diff --git a/core/archipelago/src/api/rpc/mesh/status.rs b/core/archipelago/src/api/rpc/mesh/status.rs
index 2f004ffd..688ca399 100644
--- a/core/archipelago/src/api/rpc/mesh/status.rs
+++ b/core/archipelago/src/api/rpc/mesh/status.rs
@@ -101,12 +101,21 @@ impl RpcHandler {
             detected.iter().any(|d| d == &path),
             "{path} is not a detected mesh-radio candidate port"
         );
-        let service = self.mesh_service.read().await;
-        let probe = match service.as_ref() {
-            Some(svc) => svc.probe_device(&path).await?,
-            // No mesh service yet (radio never enabled) — probe directly.
-            None => mesh::listener::probe_device(&path).await?,
-        };
+        // Only hold the mesh_service lock long enough for the quick
+        // active-path guard check — NEVER across the actual probe, which
+        // can take 15-60s across its internal collision retries. Confirmed
+        // live 2026-07-23: holding this read lock for the full probe starved
+        // a concurrent firmware-flash job's MeshService::stop() (which needs
+        // the write lock) well past its own bounded timeout, surfacing as
+        // "Mesh listener did not release the serial port" even though
+        // stop() itself was fast.
+        {
+            let service = self.mesh_service.read().await;
+            if let Some(svc) = service.as_ref() {
+                svc.ensure_probe_allowed(&path).await?;
+            }
+        }
+        let probe = mesh::listener::probe_device(&path).await?;
         Ok(serde_json::to_value(probe)?)
     }
 
diff --git a/core/archipelago/src/mesh/listener/mod.rs b/core/archipelago/src/mesh/listener/mod.rs
index b2c237bf..dc6cc18b 100644
--- a/core/archipelago/src/mesh/listener/mod.rs
+++ b/core/archipelago/src/mesh/listener/mod.rs
@@ -547,6 +547,10 @@ pub fn spawn_mesh_listener(
         let mut shutdown = shutdown;
         let mut cmd_rx = cmd_rx;
         let mut reconnect_delay = RECONNECT_DELAY_INIT;
+        // Mutable so a successful auto-detect can pin the firmware kind for
+        // the rest of this listener's lifetime — see the pin-on-first-success
+        // block below for why.
+        let mut device_kind = device_kind;
         // Backlog #12 hot-swap re-binding: each run_mesh_session call already
         // builds a fresh device struct (contacts/current_region/etc. all
         // start empty), so per-device session state is naturally isolated
@@ -618,6 +622,45 @@ pub fn spawn_mesh_listener(
                 }
             }
 
+            // Pin the firmware kind after the first successful auto-detect.
+            // Confirmed live 2026-07-23: with device_kind left unpinned (e.g.
+            // after clearing a stale pin), EVERY reconnect re-runs the full
+            // Reticulum→Meshcore→Meshtastic auto-detect cascade — each
+            // candidate past the first does its own open() with the DTR/RTS
+            // reset both boards need, so a device correctly identified as
+            // Meshtastic still gets reset once for the failed Meshcore
+            // attempt before Meshtastic's own open() resets it again. That
+            // doubled the reset count on every single reconnect indefinitely,
+            // not just during initial detection. Once auto-detect has
+            // identified the device this listener is actually talking to,
+            // there's no reason to keep guessing on subsequent reconnects —
+            // pin it, both in this task's own loop (takes effect
+            // immediately) and on disk (survives a service restart). A
+            // genuine hot-swap to different firmware is still handled: the
+            // setup modal's `mesh.probe-device` always re-probes unpinned,
+            // and the flash flow already clears this pin on its own.
+            if device_kind.is_none() {
+                let detected = state.status.read().await.device_type;
+                if detected != super::types::DeviceType::Unknown {
+                    device_kind = Some(detected);
+                    match super::load_config(&data_dir).await {
+                        Ok(mut cfg) if cfg.device_kind.is_none() => {
+                            cfg.device_kind = Some(detected);
+                            if let Err(e) = super::save_config(&data_dir, &cfg).await {
+                                warn!("Failed to persist auto-detected device_kind: {}", e);
+                            } else {
+                                info!(
+                                    kind = %detected,
+                                    "Pinned auto-detected firmware kind to avoid repeated multi-protocol resets on reconnect"
+                                );
+                            }
+                        }
+                        Ok(_) => {}
+                        Err(e) => warn!("Failed to load mesh config to persist device_kind: {}", e),
+                    }
+                }
+            }
+
             // Update status to disconnected. device_type/firmware_version are
             // reset too — they were previously left holding the LAST radio's
             // identity, so after a hot-swap the UI showed the old firmware
diff --git a/core/archipelago/src/mesh/mod.rs b/core/archipelago/src/mesh/mod.rs
index 3a549e86..4e5b201b 100644
--- a/core/archipelago/src/mesh/mod.rs
+++ b/core/archipelago/src/mesh/mod.rs
@@ -1064,19 +1064,17 @@ impl MeshService {
         self.state.peers.read().await.values().cloned().collect()
     }
 
-    /// Probe a serial port for a mesh radio without provisioning or keeping
-    /// it — powers the hot-swap "device detected" modal's current-details
-    /// view. Refuses to probe the port the live session currently occupies
-    /// (the probe would steal the serial port from under the session); a
+    /// Refuse to probe the port the live session currently occupies (the
+    /// probe would steal the serial port from under the session); a
     /// detected-but-not-connected port is fair game, accepting a benign race
-    /// with the reconnect loop (whichever loses just retries).
-    pub async fn probe_device(&self, path: &str) -> Result<listener::DeviceProbe> {
+    /// with the reconnect loop (whichever loses just retries). Split out from
+    /// the actual probe on purpose — see `probe_device`'s doc comment.
+    pub async fn ensure_probe_allowed(&self, path: &str) -> Result<()> {
         let status = self.state.status.read().await;
         if status.device_connected && status.device_path.as_deref() == Some(path) {
             anyhow::bail!("{path} is the active mesh radio — already connected");
         }
-        drop(status);
-        listener::probe_device(path).await
+        Ok(())
     }
 
     /// Get message history.
diff --git a/core/archipelago/src/mesh/serial.rs b/core/archipelago/src/mesh/serial.rs
index 32f0451b..312c79f5 100644
--- a/core/archipelago/src/mesh/serial.rs
+++ b/core/archipelago/src/mesh/serial.rs
@@ -556,14 +556,38 @@ fn likely_non_mesh_serial_device(path: &str) -> bool {
 
 /// Scan for serial devices that could be Meshcore radios.
 /// Returns paths to existing serial device files.
+///
+/// Dedupes by canonical (symlink-resolved) target: `/dev/mesh-radio` is a
+/// stable udev symlink to whatever `/dev/ttyUSB*`/`/dev/ttyACM*` node the
+/// primary radio currently enumerates as, so both names always pointed at
+/// the same candidate list entry and both passed this scan — confirmed live
+/// 2026-07-23, this made an already-connected, working radio (connected via
+/// its `/dev/mesh-radio` alias) simultaneously appear as a second, separate
+/// "detected but unclaimed" device under its raw `/dev/ttyUSBn` name. The
+/// hot-swap UI's active-session guard compares path strings, so it didn't
+/// recognize the two aliases as the same port, showed the "device detected"
+/// modal for a radio that was already set up, and probing it there opened
+/// (and DTR/RTS-reset) the exact port the live session was mid-conversation
+/// with — a continuous, UI-driven reset loop that only ran while that view
+/// was open (matches the reported "stops when I leave, resumes when I come
+/// back"). SERIAL_CANDIDATES lists `/dev/mesh-radio` first, so it wins the
+/// dedup and is what's reported when both alias and target are present.
 pub async fn detect_serial_devices() -> Vec<String> {
     let mut devices = Vec::new();
+    let mut seen_real_paths = std::collections::HashSet::new();
     for path in SERIAL_CANDIDATES {
         if tokio::fs::metadata(path).await.is_ok() {
             if likely_non_mesh_serial_device(path) {
                 debug!(path = %path, "Skipping known non-mesh serial device");
                 continue;
             }
+            let real_path = tokio::fs::canonicalize(path)
+                .await
+                .unwrap_or_else(|_| std::path::PathBuf::from(path));
+            if !seen_real_paths.insert(real_path.clone()) {
+                debug!(path = %path, real_path = %real_path.display(), "Skipping duplicate alias for an already-listed device");
+                continue;
+            }
             devices.push(path.to_string());
         }
     }