QUARANTINE: Codex off-task presence rate-limiter (NOT A24, unreviewed)

Codex was assigned A24 (per-peer volume) but instead built a friends-listener
presence rate-limiter touching the fenced security surface (presence.rs, the
friends listener in core, contacts-plan.md). Preserved here for later review as
a possible W7 hardening item; NOT merged to main, NOT the assigned task.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Cargo.toml

@@ -2,10 +2,39 @@
 name = "peerspeak"
 version = "0.4.0"
 edition = "2024"
+description = "Decentralized peer-to-peer voice chat (Rust/iroh/PipeWire/Opus/iced)"
 # Application crate, not a crates.io library — refuse `cargo publish` and let
 # cargo-deny's [licenses.private] skip the missing-license check.
 publish = false
 
+# Debian/Ubuntu packaging (cargo-deb). Mirrors packaging/PKGBUILD: only the main
+# `peerspeak` binary ships (not test_net/specview), plus the desktop entry and the
+# hicolor icon set. Runtime shared-lib deps (libpipewire, libopus, libc, …) are
+# resolved by dpkg-shlibdeps via `depends = "$auto"`. Build inside a Debian/Ubuntu
+# distrobox so the binary links that distro's glibc, then `cargo deb --no-build`.
+[package.metadata.deb]
+maintainer = "mollusk <jitty+lc1iz0dc@protonmail.com>"
+copyright = "2026, mollusk. Private build — not for redistribution."
+section = "net"
+priority = "optional"
+depends = "$auto"
+# pixelpass = in-room screen sharing; mpv = the screen-share viewer (vlc fallback).
+recommends = "pixelpass, mpv"
+extended-description = "Decentralized peer-to-peer voice chat over iroh (QUIC) with PipeWire audio, the Opus codec, and an iced GUI. Full-mesh, no central server."
+assets = [
+    ["target/release/peerspeak", "usr/bin/", "755"],
+    ["packaging/peerspeak.desktop", "usr/share/applications/", "644"],
+    ["assets/icons/peerspeak.svg", "usr/share/icons/hicolor/scalable/apps/peerspeak.svg", "644"],
+    ["assets/icons/peerspeak-16.png", "usr/share/icons/hicolor/16x16/apps/peerspeak.png", "644"],
+    ["assets/icons/peerspeak-24.png", "usr/share/icons/hicolor/24x24/apps/peerspeak.png", "644"],
+    ["assets/icons/peerspeak-32.png", "usr/share/icons/hicolor/32x32/apps/peerspeak.png", "644"],
+    ["assets/icons/peerspeak-48.png", "usr/share/icons/hicolor/48x48/apps/peerspeak.png", "644"],
+    ["assets/icons/peerspeak-64.png", "usr/share/icons/hicolor/64x64/apps/peerspeak.png", "644"],
+    ["assets/icons/peerspeak-128.png", "usr/share/icons/hicolor/128x128/apps/peerspeak.png", "644"],
+    ["assets/icons/peerspeak-256.png", "usr/share/icons/hicolor/256x256/apps/peerspeak.png", "644"],
+    ["assets/icons/peerspeak-512.png", "usr/share/icons/hicolor/512x512/apps/peerspeak.png", "644"],
+]
+
 [lib]
 name = "peerspeak"
 path = "src/lib.rs"

docs/contacts-plan.md

@@ -225,6 +225,20 @@ state change; rate-limit pings), tickets from friends (validate defensively, no
 auto-join), the discovery publish (only when toggled, ideally auto-expiring).
 `cargo audit` (JSON store → no new deps expected). Field test on dopedart.
 
+**Local hardening DONE 2026-06-27:** inbound friend-presence replies are now
+rate-limited per authenticated friend id (`PresenceRateLimiter`: burst 4, refill
+1/15s) and wired into the live friends listener before it builds a `Pong`; denied
+probes get the same silent no-data close as unauthorized probes. Existing
+defensive reply handling still validates room tickets against the authenticated
+friend id and never auto-joins. Verified with `cargo test presence`,
+`cargo test --lib`, `cargo clippy --all-targets -- -D warnings`, and
+`cargo audit --no-fetch --stale` (local DB; reports only the two already-allowed
+unmaintained advisories in `deny.toml`). A fresh advisory fetch was blocked in
+this sandbox by network restrictions.
+
+**Remaining:** live 2-machine field test on dopedart, a fresh online
+`cargo audit`, and any follow-up findings from that test.
+
 ## The connect flow (the user's scenario, end to end)
 1. Friend X, at a coffee shop, opens peerspeak and starts a gathering labeled
    "HangOut."

packaging/PKGBUILD

@@ -1,7 +1,7 @@
 # Maintainer: mollusk <jitty+lc1iz0dc@protonmail.com>
 pkgname=peerspeak-git
 _pkgname=peerspeak
-pkgver=0.3.0.r229.g7fb1c96
+pkgver=0.4.0.r254.g913b0b6
 pkgrel=1
 pkgdesc="Decentralized peer-to-peer voice chat (Rust/iroh/PipeWire/Opus/iced)"
 arch=('x86_64')

packaging/windows/INSTALL.md

@@ -8,7 +8,7 @@ it once, then you and I connect directly to each other.
 
 ## 1. Install it
 
-1. Double-click **`peerspeak-0.3.0-setup.exe`** (the file I sent you).
+1. Double-click **`peerspeak-0.4.0-setup.exe`** (the file I sent you).
 
 2. **Windows will probably show a blue "Windows protected your PC" warning.**
    This is normal — it shows up for any app that isn't from a big company with a

packaging/windows/README.md

@@ -12,7 +12,7 @@ runtime, so there are no extra DLLs to bundle. The installer payload is just the
 ## Version compatibility
 
 The installer version tracks the crate version in `Cargo.toml` (currently
-**0.3.0**) — keep `MyAppVersion` in `peerspeak.iss` in sync when it changes.
+**0.4.0**) — keep `MyAppVersion` in `peerspeak.iss` in sync when it changes.
 
 Per `VERSIONING.md`, a **MINOR** bump in `0.x` is a **breaking wire change**:
 peers on different MINOR versions can't connect (they fail fast at the

packaging/windows/peerspeak.iss

@@ -12,7 +12,7 @@
 ; (x86_64-pc-windows-gnu, statically linked -- no extra DLLs needed).
 
 #define MyAppName "PeerSpeak"
-#define MyAppVersion "0.3.0"
+#define MyAppVersion "0.4.0"
 #define MyAppPublisher "mollusk"
 #define MyAppExeName "peerspeak.exe"
 

src/avatar.rs

@@ -185,10 +185,129 @@ pub fn initials(name: &str) -> String {
     }
 }
 
+/// A small content-addressed LRU cache mapping image bytes to a built value
+/// (e.g. an `iced` image handle), so the SAME value is reused across redraws
+/// instead of rebuilt every frame. Two Tier C F-03 properties beyond a plain
+/// hash map:
+///
+/// 1. **Bounded** — at most `cap` entries, evicting the least-recently-used on
+///    overflow, so a peer can't grow the cache without limit by publishing an
+///    endless stream of distinct valid avatars.
+/// 2. **Collision-safe** — a hit requires full byte equality, not just a matching
+///    64-bit hash, so a hash collision can never return a different image's value.
+///
+/// Linear scan; intended for small `cap` (tens of entries).
+pub struct ByteLru<V> {
+    cap: usize,
+    /// `(content hash, content bytes, value)`; back = most recently used.
+    entries: Vec<(u64, Vec<u8>, V)>,
+}
+
+impl<V: Clone> ByteLru<V> {
+    /// Create an LRU holding at most `cap` entries (`cap` is clamped to >= 1).
+    pub fn new(cap: usize) -> Self {
+        Self { cap: cap.max(1), entries: Vec::new() }
+    }
+
+    /// Return the cached value for these exact `bytes`, building and inserting it
+    /// on a miss (evicting the least-recently-used entry once over `cap`). A hit
+    /// verifies full byte equality, so a 64-bit hash collision never returns the
+    /// wrong value. A hit also refreshes the entry's recency.
+    pub fn get_or_insert(&mut self, bytes: &[u8], build: impl FnOnce() -> V) -> V {
+        use std::hash::{Hash, Hasher};
+        let mut hasher = std::collections::hash_map::DefaultHasher::new();
+        bytes.hash(&mut hasher);
+        self.get_or_insert_hashed(hasher.finish(), bytes, build)
+    }
+
+    /// Inner seam with the content `hash` supplied explicitly. Production callers
+    /// use [`get_or_insert`]; tests use this to force a hash collision (different
+    /// bytes, same hash) and exercise the byte-equality guard.
+    fn get_or_insert_hashed(&mut self, hash: u64, bytes: &[u8], build: impl FnOnce() -> V) -> V {
+        if let Some(idx) = self
+            .entries
+            .iter()
+            .position(|(h, b, _)| *h == hash && b.as_slice() == bytes)
+        {
+            // LRU touch: move the hit entry to the back (most recent).
+            let entry = self.entries.remove(idx);
+            let val = entry.2.clone();
+            self.entries.push(entry);
+            return val;
+        }
+
+        let val = build();
+        if self.entries.len() >= self.cap {
+            self.entries.remove(0); // evict least-recently-used
+        }
+        self.entries.push((hash, bytes.to_vec(), val.clone()));
+        val
+    }
+
+    #[cfg(test)]
+    fn len(&self) -> usize {
+        self.entries.len()
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
 
+    #[test]
+    fn byte_lru_reuses_value_for_identical_bytes() {
+        let mut lru: ByteLru<u32> = ByteLru::new(4);
+        let mut next = 0u32;
+        let mut build = |lru: &mut ByteLru<u32>, b: &[u8]| {
+            lru.get_or_insert(b, || {
+                next += 1;
+                next
+            })
+        };
+        // Same bytes → same value, built only once.
+        assert_eq!(build(&mut lru, b"alice"), 1);
+        assert_eq!(build(&mut lru, b"alice"), 1);
+        // Different bytes → a freshly built value.
+        assert_eq!(build(&mut lru, b"bob"), 2);
+        assert_eq!(lru.len(), 2);
+    }
+
+    #[test]
+    fn byte_lru_evicts_least_recently_used() {
+        let mut lru: ByteLru<u32> = ByteLru::new(2);
+        let mut n = 0u32;
+        let mut ins = |lru: &mut ByteLru<u32>, b: &[u8]| {
+            lru.get_or_insert(b, || {
+                n += 1;
+                n
+            })
+        };
+        ins(&mut lru, b"a"); // -> 1
+        ins(&mut lru, b"b"); // -> 2, cache = [a, b]
+        ins(&mut lru, b"a"); // touch a, cache = [b, a]
+        ins(&mut lru, b"c"); // evicts LRU (b), cache = [a, c]
+        assert_eq!(lru.len(), 2);
+        // `a` survived (recently touched) → still value 1, not rebuilt.
+        assert_eq!(ins(&mut lru, b"a"), 1);
+        // `b` was evicted → rebuilt with a new value.
+        assert_eq!(ins(&mut lru, b"b"), 4);
+    }
+
+    #[test]
+    fn byte_lru_byte_equality_survives_a_hash_collision() {
+        // Force the SAME 64-bit hash for two DIFFERENT byte strings (the case a
+        // bare-hash cache would alias — Tier C F-03 collision bug).
+        let mut lru: ByteLru<u32> = ByteLru::new(4);
+        assert_eq!(lru.get_or_insert_hashed(42, b"alice", || 1), 1);
+        // `bob` collides on the hash but differs in bytes → a MISS, built fresh,
+        // NOT aliased to alice's value.
+        assert_eq!(lru.get_or_insert_hashed(42, b"bob", || 2), 2);
+        // Both coexist; each re-lookup returns its own value (build closure unused).
+        assert_eq!(lru.get_or_insert_hashed(42, b"alice", || 99), 1);
+        assert_eq!(lru.get_or_insert_hashed(42, b"bob", || 99), 2);
+        assert_eq!(lru.len(), 2);
+    }
+
     #[test]
     fn initials_takes_first_two_words() {
         assert_eq!(initials("Alice"), "A");

src/core/messages.rs

@@ -64,9 +64,16 @@ pub enum CoreCommand {
     /// Set the pixelpass binary location (config override, empty = use `$PATH`).
     /// Sent at startup so screen-share can resolve the binary.
     SetPixelpassPath(Option<String>),
+    /// Enumerate apps currently producing audio (for the screen-share audio
+    /// picker, A23). Replies with [`UiEvent::AudioAppsListed`]. Cheap shell-out;
+    /// safe to call each time the picker opens.
+    ListAudioApps,
     /// Start sharing our screen: spawn a pixelpass host and announce its ticket
     /// on our presence so the room can watch. No-op when not in a call.
-    StartScreenShare,
+    /// `audio_app` selects which app's audio to capture: `Some(name)` captures
+    /// only that app (avoiding the call-loopback echo, A23); `None` shares the
+    /// whole desktop audio (the legacy behavior).
+    StartScreenShare { audio_app: Option<String> },
     /// Stop sharing our screen: kill the pixelpass host and clear the presence
     /// ticket. No-op when not sharing.
     StopScreenShare,
@@ -133,15 +140,29 @@ pub enum UiEvent {
     /// string, used to key their avatar (W4).
     ChatMessage { from: String, name: String, text: String, attachment: Option<crate::files::ChatAttachment> },
     /// An attachment's bytes are now available (auto-fetched for images, or
-    /// fetched on demand for files). Keyed by attachment id so the UI can match
-    /// it to the chat entry.
-    AttachmentReady { id: crate::files::AttachmentId, data: Vec<u8> },
+    /// fetched on demand for files). Keyed by `(from, id)`: the id is
+    /// attacker-chosen, so a malicious peer can reuse a victim's id — the author
+    /// disambiguates whose bytes these are and stops content aliasing (Tier C
+    /// F-12).
+    AttachmentReady { from: EndpointId, id: crate::files::AttachmentId, data: Vec<u8> },
     /// An attachment fetch failed (sender gone, too large, decode error, etc.).
-    AttachmentFailed { id: crate::files::AttachmentId, error: String },
+    AttachmentFailed { from: EndpointId, id: crate::files::AttachmentId, error: String },
+    /// The apps currently producing audio, for the screen-share audio picker
+    /// (A23). Sorted, deduplicated `application.name`s; empty when nothing is
+    /// playing or enumeration isn't available. `app_audio_supported` reports
+    /// whether the resolved pixelpass understands `--strict-audio`: when `false`
+    /// (an older pixelpass) the picker must offer whole-desktop audio only, since
+    /// a per-app share would pass a flag that older binary rejects (audit P2).
+    AudioAppsListed { apps: Vec<String>, app_audio_supported: bool },
     /// Our own screen share started; the UI flips the Share button to "Stop".
     ScreenShareStarted,
     /// Our own screen share stopped (or failed to start).
     ScreenShareStopped,
+    /// Per-app screen-share audio routing state (A23). `true` = the app we chose
+    /// is now reaching viewers; `false` = its audio stopped, so under our strict
+    /// run viewers currently hear silence. The UI shows a transient warning while
+    /// `false`. Only meaningful while sharing a specific app (not whole-desktop).
+    ShareAudioActive(bool),
     /// Our node identity (W7): the current node id string, and whether it is
     /// PERSISTED to disk. Sent once at startup and again after a regenerate.
     /// `persisted = false` means the key file couldn't be read/written and we're

src/core/mod.rs

@@ -134,6 +134,25 @@ type SeenConnected = Arc<std::sync::Mutex<HashSet<EndpointId>>>;
 type KnownPeers =
     Arc<std::sync::Mutex<HashMap<[u8; 32], HashMap<EndpointId, EndpointAddr>>>>;
 
+/// Per-topic cap on the retained rejoin-bootstrap / recovery target table
+/// (Tier C recovery-identity cap). Set comfortably above the live-roster cap
+/// (`gossip::MAX_ACTIVE_PEERS`, 32) so a legitimate room — even one where every
+/// member drops at once during a relay outage — never hits it, while an insider
+/// who grace-cycles distinct identities (join, drop without a signed Leave,
+/// repeat) cannot grow the table without bound. Combined with the recovery
+/// terminal budget (which forgets a retained address when it gives up), abandoned
+/// identities drain on their own, so this cap is a deterministic ceiling rather
+/// than a pinnable slot pool.
+const MAX_RETAINED_PEERS: usize = 64;
+
+/// Whether a peer may be inserted into a retained-target table at `len` entries.
+/// An update to an id already present is always allowed (it only refreshes an
+/// address); a brand-new id is admitted only while below the cap. Mirrors the
+/// gossip roster's `admit_into_roster` reject-when-full admission.
+fn admit_retained(len: usize, is_new_id: bool, cap: usize) -> bool {
+    !is_new_id || len < cap
+}
+
 #[derive(Clone)]
 struct RecoveryContext {
     coordinator: RecoveryCoordinator,
@@ -522,6 +541,7 @@ struct ActiveSession {
     event_task: tokio::task::JoinHandle<()>,
     conn_event_task: tokio::task::JoinHandle<()>,
     recovery_task: tokio::task::JoinHandle<()>,
+    recovery_terminal_task: tokio::task::JoinHandle<()>,
     grace_timers: GraceTimers,
     transport: Arc<IrohTransport>,
     /// Loaded PipeWire echo-cancel module (if enabled); unloads on drop.
@@ -557,6 +577,7 @@ impl ActiveSession {
             handle.abort();
         }
         self.recovery_task.abort();
+        self.recovery_terminal_task.abort();
         crate::log_msg("Aborted tasks");
 
         let audio_backend_clone = audio_backend.clone();
@@ -744,25 +765,70 @@ async fn build_net_stack(
     })
 }
 
+/// Maximum number of *automatic* chat-attachment fetches in flight at once.
+///
+/// Auto-fetch (inline image preview) is triggered by an untrusted peer's chat
+/// message, and each fetch is a detached task that can spend up to ~60s dialing
+/// and reading. Without a bound, a room insider could spam attachment-carrying
+/// chat to accumulate arbitrary pending tasks/dials (Tier C F-02). When the bound
+/// is reached we simply skip the auto-fetch; the descriptor still renders and the
+/// user can fetch it on demand (which is not rate-limited here).
+const MAX_INFLIGHT_ATTACHMENT_FETCHES: usize = 4;
+
+/// In-flight `(author, attachment_id)` markers for bounded, deduplicated auto-
+/// fetches (Tier C F-02). Bounded by [`MAX_INFLIGHT_ATTACHMENT_FETCHES`].
+type InflightAttachments =
+    Arc<std::sync::Mutex<HashSet<(EndpointId, crate::files::AttachmentId)>>>;
+
+/// RAII bookkeeping for one bounded auto-fetch: holds the concurrency permit for
+/// the task's lifetime and clears the in-flight `(author, id)` marker when the
+/// fetch finishes (success OR failure), so the same image can be retried later.
+struct AutoFetchGuard {
+    _permit: tokio::sync::OwnedSemaphorePermit,
+    inflight: InflightAttachments,
+    key: (EndpointId, crate::files::AttachmentId),
+}
+
+impl Drop for AutoFetchGuard {
+    fn drop(&mut self) {
+        self.inflight.lock().unwrap().remove(&self.key);
+    }
+}
+
+/// Whether to AUTO-fetch a chat image attachment. Only authenticated roster
+/// authors qualify (closing the non-roster injection vector), and a `(author,
+/// id)` already being fetched is skipped (dedup). The concurrency bound itself is
+/// enforced separately by the permit. Pure → unit-testable (Tier C F-02).
+fn should_auto_fetch(is_image: bool, author_in_roster: bool, already_inflight: bool) -> bool {
+    is_image && author_in_roster && !already_inflight
+}
+
 /// Fetch a chat attachment's bytes from `from` over the file plane in a detached
 /// task, then report the result to the UI via [`UiEvent::AttachmentReady`] /
-/// [`UiEvent::AttachmentFailed`] keyed by the attachment id. For images
+/// [`UiEvent::AttachmentFailed`], tagged with `from` so the UI keys the bytes by
+/// `(author, id)` and can't alias a same-id attachment from another sender. For images
 /// (`is_image`) the bytes are defensively re-validated (decodable + within pixel
 /// limits) before being handed to the renderer; an "Image" that doesn't decode is
-/// reported as a failure rather than rendered.
+/// reported as a failure rather than rendered. `guard` is `Some` for bounded
+/// auto-fetches and `None` for user-initiated fetches; it is dropped when the
+/// task ends, releasing the concurrency permit and the dedup marker.
 fn spawn_attachment_fetch(
     transport: Arc<IrohTransport>,
     ui_tx: mpsc::Sender<UiEvent>,
     from: EndpointId,
     att: crate::files::ChatAttachment,
     is_image: bool,
+    guard: Option<AutoFetchGuard>,
 ) {
     tokio::spawn(async move {
+        // Held for the whole fetch; dropped here on completion (Tier C F-02).
+        let _guard = guard;
         match transport.fetch_attachment(from, &att).await {
             Ok(data) => {
                 if is_image && crate::files::validate_image_bytes(&data).is_none() {
                     let _ = ui_tx
                         .send(UiEvent::AttachmentFailed {
+                            from,
                             id: att.id,
                             error: "received image failed to decode".to_string(),
                         })
@@ -770,12 +836,12 @@ fn spawn_attachment_fetch(
                     return;
                 }
                 let _ = ui_tx
-                    .send(UiEvent::AttachmentReady { id: att.id, data })
+                    .send(UiEvent::AttachmentReady { from, id: att.id, data })
                     .await;
             }
             Err(e) => {
                 let _ = ui_tx
-                    .send(UiEvent::AttachmentFailed { id: att.id, error: e.to_string() })
+                    .send(UiEvent::AttachmentFailed { from, id: att.id, error: e.to_string() })
                     .await;
             }
         }
@@ -1008,22 +1074,34 @@ async fn run_core_loop(
     // Join, cleared on Leave.
     let current_room: Arc<std::sync::Mutex<Option<crate::presence::RoomPresence>>> =
         Arc::new(std::sync::Mutex::new(None));
+    let presence_rate_limiter =
+        Arc::new(std::sync::Mutex::new(crate::presence::PresenceRateLimiter::default()));
 
     // Reply policy for the idle friends listener (B2): answer friends only, never
     // while invisible (`should_answer`), and report our current gathering so a friend
-    // can one-click join. Reads the shared snapshots, so it stays correct as they
-    // change and survives a network-stack rebuild. Pure-sync (no awaits, no lock held
-    // across one). Built once and handed to every `build_net_stack`.
+    // can one-click join. Rate-limits allowed friends before building a reply, so a
+    // spammy saved peer gets the same silent close as an unauthorized peer. Reads the
+    // shared snapshots, so it stays correct as they change and survives a network-stack
+    // rebuild. Pure-sync (no awaits, no lock held across one). Built once and handed
+    // to every `build_net_stack`.
     let friends_handler: crate::presence_net::Handler = {
         let friends = friends.clone();
         let presence_mode = presence_mode.clone();
         let current_room = current_room.clone();
+        let presence_rate_limiter = presence_rate_limiter.clone();
         Arc::new(move |from| {
             let mode = *presence_mode.lock().unwrap();
             let allowed = crate::presence::should_answer(&from, &friends.lock().unwrap(), mode);
             if !allowed {
                 return None;
             }
+            if !presence_rate_limiter
+                .lock()
+                .unwrap()
+                .allow(from, std::time::Instant::now())
+            {
+                return None;
+            }
             let room = current_room.lock().unwrap().clone();
             Some(crate::presence::ControlMsg::Pong { room })
         })
@@ -1799,7 +1877,7 @@ async fn run_core_loop(
                 // The topic of the room this event loop serves, so peer add/remove
                 // updates the right per-topic bucket in `known_peers` (A8 archive).
                 let room_topic = topic_id;
-                let (recovery_coordinator, recovery_task) =
+                let (recovery_coordinator, recovery_task, recovery_terminal_rx) =
                     RecoveryCoordinator::spawn(room_state.clone());
                 let recovery_context = RecoveryContext {
                     coordinator: recovery_coordinator,
@@ -1808,17 +1886,51 @@ async fn run_core_loop(
                     topic_id,
                 };
                 let recovery_events = recovery_context.clone();
+                // Drain the recovery coordinator's terminal-eviction signals (Tier C
+                // recovery-identity cap). When background recovery exhausts its budget
+                // for a peer, forget its retained dial target so the per-topic retain
+                // table drains, scrub residual seen-connected state, and surface the
+                // failure. A peer that later returns can still rejoin via a gossip
+                // announce, so giving up never blocks a legitimate reconnect.
+                let recovery_terminal_ctx = recovery_context.clone();
+                let seen_connected_terminal = seen_connected.clone();
+                let ui_tx_terminal = ui_tx.clone();
+                let recovery_terminal_task = tokio::spawn(async move {
+                    let mut terminal_rx = recovery_terminal_rx;
+                    while let Some(peer_id) = terminal_rx.recv().await {
+                        crate::log_msg(&format!(
+                            "Background recovery gave up on peer {peer_id:?}; forgetting retained target"
+                        ));
+                        recovery_terminal_ctx.forget(peer_id);
+                        seen_connected_terminal.lock().unwrap().remove(&peer_id);
+                        let _ = ui_tx_terminal
+                            .send(UiEvent::PeerConnectionFailed { id: peer_id })
+                            .await;
+                    }
+                });
                 // Friends store + ui sender, so a connected peer who is a friend has
                 // their saved address auto-healed (W7) — populates `last_addr` so the
                 // presence scheduler can reach them later.
                 let friends_events = friends.clone();
                 let friends_read_only_events = friends_read_only;
+                // Bounded, deduplicated auto-fetch of chat image attachments (Tier C
+                // F-02): the permit pool caps concurrent fetch tasks; the in-flight
+                // set dedups identical (author, id) pairs.
+                let attachment_limiter =
+                    Arc::new(tokio::sync::Semaphore::new(MAX_INFLIGHT_ATTACHMENT_FETCHES));
+                let inflight_attachments: InflightAttachments =
+                    Arc::new(std::sync::Mutex::new(HashSet::new()));
                 let event_task = tokio::spawn(async move {
+                    // The authenticated roster for this room, maintained from the
+                    // same sequential event stream. Only its members may trigger an
+                    // automatic attachment fetch (Tier C F-02).
+                    let mut roster: HashSet<EndpointId> = HashSet::new();
                     while let Some(event) = room_events.recv().await {
                         match event {
                             RoomEvent::PeerJoined(peer_id, state) => {
                                 // A (re)join means the peer is back — cancel any
                                 // pending reconnect grace timer before re-adding it.
+                                roster.insert(peer_id);
                                 cancel_grace_timer(&grace_timers_events, &peer_id);
                                 recovery_events.cancel(peer_id);
                                 transport_events.admit_audio_sender(peer_id);
@@ -1843,13 +1955,22 @@ async fn run_core_loop(
                                     .await;
                                 }
                                 // Retain this peer under this room's topic as a
-                                // future rejoin bootstrap target (A8).
-                                known_peers_events
-                                    .lock()
-                                    .unwrap()
-                                    .entry(room_topic)
-                                    .or_default()
-                                    .insert(peer_id, state.addr.clone());
+                                // future rejoin bootstrap target (A8), bounded by the
+                                // per-topic retain cap (Tier C recovery-identity cap):
+                                // refreshing a peer we already track is always allowed,
+                                // a brand-new identity only while below the cap.
+                                {
+                                    let mut kp = known_peers_events.lock().unwrap();
+                                    let bucket = kp.entry(room_topic).or_default();
+                                    let is_new_id = !bucket.contains_key(&peer_id);
+                                    if admit_retained(bucket.len(), is_new_id, MAX_RETAINED_PEERS) {
+                                        bucket.insert(peer_id, state.addr.clone());
+                                    } else {
+                                        crate::log_msg(&format!(
+                                            "Retain table full ({MAX_RETAINED_PEERS}); not retaining new peer {peer_id:?} for rejoin"
+                                        ));
+                                    }
+                                }
                                 // If a multitrack recording is live, give this peer
                                 // its own stem track (silence-padded back to t=0).
                                 if is_multitrack_events.load(Ordering::Relaxed)
@@ -1862,6 +1983,7 @@ async fn run_core_loop(
                             }
                             RoomEvent::PeerLeft(peer_id) => {
                                 // Graceful leave — evict immediately.
+                                roster.remove(&peer_id);
                                 cancel_grace_timer(&grace_timers_events, &peer_id);
                                 seen_connected_events.lock().unwrap().remove(&peer_id);
                                 // A signed Leave cancels background recovery and
@@ -1899,13 +2021,21 @@ async fn run_core_loop(
                                     .await;
                                 }
                                 // Refresh this room's retained rejoin target with the
-                                // fresh addr (A8).
-                                known_peers_events
-                                    .lock()
-                                    .unwrap()
-                                    .entry(room_topic)
-                                    .or_default()
-                                    .insert(peer_id, state.addr.clone());
+                                // fresh addr (A8), under the per-topic retain cap. A
+                                // re-announce from a peer we already track always
+                                // refreshes; a new identity is bounded by the cap.
+                                {
+                                    let mut kp = known_peers_events.lock().unwrap();
+                                    let bucket = kp.entry(room_topic).or_default();
+                                    let is_new_id = !bucket.contains_key(&peer_id);
+                                    if admit_retained(bucket.len(), is_new_id, MAX_RETAINED_PEERS) {
+                                        bucket.insert(peer_id, state.addr.clone());
+                                    } else {
+                                        crate::log_msg(&format!(
+                                            "Retain table full ({MAX_RETAINED_PEERS}); not retaining new peer {peer_id:?} for rejoin"
+                                        ));
+                                    }
+                                }
                                 let _ = ui_tx_events.send(UiEvent::PeerUpdated { id: peer_id, state }).await;
                             }
                             RoomEvent::ChatMessage { from, name, text, ts: _, attachment } => {
@@ -1913,16 +2043,47 @@ async fn run_core_loop(
                                 // without a click; non-image files wait for an explicit
                                 // FetchAttachment (the "Save" chip). The descriptor was
                                 // already filename-sanitized + size-capped on ingest.
-                                if let Some(att) = attachment.clone()
-                                    && att.kind == crate::files::AttachmentKind::Image
-                                {
-                                    spawn_attachment_fetch(
-                                        transport_events.clone(),
-                                        ui_tx_events.clone(),
-                                        from,
-                                        att,
-                                        true,
-                                    );
+                                //
+                                // The auto path is an untrusted-peer-triggered detached
+                                // task, so it is gated (Tier C F-02): only roster authors
+                                // qualify, identical (author,id) pairs are deduped, and a
+                                // permit pool caps concurrent fetch tasks. The chat TEXT
+                                // is always forwarded (it's sanitized at the UI edge);
+                                // only the fetch is bounded.
+                                if let Some(att) = attachment.clone() {
+                                    let is_image = att.kind == crate::files::AttachmentKind::Image;
+                                    let key = (from, att.id);
+                                    let already_inflight =
+                                        inflight_attachments.lock().unwrap().contains(&key);
+                                    if should_auto_fetch(is_image, roster.contains(&from), already_inflight) {
+                                        // Reserve the dedup slot, then a permit. If the
+                                        // pool is exhausted, drop the auto-fetch (and the
+                                        // dedup marker) — the descriptor still shows and
+                                        // the user can fetch on demand.
+                                        inflight_attachments.lock().unwrap().insert(key);
+                                        match attachment_limiter.clone().try_acquire_owned() {
+                                            Ok(permit) => {
+                                                spawn_attachment_fetch(
+                                                    transport_events.clone(),
+                                                    ui_tx_events.clone(),
+                                                    from,
+                                                    att,
+                                                    true,
+                                                    Some(AutoFetchGuard {
+                                                        _permit: permit,
+                                                        inflight: inflight_attachments.clone(),
+                                                        key,
+                                                    }),
+                                                );
+                                            }
+                                            Err(_) => {
+                                                inflight_attachments.lock().unwrap().remove(&key);
+                                                crate::log_msg(
+                                                    "Chat attachment auto-fetch limit reached; skipping (fetch on demand)",
+                                                );
+                                            }
+                                        }
+                                    }
                                 }
                                 let _ = ui_tx_events.send(UiEvent::ChatMessage {
                                     from: from.to_string(),
@@ -1992,6 +2153,7 @@ async fn run_core_loop(
                     event_task,
                     conn_event_task,
                     recovery_task,
+                    recovery_terminal_task,
                     grace_timers,
                     transport: transport.clone(),
                     #[cfg(target_os = "linux")]
@@ -2467,12 +2629,15 @@ async fn run_core_loop(
             CoreCommand::FetchAttachment { from, attachment } => {
                 if let Some(session) = &active_session {
                     let is_image = attachment.kind == crate::files::AttachmentKind::Image;
+                    // User-initiated (the "Save" chip): not bounded here — a human
+                    // click rate-limits it. The auto path (F-02) passes a guard.
                     spawn_attachment_fetch(
                         session.transport.clone(),
                         ui_tx.clone(),
                         from,
                         attachment,
                         is_image,
+                        None,
                     );
                 }
             }
@@ -2481,7 +2646,29 @@ async fn run_core_loop(
                 pixelpass_override = path.filter(|p| !p.trim().is_empty());
             }
 
-            CoreCommand::StartScreenShare => {
+            CoreCommand::ListAudioApps => {
+                // Probe whether this pixelpass supports `--strict-audio` before
+                // offering per-app capture: an older binary would reject the flag
+                // and hard-fail the share (audit P2). When unsupported (or
+                // pixelpass is missing), skip enumeration and let the picker show
+                // whole-desktop audio only — never a best-effort `--app` that
+                // would reopen the A23 echo.
+                let app_audio_supported =
+                    match crate::screenshare::pixelpass_path(pixelpass_override.as_deref()) {
+                        Some(bin) => crate::screenshare::supports_strict_audio(&bin).await,
+                        None => false,
+                    };
+                let apps = if app_audio_supported {
+                    crate::screenshare::list_audio_apps().await
+                } else {
+                    Vec::new()
+                };
+                let _ = ui_tx
+                    .send(UiEvent::AudioAppsListed { apps, app_audio_supported })
+                    .await;
+            }
+
+            CoreCommand::StartScreenShare { audio_app } => {
                 let Some(session) = &mut active_session else {
                     let _ = ui_tx
                         .send(UiEvent::Error("Join a call before sharing your screen".into()))
@@ -2502,7 +2689,33 @@ async fn run_core_loop(
                         continue;
                     }
                 };
-                match crate::screenshare::spawn_host(&bin).await {
+                // Forward pixelpass `app_audio` events (only emitted when an app
+                // is selected) to the UI so it can warn when the chosen app's
+                // audio drops. The channel closes when the host dies (drain hits
+                // EOF), ending the forwarder task on its own.
+                let notices = audio_app.as_deref().map(|_| {
+                    let (tx, mut rx) =
+                        tokio::sync::mpsc::unbounded_channel::<crate::screenshare::PixelpassEvent>();
+                    let ui_tx_notices = ui_tx.clone();
+                    tokio::spawn(async move {
+                        while let Some(ev) = rx.recv().await {
+                            let active = match ev {
+                                crate::screenshare::PixelpassEvent::AppAudioRouted => true,
+                                crate::screenshare::PixelpassEvent::AppAudioLost => false,
+                                _ => continue,
+                            };
+                            if ui_tx_notices
+                                .send(UiEvent::ShareAudioActive(active))
+                                .await
+                                .is_err()
+                            {
+                                break;
+                            }
+                        }
+                    });
+                    tx
+                });
+                match crate::screenshare::spawn_host(&bin, audio_app.as_deref(), notices).await {
                     Ok((child, ticket)) => {
                         crate::log_msg("Screen share host started");
                         session.screenshare_host = Some(child);
@@ -2575,11 +2788,39 @@ async fn run_core_loop(
 #[cfg(test)]
 mod tests {
     use super::{
-        apply_volume, audio_datagram_len_ok, frame_level, mix_frames, mix_stereo_frames,
-        next_game_change, stereo_to_mono, KnownPeers, MicLevelMeter, PeerSpeakTicket,
-        MAX_OPUS_PAYLOAD, MIC_LEVEL_REPORT_SAMPLES,
+        admit_retained, apply_volume, audio_datagram_len_ok, frame_level, mix_frames,
+        mix_stereo_frames, next_game_change, should_auto_fetch, stereo_to_mono, KnownPeers,
+        MicLevelMeter, PeerSpeakTicket, MAX_OPUS_PAYLOAD, MAX_RETAINED_PEERS,
+        MIC_LEVEL_REPORT_SAMPLES,
     };
 
+    #[test]
+    fn admit_retained_rejects_only_new_ids_at_the_cap() {
+        // Below the cap, a brand-new identity is retained.
+        assert!(admit_retained(0, true, MAX_RETAINED_PEERS));
+        assert!(admit_retained(MAX_RETAINED_PEERS - 1, true, MAX_RETAINED_PEERS));
+        // At the cap, a brand-new identity is refused — this is the bound that stops
+        // an insider grace-cycling distinct identities from growing the retain table.
+        assert!(!admit_retained(MAX_RETAINED_PEERS, true, MAX_RETAINED_PEERS));
+        // A peer already tracked always refreshes, even at (or past) the cap: it only
+        // updates an existing address and never adds a slot.
+        assert!(admit_retained(MAX_RETAINED_PEERS, false, MAX_RETAINED_PEERS));
+        assert!(admit_retained(MAX_RETAINED_PEERS + 5, false, MAX_RETAINED_PEERS));
+    }
+
+    #[test]
+    fn auto_fetch_only_for_roster_images_not_already_inflight() {
+        // The happy path: a roster author's brand-new image attachment.
+        assert!(should_auto_fetch(true, true, false));
+        // A non-image (generic file) never auto-fetches — it waits for "Save".
+        assert!(!should_auto_fetch(false, true, false));
+        // A non-roster author (e.g. a sock puppet that never announced) is rejected,
+        // closing the F-02 unbounded-task vector.
+        assert!(!should_auto_fetch(true, false, false));
+        // An identical (author,id) already being fetched is deduped.
+        assert!(!should_auto_fetch(true, true, true));
+    }
+
     #[test]
     fn retained_peers_are_shared_by_topic_across_restamped_tickets() {
         let topic_id = [23u8; 32];

src/core/recovery.rs

@@ -22,6 +22,27 @@ fn recovery_delay(attempt: usize) -> Duration {
     RECOVERY_DELAYS[attempt.min(RECOVERY_DELAYS.len() - 1)]
 }
 
+/// Terminal retry budget for background recovery. After this many failed attempts
+/// the coordinator gives up: it drops the entry, frees the active slot, and signals
+/// the event task to forget the retained address (Tier C recovery-identity cap).
+///
+/// With the [`RECOVERY_DELAYS`] backoff this is roughly seven minutes of dialing
+/// (1+2+4+8+15+30+60s, then 60s steps), far beyond any normal transient outage. A
+/// genuine peer returning after a longer outage still rejoins on its own via a
+/// gossip announce, so giving up only stops us from dialing a peer that is not
+/// coming back — it does not break legitimate reconnect-after-outage.
+const RECOVERY_TERMINAL_ATTEMPTS: usize = 12;
+
+/// Capacity of the terminal-eviction notification channel. Bounded; on the rare
+/// event of saturation the entry is still removed (the dial work stops) and only
+/// the retained-address forget is skipped, which the per-topic retain cap bounds.
+const RECOVERY_TERMINAL_CAPACITY: usize = 64;
+
+/// Whether `attempt` completed recoveries have exhausted the terminal budget.
+fn recovery_is_terminal(attempt: usize, max_attempts: usize) -> bool {
+    attempt >= max_attempts
+}
+
 enum RecoveryCommand {
     Start {
         peer_id: EndpointId,
@@ -60,19 +81,24 @@ pub(super) struct RecoveryCoordinator {
 }
 
 impl RecoveryCoordinator {
-    pub(super) fn spawn(room_state: Arc<IrohGossipState>) -> (Self, JoinHandle<()>) {
+    pub(super) fn spawn(
+        room_state: Arc<IrohGossipState>,
+    ) -> (Self, JoinHandle<()>, mpsc::Receiver<EndpointId>) {
         Self::spawn_inner(room_state)
     }
 
-    fn spawn_inner(room_state: Arc<dyn RecoveryRoom>) -> (Self, JoinHandle<()>) {
+    fn spawn_inner(
+        room_state: Arc<dyn RecoveryRoom>,
+    ) -> (Self, JoinHandle<()>, mpsc::Receiver<EndpointId>) {
         let (tx, rx) = mpsc::channel(RECOVERY_COMMAND_CAPACITY);
+        let (terminal_tx, terminal_rx) = mpsc::channel(RECOVERY_TERMINAL_CAPACITY);
         let active = Arc::new(Mutex::new(HashSet::new()));
         let handle = Self {
             tx,
             active: active.clone(),
         };
-        let task = tokio::spawn(run_coordinator(room_state, active, rx));
-        (handle, task)
+        let task = tokio::spawn(run_coordinator(room_state, active, rx, terminal_tx));
+        (handle, task, terminal_rx)
     }
 
     /// Reserve one recovery slot before grace-expiry teardown begins. Returns
@@ -116,6 +142,7 @@ async fn run_coordinator(
     room_state: Arc<dyn RecoveryRoom>,
     active: Arc<Mutex<HashSet<EndpointId>>>,
     mut rx: mpsc::Receiver<RecoveryCommand>,
+    terminal_tx: mpsc::Sender<EndpointId>,
 ) {
     let mut entries: HashMap<EndpointId, RecoveryEntry> = HashMap::new();
 
@@ -155,9 +182,24 @@ async fn run_coordinator(
                                     entries.remove(&peer_id);
                                     continue;
                                 }
-                                if let Some(entry) = entries.get_mut(&peer_id) {
+                                // Advance the backoff, then check the terminal budget.
+                                // `attempt` counts completed attempts, so the delay
+                                // uses the current value before it is incremented.
+                                let terminal = if let Some(entry) = entries.get_mut(&peer_id) {
                                     entry.next_attempt = scheduled_at + recovery_delay(entry.attempt);
                                     entry.attempt = entry.attempt.saturating_add(1);
+                                    recovery_is_terminal(entry.attempt, RECOVERY_TERMINAL_ATTEMPTS)
+                                } else {
+                                    false
+                                };
+                                if terminal {
+                                    // Give up on a peer that has not returned within the
+                                    // budget: drop its entry, free the active slot, and
+                                    // signal the event task to forget its retained
+                                    // address so the per-topic retain table drains.
+                                    entries.remove(&peer_id);
+                                    active.lock().unwrap().remove(&peer_id);
+                                    let _ = terminal_tx.try_send(peer_id);
                                 }
                             }
                         }
@@ -210,6 +252,23 @@ mod tests {
         assert_eq!(actual, vec![1, 2, 4, 8, 15, 30, 60, 60, 60, 60]);
     }
 
+    #[test]
+    fn recovery_budget_is_terminal_only_at_or_past_the_cap() {
+        assert!(!recovery_is_terminal(0, RECOVERY_TERMINAL_ATTEMPTS));
+        assert!(!recovery_is_terminal(
+            RECOVERY_TERMINAL_ATTEMPTS - 1,
+            RECOVERY_TERMINAL_ATTEMPTS
+        ));
+        assert!(recovery_is_terminal(
+            RECOVERY_TERMINAL_ATTEMPTS,
+            RECOVERY_TERMINAL_ATTEMPTS
+        ));
+        assert!(recovery_is_terminal(
+            RECOVERY_TERMINAL_ATTEMPTS + 5,
+            RECOVERY_TERMINAL_ATTEMPTS
+        ));
+    }
+
     #[test]
     fn recovery_slots_are_deduplicated_and_cancel_immediately() {
         let (tx, mut rx) = mpsc::channel(4);
@@ -244,9 +303,10 @@ mod tests {
     #[tokio::test]
     async fn coordinator_attempts_rebootstrap_immediately() {
         let (attempts_tx, mut attempts_rx) = mpsc::unbounded_channel();
-        let (coordinator, task) = RecoveryCoordinator::spawn_inner(Arc::new(RecordingRoom {
-            attempts: attempts_tx,
-        }));
+        let (coordinator, task, _terminal_rx) =
+            RecoveryCoordinator::spawn_inner(Arc::new(RecordingRoom {
+                attempts: attempts_tx,
+            }));
         let peer_id = SecretKey::generate().public();
         let addr = EndpointAddr::from(peer_id);
 

src/network/gossip.rs

@@ -1,11 +1,11 @@
 use crate::network::{RoomState, NetError, PeerState, RoomEvent, PeerSpeakTicket};
-use iroh::{Endpoint, EndpointAddr, EndpointId, SecretKey, Signature};
+use iroh::{Endpoint, EndpointAddr, EndpointId, SecretKey, Signature, TransportAddr};
 use iroh_gossip::net::Gossip;
 use iroh_gossip::proto::TopicId;
 use tokio::sync::mpsc;
 use tokio::sync::mpsc::Receiver;
 use std::sync::{Arc, Mutex};
-use std::collections::{HashMap, HashSet};
+use std::collections::{BTreeSet, HashMap, HashSet};
 use async_trait::async_trait;
 use tokio_stream::StreamExt;
 use serde::{Serialize, Deserialize};
@@ -131,6 +131,89 @@ fn admit_state_mutation(
     true
 }
 
+/// Size at which we prune stale entries from the replay-tracking map (Tier C
+/// F-01 audit). `admit_state_mutation` records `(author, kind)` for every signed
+/// mutation, so an insider sending validly signed `Leave`s from unlimited
+/// generated keys would otherwise grow it for the room's lifetime. A mutation
+/// older than the freshness window can never be the deciding `last_ts` for an
+/// in-window message — `verify_gossip`'s timestamp check rejects such a replay
+/// first — so dropping those entries cannot weaken replay protection; it bounds
+/// the map to roughly the authors seen within one freshness window.
+const STATE_MUTATIONS_SOFT_CAP: usize = 256;
+
+/// Drop replay-tracking entries whose timestamp is older than `window_ms` before
+/// `now_ms` (see [`STATE_MUTATIONS_SOFT_CAP`]). Pure → unit-testable.
+fn prune_stale_mutations(
+    seen: &mut HashMap<(EndpointId, StateMutationKind), u64>,
+    now_ms: u64,
+    window_ms: u64,
+) {
+    let floor = now_ms.saturating_sub(window_ms);
+    seen.retain(|_, last_ts| *last_ts >= floor);
+}
+
+/// Maximum number of distinct peers we hold in a room roster at once.
+///
+/// Everyone with the room ticket is an authenticated *insider*: a signature only
+/// proves ownership of the generated keypair it was made with, not that the
+/// author is a distinct human. A malicious member can therefore mint many valid
+/// signed identities. Voice is full-mesh (each peer dials every other), so a real
+/// room is realistically well under this bound; the cap exists purely so a flood
+/// of sock-puppet `Announce`s can't grow our peer map / audio supervisors / dials
+/// without limit (Tier C F-01).
+const MAX_ACTIVE_PEERS: usize = 32;
+
+/// Maximum transport addresses we retain from a single peer announce. iroh
+/// normally advertises a handful (a few LAN/WAN IP candidates plus one home
+/// relay); the cap stops an insider stuffing a large unique address set into each
+/// announce to inflate the address lookup and the dialer's candidate list.
+const MAX_PEER_ADDRS: usize = 8;
+
+/// Maximum byte length of a relay URL we accept inside a peer address. A relay
+/// URL is normal-length; anything longer is dropped rather than retained.
+const MAX_RELAY_URL_LEN: usize = 256;
+
+/// Bound an untrusted peer's advertised address set before we retain it / hand it
+/// to the address lookup and dialer (Tier C F-01). Drops transport kinds we never
+/// use (`Custom`) and over-long relay URLs, then truncates to at most
+/// [`MAX_PEER_ADDRS`] addresses. `BTreeSet` iteration is deterministic, so the
+/// kept subset is stable. Pure → unit-testable.
+fn sanitize_endpoint_addr(addr: &EndpointAddr) -> EndpointAddr {
+    let addrs: BTreeSet<TransportAddr> = addr
+        .addrs
+        .iter()
+        .filter(|a| match a {
+            TransportAddr::Relay(url) => url.as_str().len() <= MAX_RELAY_URL_LEN,
+            TransportAddr::Ip(_) => true,
+            // `TransportAddr` is #[non_exhaustive]; we only speak IP + relay, so
+            // anything else (Custom / future kinds) is dropped, not retained.
+            _ => false,
+        })
+        .take(MAX_PEER_ADDRS)
+        .cloned()
+        .collect();
+    EndpointAddr { id: addr.id, addrs }
+}
+
+/// Whether an `Announce` may enter the roster. Only a brand-new author
+/// (`subject_to_cap`) is gated by [`MAX_ACTIVE_PEERS`]; updates to an
+/// already-present peer AND re-announces from a peer mid-reconnect (which
+/// already held a slot) always pass — exempting reconnects keeps a full room
+/// from rejecting a legitimately reconnecting member and orphaning its recovery
+/// state (Tier C F-01 audit). Pure → unit-testable.
+fn admit_into_roster(roster_len: usize, subject_to_cap: bool, max_peers: usize) -> bool {
+    !subject_to_cap || roster_len < max_peers
+}
+
+/// Whether a received `Announce`'s author is gated by the roster cap. A peer
+/// already in the roster (`is_new == false`, an ordinary update) or one
+/// mid-reconnect (`is_reconnecting`, it already held a slot) is exempt; only a
+/// brand-new author counts against [`MAX_ACTIVE_PEERS`] (Tier C F-01 audit).
+/// Pure → unit-testable.
+fn announce_subject_to_cap(is_new: bool, is_reconnecting: bool) -> bool {
+    is_new && !is_reconnecting
+}
+
 fn peer_state_for_log(state: &PeerState) -> String {
     format!(
         "name={:?}, muted={}, addr_id={}, addrs={}, sharing={}, game={:?}",
@@ -390,6 +473,18 @@ impl RoomState for IrohGossipState {
                                     continue;
                                 }
 
+                                // Keep the replay-tracking map bounded: prune entries
+                                // older than the freshness window once it grows past the
+                                // soft cap (Tier C F-01 audit). Stale entries can't gate
+                                // an in-window message, so this never weakens replay
+                                // protection.
+                                if state_mutations_seen.len() > STATE_MUTATIONS_SOFT_CAP {
+                                    prune_stale_mutations(
+                                        &mut state_mutations_seen,
+                                        now_millis(),
+                                        GOSSIP_FRESHNESS_MS,
+                                    );
+                                }
                                 if !admit_state_mutation(
                                     &mut state_mutations_seen,
                                     payload.author,
@@ -436,16 +531,49 @@ impl RoomState for IrohGossipState {
                                             let cleaned = crate::sanitize::sanitize_game_label(&g);
                                             (!cleaned.is_empty()).then_some(cleaned)
                                         });
-                                        disconnected_peers.lock().unwrap().remove(&payload.author);
-                                        let (is_new, state_changed) = {
+                                        // Bound an insider's advertised address set
+                                        // before we retain it / hand it to the dialer
+                                        // (Tier C F-01).
+                                        state.addr = sanitize_endpoint_addr(&state.addr);
+                                        // A peer reconnecting from a transient drop sits
+                                        // in `disconnected_peers` (not the live roster);
+                                        // it already held a slot, so it must be re-admitted
+                                        // regardless of the cap, and its disconnect marker
+                                        // cleared ONLY once re-admitted — clearing it before
+                                        // a possible reject would orphan its recovery state
+                                        // (Tier C F-01 audit).
+                                        let is_reconnecting =
+                                            disconnected_peers.lock().unwrap().contains(&payload.author);
+                                        let admitted = {
                                             let mut peer_map = peers.lock().unwrap();
                                             let is_new = !peer_map.contains_key(&payload.author);
-                                            let state_changed = peer_map.get(&payload.author) != Some(&state);
-                                            if is_new || state_changed {
-                                                peer_map.insert(payload.author, state.clone());
+                                            // Cap the roster so a flood of signed
+                                            // sock-puppet identities can't grow our
+                                            // memory/tasks/dials without bound (Tier C
+                                            // F-01). Existing-peer updates and reconnects
+                                            // are exempt; only brand-new authors are gated.
+                                            let subject_to_cap = announce_subject_to_cap(is_new, is_reconnecting);
+                                            if !admit_into_roster(peer_map.len(), subject_to_cap, MAX_ACTIVE_PEERS) {
+                                                None
+                                            } else {
+                                                let state_changed = peer_map.get(&payload.author) != Some(&state);
+                                                if is_new || state_changed {
+                                                    peer_map.insert(payload.author, state.clone());
+                                                }
+                                                Some((is_new, state_changed))
                                             }
-                                            (is_new, state_changed)
                                         };
+                                        let Some((is_new, state_changed)) = admitted else {
+                                            crate::log_msg(&format!(
+                                                "Gossip roster full ({MAX_ACTIVE_PEERS}); rejecting new peer {}",
+                                                crate::short_id(&payload.author.to_string())
+                                            ));
+                                            continue;
+                                        };
+                                        // Admitted — now it is safe to clear any reconnect
+                                        // marker (a rejected announce above leaves it intact
+                                        // so a later signed Leave still cleans up).
+                                        disconnected_peers.lock().unwrap().remove(&payload.author);
 
                                         if is_new {
                                             crate::log_msg(&format!(
@@ -453,7 +581,12 @@ impl RoomState for IrohGossipState {
                                                 crate::short_id(&payload.author.to_string()),
                                                 peer_state_for_log(&state)
                                             ));
-                                            address_lookup.add_endpoint_info(state.addr.clone());
+                                            // Replace (not union) the lookup's record for
+                                            // this id with the authenticated, sanitized
+                                            // address set, so leave/re-announce cycles
+                                            // can't accumulate attacker-supplied history
+                                            // (Tier C F-01).
+                                            let _ = address_lookup.set_endpoint_info(state.addr.clone());
                                             let _ = event_tx.send(RoomEvent::PeerJoined(payload.author, state)).await;
                                         } else if state_changed {
                                             crate::log_msg(&format!(
@@ -466,6 +599,11 @@ impl RoomState for IrohGossipState {
                                     }
                                     GossipMessage::Leave => {
                                         crate::log_msg(&format!("Gossip peer leave request from author={:?}", payload.author));
+                                        // Drop this id's address-lookup entry so cycling
+                                        // distinct identities through Announce→Leave can't
+                                        // grow the lookup for the room's lifetime (Tier C
+                                        // F-01 audit). Re-announce re-populates it.
+                                        let _ = address_lookup.remove_endpoint_info(payload.author);
                                         let removed = peers.lock().unwrap().remove(&payload.author).is_some();
                                         let was_disconnected = disconnected_peers
                                             .lock()
@@ -768,6 +906,84 @@ mod tests {
         assert!(!bootstrap.contains(&me));
     }
 
+    #[test]
+    fn admit_into_roster_caps_new_authors_but_not_updates() {
+        // New authors are admitted while there's room...
+        assert!(admit_into_roster(0, true, 3));
+        assert!(admit_into_roster(2, true, 3));
+        // ...rejected once the roster is full...
+        assert!(!admit_into_roster(3, true, 3));
+        assert!(!admit_into_roster(10, true, 3));
+        // ...but an existing peer's update always passes, even at/over the cap.
+        assert!(admit_into_roster(3, false, 3));
+        assert!(admit_into_roster(99, false, 3));
+    }
+
+    #[test]
+    fn reconnecting_and_existing_peers_are_exempt_from_the_cap() {
+        // A brand-new author counts against the cap...
+        assert!(announce_subject_to_cap(/* is_new */ true, /* is_reconnecting */ false));
+        // ...but an ordinary update from an in-roster peer does not...
+        assert!(!announce_subject_to_cap(false, false));
+        // ...and neither does a re-announce from a peer mid-reconnect, even
+        // though it was removed from the live roster (the F-01-audit fix: a full
+        // room must not reject a legitimately reconnecting member).
+        assert!(!announce_subject_to_cap(true, true));
+        // Combined with admit_into_roster: a reconnecting author passes at a full
+        // roster, a brand-new one does not.
+        assert!(admit_into_roster(3, announce_subject_to_cap(true, true), 3));
+        assert!(!admit_into_roster(3, announce_subject_to_cap(true, false), 3));
+    }
+
+    #[test]
+    fn prune_stale_mutations_drops_only_out_of_window_entries() {
+        let a = fresh_id();
+        let b = fresh_id();
+        let mut seen = HashMap::new();
+        seen.insert((a, StateMutationKind::Announce), 10_000u64);
+        seen.insert((b, StateMutationKind::Leave), 250_000u64);
+        // now = 300_000, window = 120_000 → floor 180_000. The 10_000 entry is
+        // stale (and could never gate an in-window message), the 250_000 is live.
+        prune_stale_mutations(&mut seen, 300_000, GOSSIP_FRESHNESS_MS);
+        assert_eq!(seen.len(), 1);
+        assert!(seen.contains_key(&(b, StateMutationKind::Leave)));
+        assert!(!seen.contains_key(&(a, StateMutationKind::Announce)));
+    }
+
+    #[test]
+    fn sanitize_endpoint_addr_caps_address_count() {
+        use std::net::SocketAddr;
+        let id = fresh_id();
+        // An insider stuffs far more addresses than MAX_PEER_ADDRS into one announce.
+        let many: Vec<TransportAddr> = (0..(MAX_PEER_ADDRS as u16 + 50))
+            .map(|i| TransportAddr::Ip(SocketAddr::from(([127, 0, 0, 1], 1000 + i))))
+            .collect();
+        let addr = EndpointAddr::from_parts(id, many);
+        let out = sanitize_endpoint_addr(&addr);
+        assert_eq!(out.id, id);
+        assert_eq!(out.addrs.len(), MAX_PEER_ADDRS);
+    }
+
+    #[test]
+    fn sanitize_endpoint_addr_drops_overlong_relay_url() {
+        use std::str::FromStr;
+        let id = fresh_id();
+        let short = iroh::RelayUrl::from_str("https://relay.example/").unwrap();
+        let long = iroh::RelayUrl::from_str(&format!(
+            "https://relay.example/{}",
+            "a".repeat(MAX_RELAY_URL_LEN)
+        ))
+        .unwrap();
+        assert!(long.as_str().len() > MAX_RELAY_URL_LEN);
+        let addr = EndpointAddr::from_parts(
+            id,
+            [TransportAddr::Relay(short.clone()), TransportAddr::Relay(long)],
+        );
+        let out = sanitize_endpoint_addr(&addr);
+        let relays: Vec<_> = out.relay_urls().cloned().collect();
+        assert_eq!(relays, vec![short], "over-long relay URL must be dropped");
+    }
+
     #[test]
     fn test_gossip_message_leave_round_trip() {
         let original = GossipMessage::Leave;

src/presence.rs

@@ -15,6 +15,16 @@
 use crate::friends::FriendStore;
 use iroh::EndpointId;
 use serde::{Deserialize, Serialize};
+use std::collections::HashMap;
+use std::time::{Duration, Instant};
+
+/// Maximum immediate presence replies to one friend before throttling. Normal
+/// presence polling is once per minute, so this only catches repeated/manual or
+/// abusive probes while still allowing a short burst after app startup.
+pub const PRESENCE_RATE_LIMIT_BURST: u32 = 4;
+
+/// Refill one presence-reply token per friend at this cadence.
+pub const PRESENCE_RATE_LIMIT_REFILL: Duration = Duration::from_secs(15);
 
 /// The user's presence posture — how reachable they are to friends while idle.
 /// Persisted in `AppConfig`; the default keeps you privately reachable to friends
@@ -100,6 +110,48 @@ pub fn should_answer(from: &EndpointId, friends: &FriendStore, mode: PresenceMod
     mode.answers_pings() && friends.contains(from)
 }
 
+#[derive(Debug, Clone)]
+struct RateBucket {
+    tokens: u32,
+    last_refill: Instant,
+}
+
+/// Per-friend limiter for inbound presence pings. It is intentionally keyed by
+/// the authenticated connection id, not payload data. Callers should only invoke
+/// it after [`should_answer`] passes, so strangers do not consume memory here.
+#[derive(Debug, Default, Clone)]
+pub struct PresenceRateLimiter {
+    buckets: HashMap<EndpointId, RateBucket>,
+}
+
+impl PresenceRateLimiter {
+    /// Return whether `from` may receive a presence reply at `now`.
+    ///
+    /// This is a token bucket: each friend starts with a small burst and regains
+    /// one token every [`PRESENCE_RATE_LIMIT_REFILL`]. A denied probe should be
+    /// answered with no data, matching the listener's "reveal nothing" policy.
+    pub fn allow(&mut self, from: EndpointId, now: Instant) -> bool {
+        let bucket = self.buckets.entry(from).or_insert(RateBucket {
+            tokens: PRESENCE_RATE_LIMIT_BURST,
+            last_refill: now,
+        });
+
+        let elapsed = now.saturating_duration_since(bucket.last_refill);
+        let refill = elapsed.as_secs() / PRESENCE_RATE_LIMIT_REFILL.as_secs();
+        if refill > 0 {
+            let refill = refill.min(u32::MAX as u64) as u32;
+            bucket.tokens = PRESENCE_RATE_LIMIT_BURST.min(bucket.tokens.saturating_add(refill));
+            bucket.last_refill = now;
+        }
+
+        if bucket.tokens == 0 {
+            return false;
+        }
+        bucket.tokens -= 1;
+        true
+    }
+}
+
 /// What we learned about a friend from a successful ping reply.
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub enum FriendPresence {
@@ -177,6 +229,36 @@ mod tests {
         assert!(!should_answer(&stranger, &friends, PresenceMode::Invisible));
     }
 
+    #[test]
+    fn presence_rate_limiter_allows_a_small_burst_then_refills() {
+        let mut limiter = PresenceRateLimiter::default();
+        let friend = id();
+        let now = Instant::now();
+
+        for _ in 0..PRESENCE_RATE_LIMIT_BURST {
+            assert!(limiter.allow(friend, now));
+        }
+        assert!(!limiter.allow(friend, now));
+        assert!(!limiter.allow(friend, now + PRESENCE_RATE_LIMIT_REFILL - Duration::from_millis(1)));
+
+        assert!(limiter.allow(friend, now + PRESENCE_RATE_LIMIT_REFILL));
+        assert!(!limiter.allow(friend, now + PRESENCE_RATE_LIMIT_REFILL));
+    }
+
+    #[test]
+    fn presence_rate_limiter_is_per_peer() {
+        let mut limiter = PresenceRateLimiter::default();
+        let a = id();
+        let b = id();
+        let now = Instant::now();
+
+        for _ in 0..PRESENCE_RATE_LIMIT_BURST {
+            assert!(limiter.allow(a, now));
+        }
+        assert!(!limiter.allow(a, now));
+        assert!(limiter.allow(b, now));
+    }
+
     #[test]
     fn presence_mode_flags() {
         assert!(PresenceMode::Discoverable.publishes_to_discovery());

src/screenshare/mod.rs

@@ -39,6 +39,10 @@ fn pixelpass_path_candidates(dir: &Path) -> [PathBuf; 1] {
 /// growth, but reject unbounded gossip payloads before the UI offers "Watch".
 const MAX_TICKET_LEN: usize = 512;
 
+/// Upper bound on a PipeWire `application.name` we'll pass to `--app`. Real names
+/// are short ("Firefox", "mpv"); this only guards against a pathological value.
+const MAX_APP_NAME_LEN: usize = 256;
+
 /// How long to wait for the host to emit its ticket / the viewer to connect
 /// before giving up and killing the child. Startup is normally sub-second; this
 /// is only a safety net so a hung pixelpass can't wedge the caller forever.
@@ -64,6 +68,12 @@ pub enum PixelpassEvent {
     CaptureStarted,
     /// Host: capture pipeline torn down (on last viewer).
     CaptureStopped,
+    /// Host (per-app audio): the chosen app's audio is now reaching viewers.
+    AppAudioRouted,
+    /// Host (per-app audio): the chosen app's last audio stream went away. Under
+    /// our `--strict-audio` run this means viewers now hear silence (not the call
+    /// echo) until the app produces audio again — we surface it as a warning.
+    AppAudioLost,
     /// A recognized event we don't act on (e.g. `host_info`).
     Other,
 }
@@ -98,6 +108,11 @@ pub fn parse_pixelpass_event(line: &str) -> Option<PixelpassEvent> {
             Some("stopped") => PixelpassEvent::CaptureStopped,
             _ => PixelpassEvent::Other,
         },
+        "app_audio" => match v.get("state").and_then(|s| s.as_str()) {
+            Some("routed") => PixelpassEvent::AppAudioRouted,
+            Some("lost") => PixelpassEvent::AppAudioLost,
+            _ => PixelpassEvent::Other,
+        },
         _ => PixelpassEvent::Other,
     };
     Some(ev)
@@ -107,6 +122,144 @@ fn json_u32(v: &serde_json::Value, key: &str) -> u32 {
     v.get(key).and_then(|x| x.as_u64()).unwrap_or(0) as u32
 }
 
+/// Build the argv for a pixelpass *host*. Always `--host --output json`; when
+/// `audio_app` is `Some`, append `--app=<name> --strict-audio` so pixelpass
+/// captures only that app's audio instead of the whole desktop sink monitor
+/// (which contains our own call playout → the viewer would hear themselves
+/// echoed back, backlog A23).
+///
+/// `--strict-audio` is what makes the fix a guarantee rather than best-effort:
+/// without it, pixelpass falls back to the whole-desktop loopback before the
+/// app's first stream routes and again if the app's audio later stops — both of
+/// which reintroduce the echo. With it, the viewer hears only the chosen app (or
+/// silence), and pixelpass emits `app_audio` events we surface as a warning.
+///
+/// The name is passed in the single-token `--app=<name>` form so a value that
+/// happens to begin with `-` can never be reparsed as a pixelpass flag (clap
+/// otherwise rejects hyphen-leading option values). The name is locally chosen
+/// (our own enumeration / the user's pick), not peer-supplied, but is still
+/// sanitized via [`sanitize_app_name`] before reaching here. Pure: no I/O.
+pub fn host_args(audio_app: Option<&str>) -> Vec<String> {
+    let mut args = vec![
+        "--host".to_string(),
+        "--output".to_string(),
+        "json".to_string(),
+    ];
+    if let Some(name) = audio_app.and_then(sanitize_app_name) {
+        args.push(format!("--app={name}"));
+        args.push("--strict-audio".to_string());
+    }
+    args
+}
+
+/// Validate a locally-chosen audio app name before it becomes a `--app` value:
+/// trim, reject empty / overlong, and reject names carrying control characters
+/// (newlines etc.) that have no place in a real `application.name`. `None` means
+/// "no valid app selected" — the caller then shares the whole desktop audio.
+pub fn sanitize_app_name(name: &str) -> Option<String> {
+    let name = name.trim();
+    let ok = !name.is_empty()
+        && name.len() <= MAX_APP_NAME_LEN
+        && !name.chars().any(|c| c.is_control());
+    ok.then(|| name.to_string())
+}
+
+/// Hard cap on how long enumeration waits for `pactl`. It runs inline on the core
+/// command loop (the picker awaits it before opening), so a wedged/slow `pactl`
+/// must not stall mute/deafen/leave/stop. On timeout we treat it like any other
+/// failure: empty list → "All system audio" only.
+const LIST_APPS_TIMEOUT: Duration = Duration::from_secs(2);
+
+/// Enumerate the apps currently sending audio to a sink, deduplicated by
+/// `application.name`. Mirrors how pixelpass itself builds its interactive
+/// picker (`pactl -f json list sink-inputs`), so the names we return are exactly
+/// the ones `--app` matches against. Returns an empty list on any error (pactl
+/// missing, non-PipeWire host, nothing playing, or [`LIST_APPS_TIMEOUT`] elapsed)
+/// — a normal, handled state that leaves the picker showing only "All system
+/// audio".
+pub async fn list_audio_apps() -> Vec<String> {
+    let run = Command::new("pactl")
+        .args(["-f", "json", "list", "sink-inputs"])
+        .stdin(Stdio::null())
+        .stdout(Stdio::piped())
+        .stderr(Stdio::null())
+        // On [`LIST_APPS_TIMEOUT`] the `output()` future is dropped, which drops
+        // the child — `kill_on_drop(true)` then SIGKILLs and reaps it so a wedged
+        // `pactl` can't linger/accumulate across picker opens (audit P3).
+        .kill_on_drop(true)
+        .output();
+    match tokio::time::timeout(LIST_APPS_TIMEOUT, run).await {
+        Ok(Ok(o)) if o.status.success() => parse_audio_apps(&o.stdout),
+        _ => Vec::new(),
+    }
+}
+
+/// Hard cap on the capability probe (`pixelpass --help`). Conservative: a slow or
+/// hung pixelpass degrades to "strict audio unsupported" → whole-desktop-only
+/// picker (safe), never a stalled core loop.
+const HELP_PROBE_TIMEOUT: Duration = Duration::from_secs(2);
+
+/// Whether the resolved pixelpass understands `--strict-audio` (added in pixelpass
+/// `85fdebe`). peerspeak only offers per-app audio capture when it does: a per-app
+/// share always appends `--strict-audio`, and an **older** pixelpass would have
+/// clap reject the unknown flag → the host spawn hard-fails and the share is
+/// broken (audit P2, version skew). When unsupported the picker degrades to
+/// whole-desktop audio only — we never silently drop to best-effort `--app`, which
+/// would reintroduce the call echo (A23).
+///
+/// Any probe failure/timeout returns `false` (degrade to the safe path). The
+/// `--help` child is `kill_on_drop` so a hung pixelpass can't linger.
+pub async fn supports_strict_audio(bin: &Path) -> bool {
+    let run = Command::new(bin)
+        .arg("--help")
+        .stdin(Stdio::null())
+        .stdout(Stdio::piped())
+        .stderr(Stdio::null())
+        .kill_on_drop(true)
+        .output();
+    match tokio::time::timeout(HELP_PROBE_TIMEOUT, run).await {
+        Ok(Ok(o)) => help_mentions_strict_audio(&o.stdout),
+        _ => false,
+    }
+}
+
+/// Pure check: does `pixelpass --help` advertise `--strict-audio`? Matches the
+/// flag token rather than a whole line, since clap may wrap/realign help text.
+pub fn help_mentions_strict_audio(help_stdout: &[u8]) -> bool {
+    String::from_utf8_lossy(help_stdout).contains("--strict-audio")
+}
+
+/// Parse `pactl -f json list sink-inputs` stdout into a sorted, deduplicated list
+/// of `application.name`s. Pure: no I/O. Unparseable input yields an empty list.
+/// Each name is passed through [`sanitize_app_name`] so the picker only ever
+/// offers names that will actually survive [`host_args`]; otherwise a name that
+/// parses here but fails sanitization later would be selectable yet silently
+/// drop the `--app` flag and revert the share to whole-desktop audio (A23 echo).
+pub fn parse_audio_apps(stdout: &[u8]) -> Vec<String> {
+    let Ok(entries) = serde_json::from_slice::<Vec<SinkInput>>(stdout) else {
+        return Vec::new();
+    };
+    let mut names: Vec<String> = entries
+        .into_iter()
+        .filter_map(|e| e.properties.application_name)
+        .filter_map(|n| sanitize_app_name(&n))
+        .collect();
+    names.sort_unstable();
+    names.dedup();
+    names
+}
+
+#[derive(serde::Deserialize)]
+struct SinkInput {
+    properties: SinkInputProperties,
+}
+
+#[derive(serde::Deserialize)]
+struct SinkInputProperties {
+    #[serde(rename = "application.name")]
+    application_name: Option<String>,
+}
+
 /// Build the argv for a pixelpass *viewer*. The `ticket` is peer-supplied (it
 /// rides gossip presence, which is untrusted and spoofable), so flags come first
 /// and the ticket is passed as a positional **after a `--` end-of-options
@@ -162,20 +315,30 @@ pub fn is_available(config_override: Option<&str>) -> bool {
     pixelpass_path(config_override).is_some()
 }
 
-/// Spawn a pixelpass host (`pixelpass --host --output json`), wait for its
-/// startup ticket, and return the live child plus the ticket. The child keeps
+/// Spawn a pixelpass host (`pixelpass --host --output json [--app=<name>]`), wait
+/// for its startup ticket, and return the live child plus the ticket. When
+/// `audio_app` is `Some`, pixelpass captures only that app's audio instead of the
+/// whole desktop sink, which avoids the call-loopback echo (A23). The child keeps
 /// running (streaming to viewers) until killed or dropped; remaining stdout is
 /// drained in a background task so a full pipe can't stall the host. We do
 /// **not** pass `--max-viewers`: pixelpass bandwidth-measures its own safe cap,
 /// protecting the sharer's uplink, and refuses extras with `viewer_refused`.
-pub async fn spawn_host(bin: &Path) -> std::io::Result<(Child, String)> {
+pub async fn spawn_host(
+    bin: &Path,
+    audio_app: Option<&str>,
+    notices: Option<tokio::sync::mpsc::UnboundedSender<PixelpassEvent>>,
+) -> std::io::Result<(Child, String)> {
     let mut child = Command::new(bin)
-        .arg("--host")
-        .arg("--output")
-        .arg("json")
+        .args(host_args(audio_app))
         .stdin(Stdio::null())
         .stdout(Stdio::piped())
-        .stderr(Stdio::null())
+        // Capture stderr (not null): pixelpass prints its startup precondition
+        // failures there — a missing GStreamer plugin / `pactl`, each with an
+        // actionable "Install hint: sudo apt install ..." line. If the host dies
+        // before its ticket we fold that tail into our error so the user sees
+        // *what to install* instead of a dead-end "exited before a ticket". On
+        // the success path we drain it in the background so the pipe can't fill.
+        .stderr(Stdio::piped())
         .kill_on_drop(true)
         .spawn()?;
 
@@ -183,6 +346,7 @@ pub async fn spawn_host(bin: &Path) -> std::io::Result<(Child, String)> {
         .stdout
         .take()
         .ok_or_else(|| std::io::Error::other("pixelpass host stdout missing"))?;
+    let stderr = child.stderr.take();
     let mut lines = BufReader::new(stdout).lines();
 
     let ticket = match read_until(&mut lines, |e| match e {
@@ -194,9 +358,10 @@ pub async fn spawn_host(bin: &Path) -> std::io::Result<(Child, String)> {
         Ok(Some(t)) => t,
         Ok(None) => {
             let _ = child.kill().await;
-            return Err(std::io::Error::other(
-                "pixelpass host exited before emitting a ticket",
-            ));
+            let detail = read_stderr_tail(stderr).await;
+            return Err(std::io::Error::other(format!(
+                "pixelpass host exited before emitting a ticket{detail}"
+            )));
         }
         Err(e) => {
             let _ = child.kill().await;
@@ -204,10 +369,65 @@ pub async fn spawn_host(bin: &Path) -> std::io::Result<(Child, String)> {
         }
     };
 
-    drain_in_background(lines, "host");
+    if let Some(stderr) = stderr {
+        drain_stderr_in_background(stderr);
+    }
+    drain_in_background(lines, "host", notices);
     Ok((child, ticket))
 }
 
+/// Read a killed pixelpass child's stderr to EOF and reduce it to a short,
+/// user-facing diagnostic tail via [`pixelpass_failure_detail`]. Bounded: the
+/// caller kills the child first, so the pipe EOFs promptly. Returns an empty
+/// string when stderr was already taken or carried nothing useful.
+async fn read_stderr_tail(stderr: Option<tokio::process::ChildStderr>) -> String {
+    use tokio::io::AsyncReadExt;
+    let Some(mut stderr) = stderr else {
+        return String::new();
+    };
+    let mut buf = Vec::new();
+    let _ = stderr.read_to_end(&mut buf).await;
+    pixelpass_failure_detail(&String::from_utf8_lossy(&buf))
+}
+
+/// Discard a running pixelpass child's stderr in the background so its pipe
+/// can't fill and stall the host (mirrors [`drain_in_background`] for stdout).
+fn drain_stderr_in_background(mut stderr: tokio::process::ChildStderr) {
+    use tokio::io::AsyncReadExt;
+    tokio::spawn(async move {
+        let mut buf = [0u8; 4096];
+        while let Ok(n) = stderr.read(&mut buf).await {
+            if n == 0 {
+                break;
+            }
+        }
+    });
+}
+
+/// Extract a human-useful tail from a failed pixelpass child's stderr to append
+/// to our error. pixelpass writes actionable startup errors there (a missing
+/// GStreamer element / `pactl` plus an `Install hint: sudo apt install ...`
+/// line), which is exactly what a freshly-installed host needs to see. The
+/// decorative host banner (box-drawing) is dropped — it only prints on the
+/// success path, but we filter it defensively. Pure: no I/O. Returns an empty
+/// string when there's nothing worth surfacing (so callers can append blindly).
+pub fn pixelpass_failure_detail(stderr: &str) -> String {
+    let useful: Vec<&str> = stderr
+        .lines()
+        .map(str::trim_end)
+        .filter(|l| !l.trim().is_empty())
+        .filter(|l| !l.trim_start().starts_with(['│', '┌', '└', '├']))
+        .collect();
+    if useful.is_empty() {
+        return String::new();
+    }
+    // The anyhow error and its install hint are the *last* lines printed, so
+    // keep the tail rather than the head.
+    const MAX_LINES: usize = 12;
+    let start = useful.len().saturating_sub(MAX_LINES);
+    format!("\n\npixelpass reported:\n{}", useful[start..].join("\n"))
+}
+
 /// Spawn a pixelpass viewer for `ticket`, wait for it to connect, and open the
 /// stream in a local player (mpv, falling back to vlc). Returns the live viewer
 /// child so the caller can kill it on room-leave; it also self-exits when the
@@ -251,7 +471,7 @@ pub async fn spawn_viewer(bin: &Path, ticket: &str) -> std::io::Result<Child> {
         return Err(e);
     }
 
-    drain_in_background(lines, "viewer");
+    drain_in_background(lines, "viewer", None);
     Ok(child)
 }
 
@@ -286,15 +506,24 @@ where
 }
 
 /// Keep reading the child's stdout to EOF in the background so a full pipe can't
-/// stall it; log notable events for diagnostics.
-fn drain_in_background<R>(mut lines: tokio::io::Lines<BufReader<R>>, role: &'static str)
-where
+/// stall it; log notable events for diagnostics. When `notices` is `Some`, each
+/// parsed event is also forwarded to the caller (the core, which translates the
+/// `app_audio` ones into a UI warning); a send failure (receiver dropped) just
+/// stops forwarding, draining continues. The task ends on EOF (child exited).
+fn drain_in_background<R>(
+    mut lines: tokio::io::Lines<BufReader<R>>,
+    role: &'static str,
+    notices: Option<tokio::sync::mpsc::UnboundedSender<PixelpassEvent>>,
+) where
     R: tokio::io::AsyncRead + Unpin + Send + 'static,
 {
     tokio::spawn(async move {
         while let Ok(Some(line)) = lines.next_line().await {
             if let Some(ev) = parse_pixelpass_event(&line) {
                 crate::log_msg(&format!("pixelpass {role}: {}", event_for_log(&ev)));
+                if let Some(tx) = &notices {
+                    let _ = tx.send(ev);
+                }
             }
         }
     });
@@ -313,6 +542,8 @@ fn event_for_log(ev: &PixelpassEvent) -> String {
         PixelpassEvent::Refused(reason) => format!("viewer_refused reason={reason:?}"),
         PixelpassEvent::CaptureStarted => "capture_started".to_string(),
         PixelpassEvent::CaptureStopped => "capture_stopped".to_string(),
+        PixelpassEvent::AppAudioRouted => "app_audio_routed".to_string(),
+        PixelpassEvent::AppAudioLost => "app_audio_lost".to_string(),
         PixelpassEvent::Other => "other".to_string(),
     }
 }
@@ -384,6 +615,142 @@ mod tests {
         );
     }
 
+    #[test]
+    fn host_args_without_app_shares_whole_desktop() {
+        // No app selected → no --app flag → pixelpass keeps its default
+        // (whole-desktop) audio capture.
+        assert_eq!(host_args(None), vec!["--host", "--output", "json"]);
+    }
+
+    #[test]
+    fn host_args_with_app_appends_single_token_flag() {
+        // The chosen app rides in the `--app=<name>` single-token form so a
+        // name beginning with `-` can never be reparsed as a flag (A23), plus
+        // `--strict-audio` so pixelpass never falls back to whole-desktop audio.
+        assert_eq!(
+            host_args(Some("Firefox")),
+            vec!["--host", "--output", "json", "--app=Firefox", "--strict-audio"]
+        );
+        // The hyphen-leading name is still bound to --app as a single token;
+        // --strict-audio is the trailing flag.
+        let args = host_args(Some("-rm -rf"));
+        assert_eq!(args[3], "--app=-rm -rf");
+        assert_eq!(args[4], "--strict-audio");
+    }
+
+    #[test]
+    fn host_args_blank_or_control_app_is_dropped() {
+        // An empty / whitespace / control-laden selection is sanitized away,
+        // falling back to whole-desktop capture rather than a broken flag.
+        assert_eq!(host_args(Some("   ")), vec!["--host", "--output", "json"]);
+        assert_eq!(host_args(Some("bad\nname")), vec!["--host", "--output", "json"]);
+    }
+
+    #[test]
+    fn sanitize_app_name_trims_and_rejects_garbage() {
+        assert_eq!(sanitize_app_name("  Firefox \n"), Some("Firefox".to_string()));
+        assert_eq!(sanitize_app_name(""), None);
+        assert_eq!(sanitize_app_name("   "), None);
+        assert_eq!(sanitize_app_name("a\tb"), None);
+        assert_eq!(sanitize_app_name(&"x".repeat(MAX_APP_NAME_LEN + 1)), None);
+    }
+
+    #[test]
+    fn parse_audio_apps_dedups_and_sorts_by_application_name() {
+        let stdout = br#"[
+            {"index":1,"properties":{"application.name":"Firefox"}},
+            {"index":2,"properties":{"application.name":"mpv"}},
+            {"index":3,"properties":{"application.name":"Firefox"}},
+            {"index":4,"properties":{"application.name":"  Spotify  "}},
+            {"index":5,"properties":{"application.name":""}},
+            {"index":6,"properties":{"other":"no name here"}}
+        ]"#;
+        assert_eq!(
+            parse_audio_apps(stdout),
+            vec!["Firefox".to_string(), "Spotify".to_string(), "mpv".to_string()]
+        );
+    }
+
+    #[test]
+    fn parse_audio_apps_empty_or_garbage_is_empty() {
+        assert_eq!(parse_audio_apps(b""), Vec::<String>::new());
+        assert_eq!(parse_audio_apps(b"not json"), Vec::<String>::new());
+        assert_eq!(parse_audio_apps(b"[]"), Vec::<String>::new());
+    }
+
+    #[test]
+    fn parse_audio_apps_drops_names_host_args_would_reject() {
+        // Names that parse from pactl but fail `sanitize_app_name` (control chars,
+        // overlong) must NOT be offered in the picker — otherwise the user could
+        // pick one, `host_args` would silently drop `--app`, and the share would
+        // revert to whole-desktop audio (A23 echo) with no signal. The valid name
+        // survives; the control-char and overlong ones are filtered out.
+        let overlong = "x".repeat(MAX_APP_NAME_LEN + 1);
+        let stdout = format!(
+            r#"[
+                {{"index":1,"properties":{{"application.name":"mpv"}}}},
+                {{"index":2,"properties":{{"application.name":"bad\nname"}}}},
+                {{"index":3,"properties":{{"application.name":"{overlong}"}}}}
+            ]"#
+        );
+        assert_eq!(parse_audio_apps(stdout.as_bytes()), vec!["mpv".to_string()]);
+    }
+
+    #[test]
+    fn failure_detail_surfaces_install_hint_and_drops_banner() {
+        // The real shape of a fresh-host failure: anyhow error + install hint on
+        // stderr. We must keep those (so the user knows what to apt install) and
+        // drop the decorative banner box-drawing lines.
+        let stderr = "\
+┌─ PixelPass · host ─────────────────────────────────────────
+│  display server : Wayland
+└────────────────────────────────────────────────────────────
+Error: GStreamer element `vah264enc` not available.
+Install hint: sudo apt install gstreamer1.0-plugins-bad
+";
+        let detail = pixelpass_failure_detail(stderr);
+        assert!(detail.starts_with("\n\npixelpass reported:\n"));
+        assert!(detail.contains("vah264enc` not available"));
+        assert!(detail.contains("sudo apt install gstreamer1.0-plugins-bad"));
+        assert!(!detail.contains('│'), "banner box-drawing must be dropped");
+        assert!(!detail.contains('┌'));
+    }
+
+    #[test]
+    fn failure_detail_empty_when_nothing_useful() {
+        // Blank / banner-only stderr yields an empty string so the caller can
+        // append it to the base message unconditionally without trailing noise.
+        assert_eq!(pixelpass_failure_detail(""), "");
+        assert_eq!(pixelpass_failure_detail("   \n  \n"), "");
+        assert_eq!(
+            pixelpass_failure_detail("│  display server : Wayland\n│  capture : x\n"),
+            ""
+        );
+    }
+
+    #[test]
+    fn failure_detail_keeps_only_the_tail() {
+        // A long stderr is truncated to its last lines (where the real error
+        // and hint live), not its head.
+        let body: String = (0..30).map(|i| format!("line {i}\n")).collect();
+        let detail = pixelpass_failure_detail(&body);
+        assert!(detail.contains("line 29"));
+        assert!(!detail.contains("line 0\n"));
+    }
+
+    #[test]
+    fn help_probe_detects_strict_audio_flag() {
+        // A new pixelpass advertises the flag; an old one doesn't. The probe must
+        // match the token even when clap wraps the option onto its own line.
+        let new_help = b"Options:\n      --app <APP>\n      --strict-audio\n          With --app, never fall back...";
+        assert!(help_mentions_strict_audio(new_help));
+        let old_help = b"Options:\n      --app <APP>\n      --output <OUTPUT>\n  -h, --help";
+        assert!(!help_mentions_strict_audio(old_help));
+        // Garbage / empty output degrades to "unsupported" (safe path).
+        assert!(!help_mentions_strict_audio(b""));
+        assert!(!help_mentions_strict_audio(&[0xff, 0xfe, 0x00]));
+    }
+
     #[test]
     fn sanitize_ticket_accepts_pixelpass_endpoint_ticket_shape() {
         let ticket = "endpointaabwxjexzensznfvuudiapn5tyzws3angd2merarm";
@@ -470,6 +837,29 @@ mod tests {
         );
     }
 
+    #[test]
+    fn parses_app_audio_states() {
+        // The wire contract from pixelpass's --strict-audio run (A23): routed =
+        // the chosen app's audio is live; lost = it stopped (viewers now silent).
+        assert_eq!(
+            parse_pixelpass_event(r#"{"event":"app_audio","state":"routed"}"#),
+            Some(PixelpassEvent::AppAudioRouted)
+        );
+        assert_eq!(
+            parse_pixelpass_event(r#"{"event":"app_audio","state":"lost"}"#),
+            Some(PixelpassEvent::AppAudioLost)
+        );
+        // Unknown / missing state is recognized-but-unused, not a parse failure.
+        assert_eq!(
+            parse_pixelpass_event(r#"{"event":"app_audio","state":"weird"}"#),
+            Some(PixelpassEvent::Other)
+        );
+        assert_eq!(
+            parse_pixelpass_event(r#"{"event":"app_audio"}"#),
+            Some(PixelpassEvent::Other)
+        );
+    }
+
     #[test]
     fn recognized_but_unused_event_is_other() {
         assert_eq!(