houndd: replace the clamscan fork with yara-x in process

The old engine shelled out to clamscan for every scan, and clamscan
reloads a 169 MB signature database on every invocation. Measured on a
68-byte EICAR file: 6.5 seconds and ~1.5 GB RSS — paid once per file,
and realtime.rs called it once per inotify event.

Replaces it with HoundEngine: yara-x compiled once at daemon start,
held in memory, one scanner reused across a whole walk, plus a verdict
cache keyed on (dev, ino, mtime, size) so an unchanged file that has
been seen before never reaches the matcher.

Measured after, same machine, same EICAR file:

  single file      6.5 s  ->  4 ms
  400 files cold      --  ->  9 ms
  400 files warm      --  ->  5 ms

Also here:

- rules.rs: hot-swappable rule store. Built-in pack is embedded so a
  fresh install detects something before it has ever reached the
  network; on-disk packs load from $HOUNDD_RULES_DIR, /var/lib/hound
  or the XDG data dir. Reload swaps an Arc, so in-flight scans are
  never torn out from under.
- cache.rs: bounded FIFO verdict cache. Any of the four key fields
  changing means rescan, so edits, truncates and replace-by-rename all
  correctly miss.
- The goodware gate: every rule is scanned against all of /usr/bin,
  /bin and /usr/sbin in CI, and a single hit fails the build. It has
  already earned its keep — it caught a reverse-shell rule that matched
  /usr/bin/sudo, which is now removed rather than tuned. A rule that
  quarantines sudo is worse than no rule at all.
- ScanEngine is Send + Sync and selection stays per-call, so
  HOUNDD_ENGINE=clamav still reaches the legacy path for comparison.
- ScanResult.skipped reports files passed over for size instead of
  quietly counting them as clean.
- Settings gain theme (auto/light/dark), tray_icon_style (color/mono),
  close_to_tray and confirm_quit, normalised daemon-side because
  clients are not trusted to send a theme we can render.

57 tests pass, up from 29.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Cargo.toml

@@ -16,6 +16,7 @@ clap = { version = "4", features = ["derive"] }
 colored = "2"
 time = { version = "0.3", features = ["serde", "std", "formatting"] }
 inotify = "0.10"
+yara-x = "1.19"
 
 [profile.release]
 lto = true

crates/hound-api/src/lib.rs

@@ -109,6 +109,11 @@ pub struct ScanResult {
     pub scanned: u64,
     pub clean: u64,
     pub infected: u64,
+    /// Files the engine deliberately did not read — over the size cap.
+    /// Reported separately because counting them as scanned would be a
+    /// lie and dropping them silently would be worse.
+    #[serde(default)]
+    pub skipped: u64,
     pub found: Vec<Found>,
 }
 
@@ -166,11 +171,67 @@ pub struct Settings {
     /// Run the signature update automatically (daemon schedules it).
     pub auto_update_signatures: bool,
 
+    // Appearance (GUI-only; the daemon stores them so the setting follows
+    // the machine rather than a per-user GUI config file, and so the CLI
+    // can read and set them too).
+    /// Window theme: "auto" (follow the desktop) | "light" | "dark".
+    #[serde(default = "default_theme")]
+    pub theme: String,
+    /// Tray icon treatment. Colour uses the 4-state ladder; monochrome
+    /// renders a single-tone glyph that follows the panel's own theme,
+    /// which is what most desktop styleguides actually ask for.
+    #[serde(default = "default_tray_style")]
+    pub tray_icon_style: String,
+
+    // Window behaviour
+    /// Closing the window hides it to the tray instead of exiting. Quitting
+    /// is only ever possible from the tray menu, and is confirmed there —
+    /// an antivirus that can be shut off by a stray click on the X is not
+    /// protecting anything.
+    #[serde(default = "default_true")]
+    pub close_to_tray: bool,
+    /// Whether the tray's Quit entry must be confirmed before it exits.
+    #[serde(default = "default_true")]
+    pub confirm_quit: bool,
+
     // Global
     /// Master switch — when true, realtime is suspended and the tray is gray.
     pub paused: bool,
 }
 
+fn default_theme() -> String {
+    "auto".into()
+}
+
+fn default_tray_style() -> String {
+    "color".into()
+}
+
+fn default_true() -> bool {
+    true
+}
+
+impl Settings {
+    /// Accepted values for [`Settings::theme`].
+    pub const THEMES: [&'static str; 3] = ["auto", "light", "dark"];
+    /// Accepted values for [`Settings::tray_icon_style`].
+    pub const TRAY_STYLES: [&'static str; 2] = ["color", "mono"];
+
+    /// Clamp free-text enum fields back to something the GUI can render.
+    ///
+    /// These arrive over a JSON socket from clients we do not control, so
+    /// an unknown value is normalised rather than trusted — a typo'd theme
+    /// must not leave the window unstyled.
+    pub fn normalise_appearance(&mut self) {
+        if !Self::THEMES.contains(&self.theme.as_str()) {
+            self.theme = default_theme();
+        }
+        if !Self::TRAY_STYLES.contains(&self.tray_icon_style.as_str()) {
+            self.tray_icon_style = default_tray_style();
+        }
+    }
+}
+
 impl Default for Settings {
     fn default() -> Self {
         Self {
@@ -180,6 +241,10 @@ impl Default for Settings {
             realtime_enabled: true,
             realtime_watch: vec!["~/Downloads".into()],
             on_detect: "quarantine".into(),
+            theme: default_theme(),
+            tray_icon_style: default_tray_style(),
+            close_to_tray: true,
+            confirm_quit: true,
             ransomware_guard: true,
             ransomware_threshold_per_min: 40,
             rootkit_enabled: true,

crates/houndd/Cargo.toml

@@ -17,3 +17,4 @@ serde.workspace = true
 serde_json.workspace = true
 time.workspace = true
 inotify.workspace = true
+yara-x.workspace = true

crates/houndd/rules/hound-builtin.yar

@@ -0,0 +1,97 @@
+/*
+ * Hound built-in starter pack.
+ *
+ * Deliberately tiny and deliberately tight. Every rule here requires
+ * several independent strings before it fires, because a false positive
+ * in an antivirus is worse than a miss — one rule that quarantines a
+ * system binary ends the product.
+ *
+ * The real corpus lands in Phase 3 (the signed Hound Linux pack, gated
+ * behind the goodware CI regression suite). This pack exists so a fresh
+ * install detects *something* before it has ever contacted the network.
+ */
+
+rule EICAR_Test_File
+{
+    meta:
+        name     = "EICAR-Test-Signature"
+        severity = "info"
+        desc     = "Industry-standard antivirus test file. Harmless."
+    strings:
+        $eicar = "X5O!P%@AP[4\\PZX54(P^)7CC)7}$EICAR-STANDARD-ANTIVIRUS-TEST-FILE!$H+H*"
+    condition:
+        $eicar
+}
+
+rule Linux_Coinminer_XMRig
+{
+    meta:
+        name     = "Linux.Coinminer.XMRig"
+        severity = "critical"
+        desc     = "XMRig cryptocurrency miner. Requires pool protocol plus two config keys."
+    strings:
+        $pool1 = "stratum+tcp://" ascii
+        $pool2 = "stratum+ssl://" ascii
+        $cfg1  = "donate-level" ascii
+        $cfg2  = "rig-id" ascii
+        $cfg3  = "randomx" ascii nocase
+        $name  = "xmrig" ascii nocase
+    condition:
+        ($pool1 or $pool2) and 2 of ($cfg*) and $name
+}
+
+rule Linux_Webshell_PHP_Eval
+{
+    meta:
+        name     = "Linux.Webshell.PHP-Eval"
+        severity = "critical"
+        desc     = "PHP webshell: request-driven eval of decoded input."
+    strings:
+        $php   = "<?php"
+        $eval1 = /eval\s*\(\s*(base64_decode|gzinflate|str_rot13|gzuncompress)\s*\(/
+        $src1  = "$_POST"
+        $src2  = "$_GET"
+        $src3  = "$_REQUEST"
+        $src4  = "$_COOKIE"
+    condition:
+        $php and $eval1 and 1 of ($src*)
+}
+
+rule Linux_Rootkit_Preload
+{
+    meta:
+        name     = "Linux.Rootkit.Preload"
+        severity = "critical"
+        desc     = "LD_PRELOAD userland rootkit: hooks libc lookup calls and hides itself."
+    strings:
+        $dlsym   = "dlsym" ascii
+        $libc    = "RTLD_NEXT" ascii
+        $hook1   = "readdir64" ascii
+        $hook2   = "readdir" ascii
+        $hook3   = "lxstat" ascii
+        $hook4   = "fopen" ascii
+        $hide1   = "ld.so.preload" ascii
+        $hide2   = "/proc/net/tcp" ascii
+    condition:
+        uint32(0) == 0x464c457f       // ELF magic
+        and $dlsym and $libc
+        and 2 of ($hook*)
+        and 1 of ($hide*)
+}
+
+/*
+ * REMOVED: Linux_Backdoor_ReverseShell_ELF
+ *
+ * It required an ELF containing "/bin/sh" plus four of
+ * {dup2, socket, connect, inet_addr, execve}. That is a perfect
+ * description of a reverse shell and also a perfect description of
+ * /usr/bin/sudo, which the goodware test caught immediately. Any
+ * dynamically linked network-capable binary imports those symbols
+ * legitimately, so no threshold tweak saves this rule — it would only
+ * move the false positive to a different binary on a different distro.
+ *
+ * Catching reverse shells properly needs either ELF structure (statically
+ * linked, tiny, no libc) or the behaviour itself, which is Phase 7's job.
+ * Left out rather than shipped loose: a rule that quarantines sudo is
+ * worse than no rule at all.
+ */

crates/houndd/src/cache.rs

@@ -0,0 +1,209 @@
+//! The verdict cache.
+//!
+//! Scanning the same unchanged bytes twice is pure waste, and on a
+//! developer machine it is nearly all of the work: a `cargo build` opens
+//! the same crate sources hundreds of times, and Phase 1's execution gate
+//! will ask for a verdict on every `execve` of every binary on the box.
+//!
+//! The key is `(dev, ino, mtime, size)`. If any of those four change the
+//! file is treated as new, so an edit, a truncate, a replace-by-rename or
+//! a move across filesystems all correctly miss the cache. Content is
+//! never hashed — hashing to avoid reading would mean reading.
+//!
+//! Eviction is FIFO with a hard capacity. An LRU would hold a slightly
+//! better working set, but FIFO costs one `VecDeque` push and cannot
+//! degrade pathologically, and the cost of a miss here is one scan.
+
+use std::collections::{HashMap, VecDeque};
+use std::fs::Metadata;
+use std::os::unix::fs::MetadataExt;
+use std::sync::{Arc, Mutex};
+
+/// Identity of a file *version*. Any field changing means rescan.
+#[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
+pub struct FileKey {
+    dev: u64,
+    ino: u64,
+    mtime: i64,
+    mtime_nsec: i64,
+    size: u64,
+}
+
+impl FileKey {
+    pub fn from_metadata(md: &Metadata) -> Self {
+        Self {
+            dev: md.dev(),
+            ino: md.ino(),
+            mtime: md.mtime(),
+            mtime_nsec: md.mtime_nsec(),
+            size: md.size(),
+        }
+    }
+}
+
+/// What a previous scan concluded. `None` is clean.
+pub type Verdict = Option<Arc<str>>;
+
+struct Inner {
+    map: HashMap<FileKey, Verdict>,
+    order: VecDeque<FileKey>,
+    hits: u64,
+    misses: u64,
+}
+
+/// A bounded, thread-safe cache of scan verdicts.
+#[derive(Clone)]
+pub struct VerdictCache {
+    inner: Arc<Mutex<Inner>>,
+    capacity: usize,
+}
+
+impl VerdictCache {
+    pub fn new(capacity: usize) -> Self {
+        Self {
+            inner: Arc::new(Mutex::new(Inner {
+                map: HashMap::with_capacity(capacity.min(4096)),
+                order: VecDeque::with_capacity(capacity.min(4096)),
+                hits: 0,
+                misses: 0,
+            })),
+            capacity: capacity.max(1),
+        }
+    }
+
+    /// Look up a verdict. The outer `Option` is cache presence; the inner
+    /// one is the verdict itself, so a cached-clean answer is
+    /// `Some(None)` and is deliberately distinguishable from a miss.
+    pub fn get(&self, key: &FileKey) -> Option<Verdict> {
+        let mut inner = self.inner.lock().expect("verdict cache poisoned");
+        match inner.map.get(key) {
+            Some(v) => {
+                let v = v.clone();
+                inner.hits += 1;
+                Some(v)
+            }
+            None => {
+                inner.misses += 1;
+                None
+            }
+        }
+    }
+
+    /// Record a verdict, evicting the oldest entry when full.
+    pub fn put(&self, key: FileKey, verdict: Verdict) {
+        let mut inner = self.inner.lock().expect("verdict cache poisoned");
+        if inner.map.insert(key, verdict).is_none() {
+            inner.order.push_back(key);
+            while inner.order.len() > self.capacity {
+                if let Some(old) = inner.order.pop_front() {
+                    inner.map.remove(&old);
+                }
+            }
+        }
+    }
+
+    /// Drop everything. Called whenever the ruleset changes — a verdict
+    /// reached under the old rules says nothing about the new ones.
+    pub fn clear(&self) {
+        let mut inner = self.inner.lock().expect("verdict cache poisoned");
+        inner.map.clear();
+        inner.order.clear();
+    }
+
+    /// `(entries, hits, misses)` for `hound status` and the benchmarks.
+    pub fn stats(&self) -> (usize, u64, u64) {
+        let inner = self.inner.lock().expect("verdict cache poisoned");
+        (inner.map.len(), inner.hits, inner.misses)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn key(ino: u64, size: u64) -> FileKey {
+        FileKey {
+            dev: 1,
+            ino,
+            mtime: 100,
+            mtime_nsec: 0,
+            size,
+        }
+    }
+
+    #[test]
+    fn miss_then_hit() {
+        let c = VerdictCache::new(8);
+        assert!(c.get(&key(1, 10)).is_none(), "cold lookup must miss");
+        c.put(key(1, 10), None);
+        assert_eq!(c.get(&key(1, 10)), Some(None), "cached clean is a hit");
+    }
+
+    #[test]
+    fn cached_infected_round_trips() {
+        let c = VerdictCache::new(8);
+        c.put(key(2, 20), Some("Linux.Coinminer.XMRig".into()));
+        let got = c.get(&key(2, 20)).expect("should hit");
+        assert_eq!(got.as_deref(), Some("Linux.Coinminer.XMRig"));
+    }
+
+    #[test]
+    fn any_field_change_misses() {
+        let c = VerdictCache::new(8);
+        c.put(key(3, 30), None);
+        // Same inode, different size — the file was rewritten.
+        assert!(c.get(&key(3, 31)).is_none());
+        // Same size, different inode — replaced by rename.
+        assert!(c.get(&key(4, 30)).is_none());
+    }
+
+    #[test]
+    fn mtime_change_misses() {
+        let c = VerdictCache::new(8);
+        let mut k = key(5, 50);
+        c.put(k, None);
+        k.mtime_nsec = 1;
+        assert!(c.get(&k).is_none(), "a nanosecond of edit is still an edit");
+    }
+
+    #[test]
+    fn evicts_fifo_at_capacity() {
+        let c = VerdictCache::new(2);
+        c.put(key(1, 1), None);
+        c.put(key(2, 2), None);
+        c.put(key(3, 3), None);
+        assert!(c.get(&key(1, 1)).is_none(), "oldest should be evicted");
+        assert!(c.get(&key(3, 3)).is_some(), "newest should be resident");
+        let (entries, _, _) = c.stats();
+        assert_eq!(entries, 2, "capacity must be honoured");
+    }
+
+    #[test]
+    fn reinsert_does_not_grow_order_queue() {
+        let c = VerdictCache::new(4);
+        for _ in 0..50 {
+            c.put(key(9, 9), None);
+        }
+        let (entries, _, _) = c.stats();
+        assert_eq!(entries, 1);
+    }
+
+    #[test]
+    fn clear_empties() {
+        let c = VerdictCache::new(4);
+        c.put(key(1, 1), None);
+        c.clear();
+        assert!(c.get(&key(1, 1)).is_none());
+        assert_eq!(c.stats().0, 0);
+    }
+
+    #[test]
+    fn stats_count_hits_and_misses() {
+        let c = VerdictCache::new(4);
+        c.get(&key(1, 1)); // miss
+        c.put(key(1, 1), None);
+        c.get(&key(1, 1)); // hit
+        let (_, hits, misses) = c.stats();
+        assert_eq!((hits, misses), (1, 1));
+    }
+}

crates/houndd/src/engine.rs

@@ -16,9 +16,14 @@ use anyhow::{Context, Result};
 use hound_api::{DbFile, ScanResult};
 use std::fs;
 use std::process::Command;
+use std::sync::OnceLock;
 
 /// What an engine implementation must answer.
-pub trait ScanEngine {
+///
+/// `Send + Sync` because the daemon holds exactly one engine for its
+/// whole life and hands it to every connection thread and to the
+/// real-time monitor.
+pub trait ScanEngine: Send + Sync {
     /// Stable id for the wire (`Status.engine`): "clamav" today, e.g.
     /// "hound-native" when the Rust engine ships.
     fn name(&self) -> &'static str;
@@ -224,6 +229,8 @@ pub fn parse_clamscan(stdout: &[u8], exit_code: i32) -> Result<ScanResult> {
         scanned,
         clean,
         infected,
+        // clamscan does not tell us what it skipped for size.
+        skipped: 0,
         found,
     })
 }
@@ -240,23 +247,52 @@ pub fn to_rfc3339(t: std::time::SystemTime) -> String {
     dt.format(&Rfc3339).unwrap_or_else(|_| "unknown".into())
 }
 
-/// The active engine, chosen at daemon startup.
+/// The active engine, chosen once at daemon startup.
 ///
-/// Default is [`ClamAvEngine`]. Set `HOUNDD_ENGINE=fake` to the
+/// Default is [`HoundEngine`](crate::native::HoundEngine) — yara-x in
-/// [`FakeEngine`] — used by the E2E test so it can drive a full
+/// process. `HOUNDD_ENGINE` overrides it:
-/// daemon lifecycle (status, scan, settings, quarantine, rootkit)
+///
-/// without requiring ClamAV or a real filesystem of .cld files.
+/// * `clamav` — the legacy `clamscan` subprocess path. Kept so the two
+///   can be compared directly, and because it still owns the Windows
+///   malware corpus that our own rules deliberately do not cover.
+/// * `fake`   — the synthetic engine the E2E test drives, so a full
+///   daemon lifecycle can run without ClamAV or a real rule pack.
 ///
 /// The whole point of the trait is that this is the only place the
 /// daemon decides *which* engine it serves.
 pub fn engine() -> &'static dyn ScanEngine {
-    if std::env::var_os("HOUNDD_ENGINE").is_some_and(|v| v == "fake") {
+    // Selection is per call — one getenv — so the choice stays live and
+    // two tests in one process cannot contaminate each other. Only the
+    // expensive engine is memoised, below.
+    match std::env::var("HOUNDD_ENGINE").as_deref() {
+        Ok("fake") => {
             static FAKE: FakeEngine = FakeEngine;
             &FAKE
-    } else {
+        }
+        Ok("clamav") => {
             static CLAMAV: ClamAvEngine = ClamAvEngine;
             &CLAMAV
         }
+        _ => native_engine(),
+    }
+}
+
+/// The native engine, built exactly once. Compiling the ruleset is the
+/// one genuinely expensive thing the daemon does at startup, so it must
+/// never happen twice.
+fn native_engine() -> &'static dyn ScanEngine {
+    static ENGINE: OnceLock<Box<dyn ScanEngine>> = OnceLock::new();
+    ENGINE
+        .get_or_init(|| match crate::native::HoundEngine::new() {
+            Ok(e) => Box::new(e) as Box<dyn ScanEngine>,
+            Err(e) => {
+                // Losing detection entirely is worse than falling back to
+                // the slow path, so say so loudly and carry on.
+                eprintln!("engine: rules failed to compile ({e}) — falling back to clamav");
+                Box::new(ClamAvEngine)
+            }
+        })
+        .as_ref()
 }
 
 /// Test engine: reports itself present, scans anything whose name
@@ -301,6 +337,7 @@ impl ScanEngine for FakeEngine {
             scanned,
             clean: scanned - infected_u,
             infected: infected_u,
+            skipped: 0,
             found,
         })
     }

crates/houndd/src/main.rs

@@ -40,11 +40,14 @@
 //! swaps in [`engine::engine`]'s fake backend (a full daemon lifecycle
 //! without ClamAV installed).
 
+mod cache;
 mod engine;
 mod events;
+mod native;
 mod quarantine;
 mod realtime;
 mod rootkit;
+mod rules;
 mod settings;
 #[cfg(test)]
 mod test_util;
@@ -205,11 +208,13 @@ fn dispatch(req: &hound_api::Request, st: &DaemonState) -> Result<Value> {
         // ── settings ──
         "settings.get" => Ok(serde_json::to_value(st.settings.get())?),
         "settings.set" => {
-            let incoming: Settings = serde_json::from_value(
+            let mut incoming: Settings = serde_json::from_value(
                 req.params
                     .clone()
                     .context("settings.set requires a params object")?,
             )?;
+            // Clients are not trusted to send a theme we can render.
+            incoming.normalise_appearance();
             st.settings
                 .set(&incoming)
                 .map_err(|e| anyhow::anyhow!("persisting settings: {e}"))?;

crates/houndd/src/native.rs

@@ -0,0 +1,462 @@
+//! Hound's own engine: yara-x, in process, with a verdict cache.
+//!
+//! This is the Phase 0 replacement for forking `clamscan`. The old path
+//! spent 6.5 seconds and ~1.5 GB of RSS on a 68-byte file because every
+//! invocation reloaded a 169 MB signature database. Here the ruleset is
+//! compiled once at daemon start, the scanner is reused across every
+//! file in a walk, and an unchanged file that has been seen before never
+//! reaches the matcher at all.
+//!
+//! Deliberate choices worth knowing about:
+//!
+//! * **Symlinks are never followed.** A directory walk that follows links
+//!   can loop, can be steered outside the requested tree by anyone who
+//!   can create a link, and re-scans the same inode repeatedly. We stat
+//!   with `symlink_metadata` and skip links entirely.
+//! * **Pseudo-filesystems are skipped.** Reading `/proc` and `/sys` is
+//!   meaningless here and reading some of their files blocks forever.
+//! * **Oversized files are counted, not read.** Reporting them as scanned
+//!   would be a lie; skipping them silently would be worse. They are
+//!   counted separately and surfaced in the summary.
+
+use anyhow::{Context, Result};
+use hound_api::{DbFile, Found, ScanResult};
+use std::path::{Path, PathBuf};
+use std::sync::Arc;
+
+use crate::cache::{FileKey, VerdictCache};
+use crate::engine::{to_rfc3339, ScanEngine};
+use crate::rules::{RuleSet, RuleStore};
+
+/// Files larger than this are skipped. Malware that matters is rarely
+/// this big, and reading disk images on every scan makes the product
+/// unusable on the developer machines we are targeting.
+const DEFAULT_MAX_FILE_BYTES: u64 = 100 * 1024 * 1024;
+
+/// How many file verdicts to remember. ~200k entries is a few tens of MB
+/// and comfortably covers a working developer tree plus the system.
+const CACHE_CAPACITY: usize = 200_000;
+
+/// Directory prefixes that are never worth walking.
+const SKIP_PREFIXES: &[&str] = &[
+    "/proc",
+    "/sys",
+    "/dev",
+    "/run",
+    "/var/lib/hound/vault",
+];
+
+pub struct HoundEngine {
+    rules: RuleStore,
+    cache: VerdictCache,
+    max_file_bytes: u64,
+}
+
+impl HoundEngine {
+    /// Compile the ruleset and build the cache. Called once, from the
+    /// engine factory, at daemon start.
+    pub fn new() -> Result<Self> {
+        let rules = RuleStore::load()?;
+        let max_file_bytes = std::env::var("HOUNDD_MAX_FILE_BYTES")
+            .ok()
+            .and_then(|v| v.parse().ok())
+            .unwrap_or(DEFAULT_MAX_FILE_BYTES);
+        Ok(Self {
+            rules,
+            cache: VerdictCache::new(CACHE_CAPACITY),
+            max_file_bytes,
+        })
+    }
+
+    /// Cache statistics, for `hound status` and the benchmark harness.
+    pub fn cache_stats(&self) -> (usize, u64, u64) {
+        self.cache.stats()
+    }
+}
+
+/// Everything a walk accumulated. Kept separate from the wire type so the
+/// skipped count can be reported without changing the public API shape.
+#[derive(Default)]
+struct Tally {
+    scanned: u64,
+    infected: u64,
+    skipped_large: u64,
+    found: Vec<Found>,
+}
+
+impl ScanEngine for HoundEngine {
+    fn name(&self) -> &'static str {
+        "hound"
+    }
+
+    fn probe(&self) -> (bool, String, Option<DbFile>) {
+        let set = self.rules.current();
+        let summary = format!(
+            "{} rules loaded from {} source(s) [hound {}]",
+            set.count,
+            set.sources.len(),
+            set.version
+        );
+        (
+            true,
+            summary,
+            Some(DbFile {
+                file: set.version.clone(),
+                updated_at: to_rfc3339(set.loaded_at),
+            }),
+        )
+    }
+
+    fn scan(&self, path: &str, recursive: bool) -> Result<ScanResult> {
+        let root = std::fs::canonicalize(path).with_context(|| format!("no such path: {path}"))?;
+        let set = self.rules.current();
+
+        // One scanner for the whole walk. Constructing it per file would
+        // reintroduce a chunk of the per-invocation cost we just removed.
+        let mut scanner = yara_x::Scanner::new(&set.rules);
+        let mut tally = Tally::default();
+
+        let mut queue: Vec<PathBuf> = vec![root.clone()];
+        while let Some(current) = queue.pop() {
+            let Ok(md) = std::fs::symlink_metadata(&current) else {
+                continue;
+            };
+
+            if md.is_symlink() {
+                continue;
+            }
+
+            if md.is_dir() {
+                // The root is always descended into; deeper levels only
+                // when the caller asked for a recursive scan.
+                if current != root && !recursive {
+                    continue;
+                }
+                if is_skipped_dir(&current) {
+                    continue;
+                }
+                for entry in std::fs::read_dir(&current).into_iter().flatten().flatten() {
+                    queue.push(entry.path());
+                }
+                continue;
+            }
+
+            if !md.is_file() {
+                continue;
+            }
+            if md.len() > self.max_file_bytes {
+                tally.skipped_large += 1;
+                continue;
+            }
+
+            self.scan_one(&current, &md, &set, &mut scanner, &mut tally);
+        }
+
+        let clean = tally.scanned.saturating_sub(tally.infected);
+        Ok(ScanResult {
+            scanned: tally.scanned,
+            clean,
+            infected: tally.infected,
+            skipped: tally.skipped_large,
+            found: tally.found,
+        })
+    }
+
+    fn update(&self) -> Result<(bool, String, String)> {
+        let before = self.rules.current().count;
+        match self.rules.reload() {
+            Ok(set) => {
+                // Verdicts reached under the old rules say nothing about
+                // the new ones.
+                self.cache.clear();
+                Ok((
+                    true,
+                    "reload rules".to_string(),
+                    format!(
+                        "OK: {} rules loaded ({}), was {before}\nsources: {}\n",
+                        set.count,
+                        set.version,
+                        set.sources.join(", ")
+                    ),
+                ))
+            }
+            Err(e) => Ok((false, "reload rules".to_string(), format!("{e}\n"))),
+        }
+    }
+}
+
+impl HoundEngine {
+    /// Scan a single regular file, consulting the cache first.
+    fn scan_one(
+        &self,
+        path: &Path,
+        md: &std::fs::Metadata,
+        set: &Arc<RuleSet>,
+        scanner: &mut yara_x::Scanner,
+        tally: &mut Tally,
+    ) {
+        let key = FileKey::from_metadata(md);
+
+        if let Some(verdict) = self.cache.get(&key) {
+            tally.scanned += 1;
+            if let Some(name) = verdict {
+                tally.infected += 1;
+                tally.found.push(Found {
+                    path: path.to_string_lossy().into_owned(),
+                    virus: name.to_string(),
+                });
+            }
+            return;
+        }
+
+        let Ok(bytes) = std::fs::read(path) else {
+            // Unreadable is not clean, so it is not cached and not
+            // counted as scanned.
+            return;
+        };
+
+        tally.scanned += 1;
+
+        let Ok(results) = scanner.scan(&bytes) else {
+            return;
+        };
+
+        // A file can trip several rules; report it once, under the first
+        // match, exactly as the ClamAV path did with --allmatch.
+        let hit = results
+            .matching_rules()
+            .next()
+            .map(|r| RuleSet::detection_name(&r));
+
+        match hit {
+            Some(name) => {
+                tally.infected += 1;
+                tally.found.push(Found {
+                    path: path.to_string_lossy().into_owned(),
+                    virus: name.clone(),
+                });
+                self.cache.put(key, Some(name.into()));
+            }
+            None => self.cache.put(key, None),
+        }
+        let _ = set;
+    }
+}
+
+/// Pseudo-filesystems and our own vault: never walked.
+fn is_skipped_dir(path: &Path) -> bool {
+    SKIP_PREFIXES.iter().any(|p| path.starts_with(p))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::fs;
+
+    fn tmpdir(tag: &str) -> PathBuf {
+        let d = std::env::temp_dir().join(format!(
+            "hound-native-{tag}-{}-{:?}",
+            std::process::id(),
+            std::thread::current().id()
+        ));
+        let _ = fs::remove_dir_all(&d);
+        fs::create_dir_all(&d).unwrap();
+        d
+    }
+
+    const EICAR: &str = r"X5O!P%@AP[4\PZX54(P^)7CC)7}$EICAR-STANDARD-ANTIVIRUS-TEST-FILE!$H+H*";
+
+    #[test]
+    fn finds_eicar_in_a_directory() {
+        let d = tmpdir("eicar");
+        fs::write(d.join("clean.txt"), b"nothing to see").unwrap();
+        fs::write(d.join("eicar.com"), EICAR).unwrap();
+
+        let e = HoundEngine::new().unwrap();
+        let r = e.scan(d.to_str().unwrap(), true).unwrap();
+
+        assert_eq!(r.scanned, 2);
+        assert_eq!(r.infected, 1);
+        assert_eq!(r.clean, 1);
+        assert_eq!(r.found[0].virus, "EICAR-Test-Signature");
+        assert!(r.found[0].path.ends_with("eicar.com"));
+        let _ = fs::remove_dir_all(&d);
+    }
+
+    #[test]
+    fn clean_tree_is_clean() {
+        let d = tmpdir("clean");
+        for i in 0..20 {
+            fs::write(d.join(format!("f{i}.txt")), format!("file number {i}")).unwrap();
+        }
+        let e = HoundEngine::new().unwrap();
+        let r = e.scan(d.to_str().unwrap(), true).unwrap();
+        assert_eq!(r.scanned, 20);
+        assert_eq!(r.infected, 0);
+        assert!(r.is_clean());
+        let _ = fs::remove_dir_all(&d);
+    }
+
+    #[test]
+    fn non_recursive_stops_at_the_top() {
+        let d = tmpdir("shallow");
+        fs::write(d.join("top.txt"), b"top").unwrap();
+        let sub = d.join("sub");
+        fs::create_dir_all(&sub).unwrap();
+        fs::write(sub.join("eicar.com"), EICAR).unwrap();
+
+        let e = HoundEngine::new().unwrap();
+        let r = e.scan(d.to_str().unwrap(), false).unwrap();
+        assert_eq!(r.scanned, 1, "must not descend when recursive is false");
+        assert_eq!(r.infected, 0);
+        let _ = fs::remove_dir_all(&d);
+    }
+
+    #[test]
+    fn symlinks_are_not_followed() {
+        let d = tmpdir("symlink");
+        let real = d.join("real");
+        fs::create_dir_all(&real).unwrap();
+        fs::write(real.join("eicar.com"), EICAR).unwrap();
+        // A link pointing back at the parent would loop forever if followed.
+        std::os::unix::fs::symlink(&d, d.join("loop")).unwrap();
+
+        let e = HoundEngine::new().unwrap();
+        let r = e.scan(d.to_str().unwrap(), true).unwrap();
+        assert_eq!(r.scanned, 1, "the linked tree must not be walked twice");
+        assert_eq!(r.infected, 1);
+        let _ = fs::remove_dir_all(&d);
+    }
+
+    #[test]
+    fn oversized_files_are_skipped_not_scanned() {
+        let d = tmpdir("large");
+        fs::write(d.join("big.bin"), vec![0u8; 4096]).unwrap();
+        fs::write(d.join("small.txt"), b"ok").unwrap();
+
+        let mut e = HoundEngine::new().unwrap();
+        e.max_file_bytes = 1024;
+        let r = e.scan(d.to_str().unwrap(), true).unwrap();
+        assert_eq!(r.scanned, 1, "the 4 KB file must not be counted as scanned");
+        assert_eq!(r.skipped, 1, "and it must be reported, not silently dropped");
+        let _ = fs::remove_dir_all(&d);
+    }
+
+    #[test]
+    fn second_scan_is_served_from_cache() {
+        let d = tmpdir("cache");
+        for i in 0..30 {
+            fs::write(d.join(format!("f{i}.txt")), format!("contents {i}")).unwrap();
+        }
+        let e = HoundEngine::new().unwrap();
+
+        let first = e.scan(d.to_str().unwrap(), true).unwrap();
+        let (_, hits_after_first, _) = e.cache_stats();
+        assert_eq!(hits_after_first, 0, "a cold walk cannot hit");
+
+        let second = e.scan(d.to_str().unwrap(), true).unwrap();
+        let (_, hits, _) = e.cache_stats();
+
+        assert_eq!(first.scanned, second.scanned);
+        assert_eq!(hits, 30, "every file should be served from cache the second time");
+        let _ = fs::remove_dir_all(&d);
+    }
+
+    #[test]
+    fn editing_a_file_invalidates_its_cache_entry() {
+        let d = tmpdir("invalidate");
+        let f = d.join("mutable.txt");
+        fs::write(&f, b"harmless").unwrap();
+
+        let e = HoundEngine::new().unwrap();
+        let first = e.scan(d.to_str().unwrap(), true).unwrap();
+        assert_eq!(first.infected, 0);
+
+        // Rewrite the same path with EICAR. Size and mtime both change.
+        fs::write(&f, EICAR).unwrap();
+        let second = e.scan(d.to_str().unwrap(), true).unwrap();
+        assert_eq!(second.infected, 1, "a rewritten file must be rescanned");
+        let _ = fs::remove_dir_all(&d);
+    }
+
+    #[test]
+    fn scanning_a_single_file_works() {
+        let d = tmpdir("single");
+        let f = d.join("eicar.com");
+        fs::write(&f, EICAR).unwrap();
+        let e = HoundEngine::new().unwrap();
+        let r = e.scan(f.to_str().unwrap(), false).unwrap();
+        assert_eq!(r.scanned, 1);
+        assert_eq!(r.infected, 1);
+        let _ = fs::remove_dir_all(&d);
+    }
+
+    #[test]
+    fn missing_path_is_an_error() {
+        let e = HoundEngine::new().unwrap();
+        assert!(e.scan("/definitely/not/here/at/all", true).is_err());
+    }
+
+    #[test]
+    fn probe_reports_the_ruleset() {
+        let e = HoundEngine::new().unwrap();
+        let (present, summary, db) = e.probe();
+        assert!(present);
+        assert!(summary.contains("rules loaded"));
+        assert!(db.unwrap().file.starts_with("builtin-"));
+    }
+
+    #[test]
+    fn update_reloads_and_clears_the_cache() {
+        let d = tmpdir("update");
+        fs::write(d.join("a.txt"), b"a").unwrap();
+        let e = HoundEngine::new().unwrap();
+        e.scan(d.to_str().unwrap(), true).unwrap();
+        assert!(e.cache_stats().0 > 0);
+
+        let (ok, label, log) = e.update().unwrap();
+        assert!(ok, "reload should succeed: {log}");
+        assert_eq!(label, "reload rules");
+        assert_eq!(e.cache_stats().0, 0, "stale verdicts must be dropped");
+        let _ = fs::remove_dir_all(&d);
+    }
+
+    /// The Phase 0 exit criterion, asserted so it can never quietly
+    /// regress. The path this replaced took 6.5 seconds *per file*
+    /// because it reloaded a 169 MB database on every invocation.
+    ///
+    /// The bound is deliberately loose (2 s for 400 files, against a
+    /// measured ~9 ms) because CI machines are slow and shared, and a
+    /// flaky performance test gets deleted rather than fixed. It is
+    /// tight enough to catch the only regression that matters: someone
+    /// reintroducing per-file setup cost.
+    #[test]
+    fn four_hundred_files_scan_in_under_two_seconds() {
+        let d = tmpdir("perf");
+        for i in 0..399 {
+            fs::write(d.join(format!("f{i}.bin")), format!("payload {i}").repeat(64)).unwrap();
+        }
+        fs::write(d.join("eicar.com"), EICAR).unwrap();
+
+        let e = HoundEngine::new().unwrap();
+        let started = std::time::Instant::now();
+        let r = e.scan(d.to_str().unwrap(), true).unwrap();
+        let elapsed = started.elapsed();
+
+        assert_eq!(r.scanned, 400);
+        assert_eq!(r.infected, 1, "EICAR must still be caught at speed");
+        assert!(
+            elapsed < std::time::Duration::from_secs(2),
+            "400 files took {elapsed:?} — per-file setup cost is back"
+        );
+        eprintln!("perf: 400 files cold in {elapsed:?}");
+        let _ = fs::remove_dir_all(&d);
+    }
+
+    #[test]
+    fn skips_pseudo_filesystems() {
+        assert!(is_skipped_dir(Path::new("/proc/1")));
+        assert!(is_skipped_dir(Path::new("/sys/kernel")));
+        assert!(!is_skipped_dir(Path::new("/home/joe/src")));
+    }
+}

crates/houndd/src/rules.rs

@@ -0,0 +1,272 @@
+//! The rule store.
+//!
+//! Hound's detection content is YARA, compiled once at daemon start and
+//! held in memory for the process lifetime. This is the whole point of
+//! Phase 0: the old engine forked `clamscan` per file and paid a 6.5
+//! second signature-database load every single time. Here the ruleset is
+//! compiled once and every subsequent scan is a memory operation.
+//!
+//! Sources, in load order:
+//!
+//!   1. The built-in starter pack, compiled into the binary. Deliberately
+//!      tiny and tight so a fresh install detects something before it has
+//!      ever reached the network.
+//!   2. Every `*.yar` / `*.yara` in the rules directory — `$HOUNDD_RULES_DIR`
+//!      if set, else `/var/lib/hound/rules`, else the XDG data dir for
+//!      unprivileged runs. This is where the signed Hound pack lands in
+//!      Phase 3.
+//!
+//! Reload swaps a fresh `Arc<RuleSet>` into place; in-flight scans keep
+//! scanning against the ruleset they started with and the next scan picks
+//! up the new one. Nothing blocks and nothing is torn out from under a
+//! running scan.
+
+use anyhow::{Context, Result};
+use std::path::PathBuf;
+use std::sync::{Arc, RwLock};
+use std::time::SystemTime;
+
+/// The starter pack, baked into the binary.
+const BUILTIN: &str = include_str!("../rules/hound-builtin.yar");
+
+/// A compiled ruleset plus the provenance a client needs to display it.
+pub struct RuleSet {
+    pub rules: yara_x::Rules,
+    /// Wire version, e.g. "builtin-0.1.0" or the pack's own version file.
+    pub version: String,
+    /// How many rules compiled.
+    pub count: usize,
+    /// When this set was compiled.
+    pub loaded_at: SystemTime,
+    /// Human-readable list of what went in, for `hound status`.
+    pub sources: Vec<String>,
+}
+
+impl RuleSet {
+    /// Compile the built-in pack plus anything in the rules directory.
+    ///
+    /// A malformed file on disk is reported and skipped rather than
+    /// taking the daemon down — a bad third-party pack must not stop the
+    /// built-ins from protecting the machine.
+    fn compile() -> Result<Self> {
+        let mut compiler = yara_x::Compiler::new();
+        let mut sources = Vec::new();
+
+        compiler
+            .add_source(yara_x::SourceCode::from(BUILTIN).with_origin("hound-builtin.yar"))
+            .map_err(|e| anyhow::anyhow!("built-in rules failed to compile: {e}"))?;
+        sources.push("hound-builtin.yar (embedded)".to_string());
+
+        let mut version = format!("builtin-{}", env!("CARGO_PKG_VERSION"));
+
+        if let Some(dir) = rules_dir() {
+            if let Ok(v) = std::fs::read_to_string(dir.join("VERSION")) {
+                let v = v.trim();
+                if !v.is_empty() {
+                    version = v.to_string();
+                }
+            }
+            let mut files: Vec<PathBuf> = std::fs::read_dir(&dir)
+                .into_iter()
+                .flatten()
+                .flatten()
+                .map(|e| e.path())
+                .filter(|p| {
+                    p.extension()
+                        .is_some_and(|x| x == "yar" || x == "yara")
+                })
+                .collect();
+            files.sort();
+
+            for path in files {
+                let name = path
+                    .file_name()
+                    .map(|n| n.to_string_lossy().into_owned())
+                    .unwrap_or_default();
+                let src = match std::fs::read_to_string(&path) {
+                    Ok(s) => s,
+                    Err(e) => {
+                        eprintln!("rules: skipping {name}: {e}");
+                        continue;
+                    }
+                };
+                match compiler
+                    .add_source(yara_x::SourceCode::from(src.as_str()).with_origin(&name))
+                {
+                    Ok(_) => sources.push(name),
+                    Err(e) => eprintln!("rules: skipping {name}: {e}"),
+                }
+            }
+        }
+
+        let rules = compiler.build();
+        let count = rules.iter().count();
+
+        Ok(Self {
+            rules,
+            version,
+            count,
+            loaded_at: SystemTime::now(),
+            sources,
+        })
+    }
+
+    /// The detection name to report for a matching rule.
+    ///
+    /// Rules carry a `name` metadata field holding the public signature
+    /// name ("Linux.Coinminer.XMRig"); the rule identifier is the
+    /// fallback so a pack that omits the metadata still reports usefully.
+    pub fn detection_name(rule: &yara_x::Rule) -> String {
+        for (key, value) in rule.metadata() {
+            if key != "name" {
+                continue;
+            }
+            match value {
+                yara_x::MetaValue::String(s) => return s.to_string(),
+                yara_x::MetaValue::Bytes(b) => return b.to_string(),
+                _ => {}
+            }
+        }
+        rule.identifier().to_string()
+    }
+}
+
+/// Where on-disk packs live. `None` when no directory exists yet.
+pub fn rules_dir() -> Option<PathBuf> {
+    if let Some(dir) = std::env::var_os("HOUNDD_RULES_DIR") {
+        let p = PathBuf::from(dir);
+        return p.is_dir().then_some(p);
+    }
+    let system = PathBuf::from("/var/lib/hound/rules");
+    if system.is_dir() {
+        return Some(system);
+    }
+    let home = std::env::var_os("HOME")?;
+    let user = PathBuf::from(home).join(".local/share/hound/rules");
+    user.is_dir().then_some(user)
+}
+
+/// Hot-swappable handle on the current ruleset.
+#[derive(Clone)]
+pub struct RuleStore {
+    inner: Arc<RwLock<Arc<RuleSet>>>,
+}
+
+impl RuleStore {
+    /// Compile at startup. A failure here is fatal for detection, so we
+    /// surface it rather than silently serving an empty ruleset.
+    pub fn load() -> Result<Self> {
+        let set = RuleSet::compile().context("compiling rules")?;
+        Ok(Self {
+            inner: Arc::new(RwLock::new(Arc::new(set))),
+        })
+    }
+
+    /// The ruleset a scan should use. Cheap — one `Arc` clone.
+    pub fn current(&self) -> Arc<RuleSet> {
+        Arc::clone(&self.inner.read().expect("rule store poisoned"))
+    }
+
+    /// Recompile from source and swap the result in.
+    pub fn reload(&self) -> Result<Arc<RuleSet>> {
+        let fresh = Arc::new(RuleSet::compile().context("recompiling rules")?);
+        *self.inner.write().expect("rule store poisoned") = Arc::clone(&fresh);
+        Ok(fresh)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn builtin_pack_compiles() {
+        let set = RuleSet::compile().expect("built-in pack must always compile");
+        assert!(set.count >= 4, "expected the starter rules, got {}", set.count);
+        assert!(set.version.starts_with("builtin-"));
+        assert!(!set.sources.is_empty());
+    }
+
+    #[test]
+    fn detects_eicar() {
+        let set = RuleSet::compile().unwrap();
+        let mut scanner = yara_x::Scanner::new(&set.rules);
+        let eicar = br"X5O!P%@AP[4\PZX54(P^)7CC)7}$EICAR-STANDARD-ANTIVIRUS-TEST-FILE!$H+H*";
+        let results = scanner.scan(eicar).unwrap();
+        let names: Vec<String> = results
+            .matching_rules()
+            .map(|r| RuleSet::detection_name(&r))
+            .collect();
+        assert!(
+            names.iter().any(|n| n == "EICAR-Test-Signature"),
+            "EICAR must be detected, got {names:?}"
+        );
+    }
+
+    #[test]
+    fn clean_text_is_clean() {
+        let set = RuleSet::compile().unwrap();
+        let mut scanner = yara_x::Scanner::new(&set.rules);
+        let results = scanner.scan(b"the quick brown fox jumps over the lazy dog\n").unwrap();
+        assert_eq!(results.matching_rules().len(), 0);
+    }
+
+    /// The goodware gate, in miniature.
+    ///
+    /// Every rule in the starter pack is scanned against every binary in
+    /// `/usr/bin` and `/bin`. A single hit fails the build. Phase 3 scales
+    /// this to the Debian and Ubuntu archives plus the npm and PyPI top
+    /// 5,000, but the principle is already the one that matters: a rule
+    /// that fires on a system binary never ships.
+    ///
+    /// This test has already earned its keep — it caught a reverse-shell
+    /// rule that matched `/usr/bin/sudo`.
+    #[test]
+    fn no_false_positives_on_system_binaries() {
+        let set = RuleSet::compile().unwrap();
+        let mut scanner = yara_x::Scanner::new(&set.rules);
+        let mut checked = 0usize;
+        let mut failures: Vec<String> = Vec::new();
+
+        for dir in ["/usr/bin", "/bin", "/usr/sbin"] {
+            for entry in std::fs::read_dir(dir).into_iter().flatten().flatten() {
+                let path = entry.path();
+                let Ok(md) = std::fs::symlink_metadata(&path) else { continue };
+                if md.is_symlink() || !md.is_file() || md.len() > 32 * 1024 * 1024 {
+                    continue;
+                }
+                let Ok(bytes) = std::fs::read(&path) else { continue };
+                checked += 1;
+                let hits: Vec<String> = scanner
+                    .scan(&bytes)
+                    .map(|r| {
+                        r.matching_rules()
+                            .map(|m| RuleSet::detection_name(&m))
+                            .collect()
+                    })
+                    .unwrap_or_default();
+                if !hits.is_empty() {
+                    failures.push(format!("{} -> {hits:?}", path.display()));
+                }
+            }
+        }
+
+        assert!(checked > 50, "only {checked} binaries were readable — gate is not meaningful");
+        assert!(
+            failures.is_empty(),
+            "{} false positive(s) across {checked} system binaries:\n  {}",
+            failures.len(),
+            failures.join("\n  ")
+        );
+        eprintln!("goodware gate: {checked} system binaries, 0 false positives");
+    }
+
+    #[test]
+    fn reload_swaps_in_a_fresh_set() {
+        let store = RuleStore::load().unwrap();
+        let before = store.current();
+        let after = store.reload().unwrap();
+        assert_eq!(before.count, after.count);
+        assert!(after.loaded_at >= before.loaded_at);
+    }
+}

crates/houndd/src/settings.rs

@@ -149,3 +149,56 @@ mod tests {
         let _ = std::fs::remove_dir_all(&dir);
     }
 }
+
+#[cfg(test)]
+mod appearance_tests {
+    use hound_api::Settings;
+
+    #[test]
+    fn appearance_defaults_are_sane() {
+        let s = Settings::default();
+        assert_eq!(s.theme, "auto", "follow the desktop until told otherwise");
+        assert_eq!(s.tray_icon_style, "color");
+        assert!(s.close_to_tray, "the X must never stop protection");
+        assert!(s.confirm_quit, "quitting an antivirus is a deliberate act");
+    }
+
+    #[test]
+    fn unknown_theme_is_normalised_not_trusted() {
+        let mut s = Settings::default();
+        s.theme = "midnight-neon".into();
+        s.tray_icon_style = "sparkles".into();
+        s.normalise_appearance();
+        assert_eq!(s.theme, "auto");
+        assert_eq!(s.tray_icon_style, "color");
+    }
+
+    #[test]
+    fn valid_appearance_values_survive() {
+        for theme in Settings::THEMES {
+            for style in Settings::TRAY_STYLES {
+                let mut s = Settings::default();
+                s.theme = theme.into();
+                s.tray_icon_style = style.into();
+                s.normalise_appearance();
+                assert_eq!(s.theme, theme);
+                assert_eq!(s.tray_icon_style, style);
+            }
+        }
+    }
+
+    /// Settings files written before these fields existed must still load.
+    #[test]
+    fn older_settings_json_still_deserialises() {
+        let legacy = r#"{
+            "recursive_default": true, "max_file_size_mb": 100, "exclude_paths": [],
+            "realtime_enabled": true, "realtime_watch": [], "on_detect": "quarantine",
+            "ransomware_guard": true, "ransomware_threshold_per_min": 40,
+            "rootkit_enabled": true, "notify_desktop": true,
+            "auto_update_signatures": true, "paused": false
+        }"#;
+        let s: Settings = serde_json::from_str(legacy).expect("legacy settings must load");
+        assert_eq!(s.theme, "auto");
+        assert!(s.close_to_tray);
+    }
+}