diff --git a/Sources/KeyPathHelper/HelperService.swift b/Sources/KeyPathHelper/HelperService.swift
index c1c588cf0..15d179d8a 100644
--- a/Sources/KeyPathHelper/HelperService.swift
+++ b/Sources/KeyPathHelper/HelperService.swift
@@ -879,6 +879,8 @@ extension HelperService {
             <true/>
             <key>KeepAlive</key>
             <false/>
+            <key>ProcessType</key>
+            <string>Interactive</string>
             <key>StandardOutPath</key>
             <string>/var/log/com.keypath.kanata.stdout.log</string>
             <key>StandardErrorPath</key>
@@ -914,6 +916,8 @@ extension HelperService {
             <true/>
             <key>KeepAlive</key>
             <true/>
+            <key>ProcessType</key>
+            <string>Interactive</string>
             <key>StandardOutPath</key>
             <string>/var/log/karabiner-vhid-daemon.log</string>
             <key>StandardErrorPath</key>
diff --git a/Sources/KeyPathInstallationWizard/Core/PlistGenerator.swift b/Sources/KeyPathInstallationWizard/Core/PlistGenerator.swift
index c92a37b4c..46e0dff9b 100644
--- a/Sources/KeyPathInstallationWizard/Core/PlistGenerator.swift
+++ b/Sources/KeyPathInstallationWizard/Core/PlistGenerator.swift
@@ -105,6 +105,8 @@ public enum PlistGenerator {
             <true/>
             <key>KeepAlive</key>
             <false/>
+            <key>ProcessType</key>
+            <string>Interactive</string>
             <key>StandardOutPath</key>
             <string>/var/log/com.keypath.kanata.stdout.log</string>
             <key>StandardErrorPath</key>
@@ -138,9 +140,16 @@ public enum PlistGenerator {
     /// - Automatic restart (KeepAlive)
     /// - Logging to /var/log/karabiner-vhid-daemon.log
     /// - Throttle protection to prevent rapid restart loops
+    /// - Interactive process type so the daemon is not starved under CPU load
     ///
     /// This daemon is required for Kanata to access the virtual keyboard device.
     ///
+    /// ProcessType=Interactive matches the Kanata daemon plist. Without it,
+    /// heavy system load (e.g. parallel builds) can starve the daemon long
+    /// enough to miss the pqrs client's 3s heartbeat, dropping Kanata's output
+    /// connection mid-keystroke and leaving a key stuck down (autorepeat
+    /// bursts). See docs/bugs/MAL-57-duplicate-keypresses.md.
+    ///
     /// - Returns: Complete plist XML string ready to write to disk
     public static func generateVHIDDaemonPlist() -> String {
         """
@@ -158,6 +167,8 @@ public enum PlistGenerator {
             <true/>
             <key>KeepAlive</key>
             <true/>
+            <key>ProcessType</key>
+            <string>Interactive</string>
             <key>StandardOutPath</key>
             <string>/var/log/karabiner-vhid-daemon.log</string>
             <key>StandardErrorPath</key>
diff --git a/Sources/KeyPathInstallationWizard/Core/ServiceHealthChecker.swift b/Sources/KeyPathInstallationWizard/Core/ServiceHealthChecker.swift
index 035aa0e72..984c843a5 100644
--- a/Sources/KeyPathInstallationWizard/Core/ServiceHealthChecker.swift
+++ b/Sources/KeyPathInstallationWizard/Core/ServiceHealthChecker.swift
@@ -873,7 +873,12 @@ public final class ServiceHealthChecker: @unchecked Sendable {
         return Foundation.FileManager().fileExists(atPath: plistPath)
     }
 
-    /// Verifies that the installed VHID LaunchDaemon plist points to the DriverKit daemon path.
+    /// Verifies that the installed VHID LaunchDaemon plist points to the DriverKit daemon path
+    /// and carries the required ProcessType=Interactive key.
+    ///
+    /// ProcessType=Interactive keeps the daemon from being starved under CPU
+    /// load (MAL-57 stuck-key autorepeat). A plist without it predates the fix
+    /// and should be treated as misconfigured so repair rewrites it.
     ///
     /// - Returns: `true` if the plist is correctly configured
     public func isVHIDDaemonConfiguredCorrectly() -> Bool {
@@ -889,11 +894,12 @@ public final class ServiceHealthChecker: @unchecked Sendable {
             "/Library/Application Support/org.pqrs/Karabiner-DriverKit-VirtualHIDDevice/Applications/Karabiner-VirtualHIDDevice-Daemon.app/Contents/MacOS/Karabiner-VirtualHIDDevice-Daemon"
 
         if let args = dict["ProgramArguments"] as? [String], let first = args.first {
-            let ok = first == expectedPath
+            let pathOK = first == expectedPath
+            let processTypeOK = (dict["ProcessType"] as? String) == "Interactive"
             AppLogger.shared.log(
-                "🔍 [ServiceHealthChecker] VHID plist ProgramArguments[0]=\(first) | expected=\(expectedPath) | ok=\(ok)"
+                "🔍 [ServiceHealthChecker] VHID plist ProgramArguments[0]=\(first) | pathOK=\(pathOK) | processTypeOK=\(processTypeOK)"
             )
-            return ok
+            return pathOK && processTypeOK
         }
         AppLogger.shared.log(
             "🔍 [ServiceHealthChecker] VHID plist ProgramArguments missing or malformed"
diff --git a/Tests/KeyPathTests/Services/PlistGeneratorTests.swift b/Tests/KeyPathTests/Services/PlistGeneratorTests.swift
index 8695e715a..48a382386 100644
--- a/Tests/KeyPathTests/Services/PlistGeneratorTests.swift
+++ b/Tests/KeyPathTests/Services/PlistGeneratorTests.swift
@@ -212,6 +212,21 @@ final class PlistGeneratorTests: XCTestCase {
         XCTAssertTrue(plist.contains(PlistGenerator.vhidDaemonPath))
     }
 
+    /// ProcessType=Interactive prevents the VHID daemon from being starved
+    /// under CPU load, which drops Kanata's output connection mid-keystroke
+    /// and causes stuck-key autorepeat bursts (MAL-57).
+    func testGenerateVHIDDaemonPlistUsesInteractiveProcessType() {
+        let plist = PlistGenerator.generateVHIDDaemonPlist()
+
+        guard let data = plist.data(using: .utf8),
+              let dict = try? PropertyListSerialization.propertyList(from: data, options: [], format: nil) as? [String: Any]
+        else {
+            XCTFail("Generated VHID daemon plist is not valid XML/plist format")
+            return
+        }
+        XCTAssertEqual(dict["ProcessType"] as? String, "Interactive")
+    }
+
     func testGenerateVHIDDaemonPlistValidXML() {
         let plist = PlistGenerator.generateVHIDDaemonPlist()
 
diff --git a/Tests/KeyPathTests/Services/ServiceHealthCheckerTests.swift b/Tests/KeyPathTests/Services/ServiceHealthCheckerTests.swift
index 10b848eb5..abe4b6b46 100644
--- a/Tests/KeyPathTests/Services/ServiceHealthCheckerTests.swift
+++ b/Tests/KeyPathTests/Services/ServiceHealthCheckerTests.swift
@@ -77,10 +77,13 @@ final class ServiceHealthCheckerTests: XCTestCase {
         FileManager.default.createFile(atPath: url.path, contents: Data(), attributes: nil)
     }
 
-    private func writeVHIDPlist(programPath: String) throws {
-        let dict: [String: Any] = [
+    private func writeVHIDPlist(programPath: String, processType: String? = "Interactive") throws {
+        var dict: [String: Any] = [
             "ProgramArguments": [programPath]
         ]
+        if let processType {
+            dict["ProcessType"] = processType
+        }
         let url = tempLaunchDaemonsDir.appendingPathComponent("\(ServiceHealthChecker.vhidDaemonServiceID).plist")
         let data = try PropertyListSerialization.data(fromPropertyList: dict, format: .xml, options: 0)
         try data.write(to: url)
@@ -269,4 +272,13 @@ final class ServiceHealthCheckerTests: XCTestCase {
         try writeVHIDPlist(programPath: "/wrong/path")
         XCTAssertFalse(checker.isVHIDDaemonConfiguredCorrectly())
     }
+
+    /// Plists from before the MAL-57 starvation fix lack ProcessType=Interactive
+    /// and must report misconfigured so repair rewrites them.
+    func testIsVHIDDaemonConfiguredCorrectlyReturnsFalseWithoutProcessType() throws {
+        let expectedPath =
+            "/Library/Application Support/org.pqrs/Karabiner-DriverKit-VirtualHIDDevice/Applications/Karabiner-VirtualHIDDevice-Daemon.app/Contents/MacOS/Karabiner-VirtualHIDDevice-Daemon"
+        try writeVHIDPlist(programPath: expectedPath, processType: nil)
+        XCTAssertFalse(checker.isVHIDDaemonConfiguredCorrectly())
+    }
 }
diff --git a/dev-tools/debug/debug-plist-validation.sh b/dev-tools/debug/debug-plist-validation.sh
index c3bc97933..efeebb49c 100755
--- a/dev-tools/debug/debug-plist-validation.sh
+++ b/dev-tools/debug/debug-plist-validation.sh
@@ -61,6 +61,8 @@ cat > "$VHID_DAEMON_TEMP" << 'EOF'
     <true/>
     <key>KeepAlive</key>
     <true/>
+    <key>ProcessType</key>
+    <string>Interactive</string>
     <key>StandardOutPath</key>
     <string>/var/log/karabiner-vhid-daemon.log</string>
     <key>StandardErrorPath</key>
diff --git a/docs/bugs/MAL-57-duplicate-keypresses.md b/docs/bugs/MAL-57-duplicate-keypresses.md
index af369bff5..981bf703c 100644
--- a/docs/bugs/MAL-57-duplicate-keypresses.md
+++ b/docs/bugs/MAL-57-duplicate-keypresses.md
@@ -10,6 +10,104 @@ as previous conclusions, not final truth.
 Latest investigation summary:
 
 - [`docs/analysis/2026-03-07-duplicate-key-under-load-investigation.md`](/Users/malpern/local-code/KeyPath/.worktrees/duplicate-key-investigation/docs/analysis/2026-03-07-duplicate-key-under-load-investigation.md)
+- 2026-06-10 live incident evidence below (strongest capture so far; mechanism confirmed end to end)
+
+## 2026-06-10 Incident Evidence (mechanism confirmed)
+
+Two live incidents were captured with full logs on all three layers (kanata, the
+Karabiner VHID daemon, and KeyPath). Both produced multi-second autorepeat bursts
+of a single letter while typing ("togggggg…", "hteeeeee…"). Machine context: several
+agent worktrees running parallel Swift builds and `quick-deploy.sh` at the time of
+both incidents. The KeyPath GUI (and therefore the stuck-key detector/recovery)
+was **not** running.
+
+### Timeline (from `/var/log/com.keypath.kanata.stdout.log` and `/var/log/karabiner/virtual_hid_device_service.log`)
+
+Incident 1 — 21:42:17 (the "g" burst):
+
+```
+driver connected: false / driver connected: true   (repeated flapping for minutes prior)
+21:42:17.3224 [WARN] output backend unavailable during write — releasing input devices
+21:42:17.3265 [WARN] output backend suspended: recovery-entry
+21:42:18.4464 [INFO] output backend and console session ready — re-grabbing input devices
+21:42:18.5345 [KeyInput] sent key=backspace ...      <- user deleting the repeated g's
+21:42:18.7421 managed repeat BSpace (x8)
+```
+
+Incident 2 — 22:00:08-09 (the "e" burst). Kanata logged the user typing
+"check hte logs … and", then:
+
+```
+22:00:08.5099 [KeyInput] sent key=e action=Press
+22:00:08.6263 [KeyInput] sent key=e action=Release   <- written into a dying socket; never took effect
+22:00:09.6736 [KeyInput] sent key=d action=Press
+connected                                            <- pqrs client reconnected
+22:00:09.7820 [WARN] output backend unavailable during write — releasing input devices
+22:00:09.7821 [WARN] dropping KEY_D Release: output backend unavailable (will recover)
+driver connected: false
+driver connected: true
+22:00:14.2039 [INFO] output backend and console session ready — re-grabbing input devices
+```
+
+Daemon side (`virtual_hid_device_service.log`): new client socket connected at
+22:00:10.150; the old client (and with it the virtual keyboard state holding the
+stuck key) was not torn down until 22:00:11.378-11.744. From the lost `e` release
+at 22:00:08.6 that is ~3 seconds of OS-level autorepeat.
+
+Both the dext (pid 736) and the Karabiner-VirtualHIDDevice-Daemon (pid 814) ran
+continuously through both incidents — no crashes. The failure is purely the
+socket between kanata and the daemon.
+
+### Confirmed causal chain
+
+1. Kanata talks to the VHID daemon via the pqrs client in the `karabiner-driverkit`
+   crate. The client pings the daemon every 3s and declares the connection dead on a
+   missed heartbeat deadline (`virtual_hid_device_service/client.hpp`:
+   `set_server_check_interval(3000ms)`, `next_heartbeat_deadline_exceeded`).
+2. Under CPU load the heartbeat misses and the connection drops. Key releases
+   in flight are lost: either written into the dying socket (the `e`) or
+   explicitly dropped by `drop_if_sink_disconnected` in
+   `External/kanata/src/oskbd/macos.rs` (the `d` — `dropping KEY_D Release`).
+3. The virtual HID keyboard keeps the key logically down until the daemon tears
+   down the old client (~1.2s after the new client connects). macOS autorepeats
+   the stuck key for the whole window.
+4. The fork's recovery (`release_tracked_output_keys` in
+   `External/kanata/src/kanata/macos.rs`) runs at re-grab — *after* the damage
+   window has already closed on its own, so it cannot shorten the burst.
+
+### Why the daemon starves: launchd priority asymmetry
+
+The kanata daemon plist sets `ProcessType=Interactive`; the KeyPath-installed
+`com.keypath.karabiner-vhiddaemon.plist` set **no** `ProcessType`, so under load
+macOS may deprioritize exactly the process that must answer heartbeats and
+process key reports.
+
+### Fix layers
+
+- **Layer 1 (done, this repo):** `ProcessType=Interactive` added to the VHID
+  daemon plist in `PlistGenerator.generateVHIDDaemonPlist()`, the helper's
+  copy in `HelperService.swift`, the legacy kanata generators (the shipped
+  SMAppService plist `Sources/KeyPathApp/com.keypath.kanata.plist` already had
+  it), and the debug script template. `isVHIDDaemonConfiguredCorrectly()` now
+  also requires the key, so a pre-fix plist reports misconfigured and repair
+  rewrites it (`installOrRepairVHIDServices` / `repairVHIDDaemonServices`
+  rewrite unconditionally). Note: that check currently feeds only the repair
+  postflight — wiring plist-content validation into `getServiceStatus()` so the
+  wizard proactively flags stale plists on old installs is a follow-up.
+  Expected effect: fewer heartbeat-miss disconnects, so far fewer incidents.
+- **Layer 2 (open, kanata fork):** shrink the autorepeat window. On reconnect the
+  client's `connected` callback only calls `virtual_hid_keyboard_initialize`;
+  also issuing `virtual_hid_keyboard_reset` would clear stuck keys immediately
+  (~10.2s in the timeline above instead of 11.7s). Force-closing the old client
+  object as soon as `sink_ready` flips false would help further.
+- **Layer 3 (open, kanata fork):** never silently drop a Release.
+  `drop_if_sink_disconnected` treats all events equally; dropped Releases are the
+  toxic case and should be remembered and replayed after reconnect
+  (`output_pressed_since` tracking already exists).
+
+Residual gap no fix can fully close: a release written into a socket that is
+dying but not yet declared dead is unrecoverable in transit; Layer 2 bounds the
+damage to the reconnect latency.
 
 ## Problem Statement