diff --git a/src/backend/FSEvents.zig b/src/backend/FSEvents.zig
index 10094c7..53dc69d 100644
--- a/src/backend/FSEvents.zig
+++ b/src/backend/FSEvents.zig
@@ -220,6 +220,11 @@ fn callback(
         // a single event.  Emit one change call per applicable flag so
         // callers see all relevant event types (e.g. created + modified).
         const ot: ObjectType = if (flags & kFSEventStreamEventFlagItemIsDir != 0) .dir else .file;
+        // Handler errors are silently ignored: this callback runs on a GCD
+        // dispatch thread managed by the OS, so there is no way to propagate
+        // an error back to the caller.  Stopping the stream from inside the
+        // callback would require a separate signal channel and is not worth
+        // the complexity; the stream will keep delivering future events.
         if (flags & kFSEventStreamEventFlagItemCreated != 0) {
             ctx.handler.change(path, .created, ot) catch {};
         }
diff --git a/src/backend/INotify.zig b/src/backend/INotify.zig
index c71692b..db2864d 100644
--- a/src/backend/INotify.zig
+++ b/src/backend/INotify.zig
@@ -118,7 +118,10 @@ pub fn Create(comptime variant: InterfaceType) type {
                 .{ .fd = self.stop_pipe[0], .events = std.posix.POLL.IN, .revents = 0 },
             };
             while (true) {
-                _ = std.posix.poll(&pfds, -1) catch return;
+                _ = std.posix.poll(&pfds, -1) catch |e| {
+                    std.log.err("nightwatch: poll failed: {s}, stopping watch thread", .{@errorName(e)});
+                    return;
+                };
                 if (pfds[1].revents & std.posix.POLL.IN != 0) return; // stop signal
                 if (pfds[0].revents & std.posix.POLL.IN != 0) {
                     self.handle_read_ready(allocator) catch |e| {
diff --git a/src/backend/KQueue.zig b/src/backend/KQueue.zig
index 97fbdff..4cbe60f 100644
--- a/src/backend/KQueue.zig
+++ b/src/backend/KQueue.zig
@@ -112,7 +112,10 @@ fn thread_fn(self: *@This(), allocator: std.mem.Allocator) void {
     var events: [64]std.posix.Kevent = undefined;
     while (true) {
         // Block indefinitely until kqueue has events.
-        const n = std.posix.kevent(self.kq, &.{}, &events, null) catch break;
+        const n = std.posix.kevent(self.kq, &.{}, &events, null) catch |e| {
+            std.log.err("nightwatch: kevent failed: {s}, stopping watch thread", .{@errorName(e)});
+            break;
+        };
         for (events[0..n]) |ev| {
             if (ev.filter == EVFILT_READ) return; // shutdown pipe readable, exit
             if (ev.filter != EVFILT_VNODE) continue;
@@ -170,7 +173,8 @@ fn thread_fn(self: *@This(), allocator: std.mem.Allocator) void {
                     return;
                 };
             } else if (ev.fflags & NOTE_WRITE != 0) {
-                self.scan_dir(allocator, dir_path) catch {};
+                self.scan_dir(allocator, dir_path) catch |e|
+                    std.log.err("nightwatch: scan_dir failed: {s}", .{@errorName(e)});
             }
         }
     }
@@ -228,10 +232,16 @@ fn scan_dir(self: *@This(), allocator: std.mem.Allocator, dir_path: []const u8)
             }
         }
 
-        const gop = self.snapshots.getOrPut(allocator, dir_path) catch |e| {
-            return e;
-        };
-        if (!gop.found_existing) gop.value_ptr.* = .empty;
+        const gop = self.snapshots.getOrPut(allocator, dir_path) catch |e| return e;
+        if (!gop.found_existing) {
+            // dir_path points into a stack buffer; dupe it into allocator memory
+            // so the snapshot key outlives the current thread_fn iteration.
+            gop.key_ptr.* = allocator.dupe(u8, dir_path) catch |e| {
+                _ = self.snapshots.remove(dir_path);
+                return e;
+            };
+            gop.value_ptr.* = .empty;
+        }
         const snapshot = gop.value_ptr;
 
         var cit = current_files.iterator();
diff --git a/src/backend/KQueueDir.zig b/src/backend/KQueueDir.zig
index 594dd6e..ee8e0b2 100644
--- a/src/backend/KQueueDir.zig
+++ b/src/backend/KQueueDir.zig
@@ -97,7 +97,10 @@ fn thread_fn(self: *@This(), allocator: std.mem.Allocator) void {
     var events: [64]std.posix.Kevent = undefined;
     while (true) {
         // Block indefinitely until kqueue has events.
-        const n = std.posix.kevent(self.kq, &.{}, &events, null) catch break;
+        const n = std.posix.kevent(self.kq, &.{}, &events, null) catch |e| {
+            std.log.err("nightwatch: kevent failed: {s}, stopping watch thread", .{@errorName(e)});
+            break;
+        };
         for (events[0..n]) |ev| {
             if (ev.filter == EVFILT_READ) return; // shutdown pipe readable, exit
             if (ev.filter != EVFILT_VNODE) continue;
@@ -151,7 +154,8 @@ fn thread_fn(self: *@This(), allocator: std.mem.Allocator) void {
                         return;
                     };
                 } else if (ev.fflags & NOTE_WRITE != 0) {
-                    self.scan_dir(allocator, watch_path) catch {};
+                    self.scan_dir(allocator, watch_path) catch |e|
+                        std.log.err("nightwatch: scan_dir failed: {s}", .{@errorName(e)});
                 }
             }
         }
@@ -216,10 +220,16 @@ fn scan_dir(self: *@This(), allocator: std.mem.Allocator, dir_path: []const u8)
             }
         }
 
-        const gop = self.snapshots.getOrPut(allocator, dir_path) catch |e| {
-            return e;
-        };
-        if (!gop.found_existing) gop.value_ptr.* = .empty;
+        const gop = self.snapshots.getOrPut(allocator, dir_path) catch |e| return e;
+        if (!gop.found_existing) {
+            // dir_path points into a stack buffer; dupe it into allocator memory
+            // so the snapshot key outlives the current thread_fn iteration.
+            gop.key_ptr.* = allocator.dupe(u8, dir_path) catch |e| {
+                _ = self.snapshots.remove(dir_path);
+                return e;
+            };
+            gop.value_ptr.* = .empty;
+        }
         const snapshot = gop.value_ptr;
 
         var cit = current_files.iterator();
diff --git a/src/backend/Windows.zig b/src/backend/Windows.zig
index a0bc766..89bd94c 100644
--- a/src/backend/Windows.zig
+++ b/src/backend/Windows.zig
@@ -198,7 +198,8 @@ fn thread_fn(
                     // the `info` pointer (which points into w.buf) a dangling reference.
                     const next_entry_offset = info.NextEntryOffset;
                     watches_mutex.unlock();
-                    handler.change(full_path, event_type, object_type) catch {
+                    handler.change(full_path, event_type, object_type) catch |e| {
+                        std.log.err("nightwatch: handler returned {s}, stopping watch thread", .{@errorName(e)});
                         watches_mutex.lock();
                         break;
                     };