when_logged_in_but_reader_missing() {
+        // Avatar shows the session is still valid; reader didn't
+        // render — site is serving an interstitial or the layout
+        // momentarily shifted. Retry, don't blame the session.
+        let html = r#"
+            <html><body>
+              <header><div id="logo">Catalog</div><div id="avatar_menu"></div></header>
+              <main>Site maintenance — back in 5 minutes.</main>
+            </body></html>
+        "#;
+        assert_eq!(classify_chapter_probe(html), ChapterProbe::Transient);
+    }
+
+    #[test]
+    fn classify_chapter_probe_transient_on_broken_page_body() {
+        let html =
+            "<html><body><p>we're sorry, the request file are not found.</p></body></html>";
+        assert_eq!(classify_chapter_probe(html), ChapterProbe::Transient);
+    }
+
+    #[test]
+    fn classify_chapter_probe_does_not_misfire_on_avatar_alone_without_reader() {
+        // Regression for the original bug: the binary
+        // find_element("#avatar_menu") check treated "no avatar" as
+        // session-expired even when a transient hiccup was the real
+        // cause. classify_chapter_probe must NOT trip on that pattern
+        // when pic_container *is* present.
+        let html = r#"
+            <html><body>
+              <a id="pic_container">
+                <img id="page1" src="https://cdn/1.jpg">
+              </a>
+            </body></html>
+        "#;
+        assert_eq!(classify_chapter_probe(html), ChapterProbe::Ok);
+    }
+
    #[test]
    fn classify_probe_trusts_broken_body_over_stray_avatar_match() {
        // Defensive: if a broken-page body somehow contains an
--- a/frontend/package.json
+++ b/frontend/package.json
@@ -1,6 +1,6 @@
 {
  "name": "mangalord-frontend",
-  "version": "0.34.0",
+  "version": "0.34.1",
  "private": true,
  "type": "module",
  "scripts": {
--- a/frontend/src/hooks.server.test.ts
+++ b/frontend/src/hooks.server.test.ts
@@ -118,77 +118,4 @@ describe('hooks.server proxy', () => {
        expect(body.error.code).toBe('upstream_unavailable');
        expect(errSpy).toHaveBeenCalled();
    });
-
-    it('strips every hop-by-hop header listed in RFC 7230 §6.1', async () => {
-        // Defence in depth: axum doesn't emit these, but a future
-        // middleware that did would otherwise leak per-connection
-        // state across the proxy boundary.
-        fetchSpy.mockResolvedValueOnce(new Response('[]', { status: 200 }));
-        const resolve = vi.fn();
-        await handle({
-            event: makeEvent('/api/v1/health', {
-                headers: {
-                    host: 'app.example.com',
-                    'content-length': '0',
-                    connection: 'keep-alive',
-                    'keep-alive': 'timeout=5',
-                    'proxy-authenticate': 'Basic realm=x',
-                    'proxy-authorization': 'Basic xyz',
-                    te: 'trailers',
-                    trailer: 'Expires',
-                    'transfer-encoding': 'chunked',
-                    upgrade: 'websocket',
-                    // A non-hop-by-hop header to ensure non-targets
-                    // aren't accidentally stripped.
-                    'x-custom': 'pass-through'
-                }
-            }),
-            resolve
-        });
-        const init = fetchSpy.mock.calls[0][1] as RequestInit;
-        const headers = init.headers as Headers;
-        for (const h of [
-            'host',
-            'content-length',
-            'connection',
-            'keep-alive',
-            'proxy-authenticate',
-            'proxy-authorization',
-            'te',
-            'trailer',
-            'transfer-encoding',
-            'upgrade'
-        ]) {
-            expect(headers.get(h), `${h} should be stripped`).toBeNull();
-        }
-        expect(headers.get('x-custom')).toBe('pass-through');
-    });
-
-    it('aborts and returns 502 when the upstream stalls past the timeout', async () => {
-        const errSpy = vi.spyOn(console, 'error').mockImplementation(() => {});
-        // Simulate an aborted fetch (AbortController.abort() raises a
-        // DOMException with name 'AbortError' on Node's fetch). The
-        // handler should treat it as the same upstream_unavailable
-        // 502 it uses for any other network failure.
-        const abortErr = new DOMException('aborted', 'AbortError');
-        fetchSpy.mockRejectedValueOnce(abortErr);
-
-        const resolve = vi.fn();
-        const resp = await handle({ event: makeEvent('/api/v1/slow'), resolve });
-        expect(resp.status).toBe(502);
-        const body = await resp.json();
-        expect(body.error.code).toBe('upstream_unavailable');
-        expect(errSpy).toHaveBeenCalled();
-    });
-
-    it('attaches an AbortSignal to the upstream fetch so it can time out', async () => {
-        fetchSpy.mockResolvedValueOnce(new Response('[]', { status: 200 }));
-        const resolve = vi.fn();
-        await handle({ event: makeEvent('/api/v1/health'), resolve });
-        const init = fetchSpy.mock.calls[0][1] as RequestInit;
-        expect(init.signal).toBeInstanceOf(AbortSignal);
-        // The signal hasn't fired (handler returned in time), but its
-        // presence is the contract this test is pinning.
-        expect(init.signal?.aborted).toBe(false);
-    });
 });
--- a/frontend/src/hooks.server.ts
+++ b/frontend/src/hooks.server.ts
@@ -12,66 +12,20 @@ import type { Handle } from '@sveltejs/kit';

 const BACKEND_URL = process.env.BACKEND_URL ?? 'http://localhost:8080';

-/**
- * Hop-by-hop headers per RFC 7230 §6.1. These are scoped to a single
- * transport-level connection and must not be forwarded by a proxy.
- * Plus `host` and `content-length`: `host` would mislead the backend
- * about its origin, and `content-length` is recomputed by the upstream
- * fetch from the body stream.
- */
-const HOP_BY_HOP_HEADERS = [
-    'host',
-    'content-length',
-    'connection',
-    'keep-alive',
-    'proxy-authenticate',
-    'proxy-authorization',
-    'te',
-    'trailer',
-    'transfer-encoding',
-    'upgrade'
-];
-
-/**
- * Cap each proxied request at 5 minutes. The bound exists to surface
- * a wedged backend (stuck on a slow DB query, deadlocked, etc.) as a
- * 502 rather than letting the browser request hang indefinitely.
- *
- * The default leans toward the slow-upload end of the spectrum: at a
- * 1 Mbps upstream, a 200 MiB chapter upload (the default
- * `MAX_REQUEST_BYTES` cap) needs ~27 minutes; 300 s covers the more
- * realistic 25 Mbps urban-broadband case (~64 s for the same upload)
- * with comfortable headroom. Operators serving very slow clients
- * should raise `BACKEND_PROXY_TIMEOUT_MS`; operators behind a
- * tighter upstream proxy may want to lower it. A future improvement
- * is an idle-based timeout (reset per chunk) instead of this
- * wall-clock budget — that's a fair bit more code, deferred.
- */
-const PROXY_TIMEOUT_MS = (() => {
-    const raw = process.env.BACKEND_PROXY_TIMEOUT_MS;
-    const n = raw ? Number(raw) : 300_000;
-    return Number.isFinite(n) && n > 0 ? n : 300_000;
-})();
-
 export const handle: Handle = async ({ event, resolve }) => {
    if (event.url.pathname.startsWith('/api/')) {
        const target = `${BACKEND_URL}${event.url.pathname}${event.url.search}`;

+        // Strip hop-by-hop headers — `host` would mislead the backend
+        // about the origin, and `content-length` will be recomputed.
        const headers = new Headers(event.request.headers);
-        for (const h of HOP_BY_HOP_HEADERS) headers.delete(h);
-
-        // AbortController times the upstream fetch out so a backend
-        // wedged on a slow DB query doesn't keep the browser request
-        // hanging forever. The `signal` is also wired into the
-        // RequestInit so the body stream is cancelled cleanly.
-        const ctrl = new AbortController();
-        const timeoutHandle = setTimeout(() => ctrl.abort(), PROXY_TIMEOUT_MS);
+        headers.delete('host');
+        headers.delete('content-length');

        const init: RequestInit & { duplex?: 'half' } = {
            method: event.request.method,
            headers,
-            redirect: 'manual',
-            signal: ctrl.signal
+            redirect: 'manual'
        };
        if (event.request.method !== 'GET' && event.request.method !== 'HEAD') {
            init.body = event.request.body;
@@ -85,13 +39,11 @@ export const handle: Handle = async ({ event, resolve }) => {
            upstream = await fetch(target, init);
        } catch (e) {
            // Network-layer failure (DNS / connection refused / TLS
-            // handshake / abort by timeout) — most commonly "backend
-            // container restarting". SvelteKit's default 500 would be
-            // an HTML page that client.ts can't .json(), which masks
-            // the real cause. Emit the standard envelope with a
-            // dedicated code instead.
+            // handshake) — most commonly "backend container restarting".
+            // SvelteKit's default 500 would be an HTML page that
+            // client.ts can't .json(), which masks the real cause. Emit
+            // the standard envelope with a dedicated code instead.
            console.error('Proxy to backend failed:', e);
-            clearTimeout(timeoutHandle);
            return new Response(
                JSON.stringify({
                    error: {
@@ -106,7 +58,6 @@ export const handle: Handle = async ({ event, resolve }) => {
            );
        }

-        clearTimeout(timeoutHandle);
        return new Response(upstream.body, {
            status: upstream.status,
            statusText: upstream.statusText,