diff --git a/system/services/volume-manager/partition.zig b/system/services/volume-manager/partition.zig
index 00ea139..f631e48 100644
--- a/system/services/volume-manager/partition.zig
+++ b/system/services/volume-manager/partition.zig
@@ -56,6 +56,13 @@ pub fn firstVolume(block0: []const u8, device_blocks: u64) ?Volume {
         const start = std.mem.readInt(u32, entry[8..12], .little);
         const size = std.mem.readInt(u32, entry[12..16], .little);
         if (kind == 0 or start == 0 or size == 0) continue;
+        // These bytes come off an untrusted removable medium. A partition that
+        // does not fit inside the device is not a partition — skip it. This is
+        // where the driver's confinement-safety invariant is established: the
+        // clamp's overflow-safety rests on base + count staying inside the
+        // device (usb-storage.zig resolveTransfer), which only holds because the
+        // range handed down is validated here. The subtraction cannot overflow.
+        if (start > device_blocks or device_blocks - start < size) continue;
         return .{ .base_lba = start, .block_count = size, .identity = identityOf(block0, index) };
     }
     // No partition entries: a bare FAT spanning the device.
@@ -90,3 +97,20 @@ test "no boot signature is no volume" {
     const block0 = [_]u8{0} ** 512;
     try std.testing.expect(firstVolume(&block0, 65536) == null);
 }
+
+test "a partition that runs past the device is skipped, not trusted" {
+    var block0 = [_]u8{0} ** 512;
+    block0[510] = 0x55;
+    block0[511] = 0xAA;
+    // partition 0: start 0xFFFFFF00, size 0x400 — far past a 200000-block device.
+    block0[446 + 4] = 0x0c;
+    std.mem.writeInt(u32, block0[446 + 8 ..][0..4], 0xFFFFFF00, .little);
+    std.mem.writeInt(u32, block0[446 + 12 ..][0..4], 0x400, .little);
+    // partition 1: start 2048, size 1000 — fits.
+    block0[462 + 4] = 0x0c;
+    std.mem.writeInt(u32, block0[462 + 8 ..][0..4], 2048, .little);
+    std.mem.writeInt(u32, block0[462 + 12 ..][0..4], 1000, .little);
+    const v = firstVolume(&block0, 200000).?;
+    try std.testing.expectEqual(@as(u64, 2048), v.base_lba); // the fitting one, not the overflowing one
+    try std.testing.expectEqual(@as(u64, 1000), v.block_count);
+}
diff --git a/system/services/volume-manager/volume-manager.zig b/system/services/volume-manager/volume-manager.zig
index bc10b28..c13ff09 100644
--- a/system/services/volume-manager/volume-manager.zig
+++ b/system/services/volume-manager/volume-manager.zig
@@ -56,8 +56,25 @@ var bounce: memory.DmaRegion = undefined;
 var bounce_ready = false;
 var probed = false;
 var volume: ?Volume = null;
+/// The storage channel, acquired ONCE and kept — re-acquiring on every probe
+/// retry would leak a handle per attempt on a medium-absent device.
+var storage_device: ?block.Device = null;
+var logged_no_volume = false;
 const probe_retry_ms = 500;
 
+// Filesystem supervision, mirroring the device manager's (device-manager.zig):
+// a clean exit is not restarted, a fault restarts with backoff, and a fast
+// crash loop gives up rather than spinning. Without this a faulting filesystem
+// respawns in a zero-delay loop.
+const fast_death_ns: u64 = 2_000_000_000;
+const crash_loop_cap: u32 = 3;
+const backoff_base_ms: u64 = 300;
+var fs_restarts: u32 = 0;
+var fs_spawn_ns: u64 = 0;
+var fs_failed = false;
+/// Set when a backoff timer is pending so its tick respawns rather than probes.
+var restart_pending = false;
+
 fn acquireStorage() ?block.Device {
     const manager = manager_handle orelse opened: {
         const handle = channel.openEndpoint("device-manager") orelse return null;
@@ -94,26 +111,45 @@ fn acquireStorage() ?block.Device {
 /// runs, so its first read is already bounded; the volume manager is the
 /// confinement controller (it defines the first range on the device).
 fn spawnFilesystem(v: *Volume) void {
+    if (fs_failed) return;
     const pid = process.spawnSupervised(filesystem_binary, &.{"1"}, service_endpoint) orelse {
-        _ = logging.write("volume-manager: could not spawn the filesystem\n");
+        _ = logging.write("volume-manager: could not spawn the filesystem; retrying\n");
+        armRestart();
         return;
     };
     if (!v.storage.defineRange(pid, v.base_lba, v.block_count)) {
-        _ = logging.write("volume-manager: could not confine the filesystem to its volume\n");
+        _ = logging.write("volume-manager: could not confine the filesystem to its volume; retrying\n");
         _ = process.kill(pid);
+        armRestart();
         return;
     }
     v.filesystem_pid = pid;
+    fs_spawn_ns = time.clock();
     std.log.info("volume 0x{x} -> {s} (pid {d}), lba {d}, {d} blocks", .{ v.identity, filesystem_binary, pid, v.base_lba, v.block_count });
 }
 
+/// Arm a one-shot timer to (re)spawn the filesystem after backoff — used both
+/// when a spawn step fails and when a running filesystem faults. Distinguished
+/// from the probe timer by `restart_pending`.
+fn armRestart() void {
+    const delay = if (fs_restarts == 0) backoff_base_ms else backoff_base_ms << @intCast(@min(fs_restarts - 1, 5));
+    restart_pending = true;
+    _ = time.timerOnce(service_endpoint, delay);
+}
+
 fn tryProbe() void {
     if (probed) return;
     if (!bounce_ready) {
         bounce = memory.dmaAlloc(512, memory.dma_coherent | memory.dma_shareable) orelse return;
         bounce_ready = true;
     }
-    const device = acquireStorage() orelse return;
+    // Acquire the storage channel once and keep it: a fresh consumer-hello per
+    // retry would leak a handle every 500 ms on a device whose medium is absent.
+    const device = storage_device orelse acquired: {
+        const d = acquireStorage() orelse return;
+        storage_device = d;
+        break :acquired d;
+    };
     if (bounce.handle) |handle| {
         if (!device.attach(handle)) return;
         _ = ipc.close(handle);
@@ -123,8 +159,13 @@ fn tryProbe() void {
     if (!device.read(0, 1, bounce.physical)) return;
     const sector: [*]const u8 = @ptrFromInt(bounce.virtual);
     const found = partition.firstVolume(sector[0..512], geometry.block_count) orelse {
-        _ = logging.write("volume-manager: no volume found on the storage device\n");
-        probed = true;
+        // No volume yet. On removable media this can mean no medium is present —
+        // keep polling so an inserted medium is picked up (the removal-lifecycle
+        // trigger). Log once, and do NOT terminate the probe.
+        if (!logged_no_volume) {
+            _ = logging.write("volume-manager: no volume on the storage device yet\n");
+            logged_no_volume = true;
+        }
         return;
     };
     volume = .{ .storage = device, .base_lba = found.base_lba, .block_count = found.block_count, .identity = found.identity, .id = volume_id };
@@ -169,22 +210,40 @@ fn initialise(endpoint: ipc.Handle) bool {
 fn onNotification(badge: u64) void {
     const got = ipc.Received{ .len = 0, .badge = badge, .cap = null };
     if (got.isTimer()) {
+        // One timer signal, two jobs, told apart by state: a pending backoff
+        // restart, otherwise the probe retry.
+        if (restart_pending) {
+            restart_pending = false;
+            if (volume) |*v| spawnFilesystem(v);
+            return;
+        }
         tryProbe();
         if (!probed) _ = time.timerOnce(service_endpoint, probe_retry_ms);
         return;
     }
-    // A filesystem died. Its old range is reclaimed by the driver on the same
-    // death; respawn it, confined afresh to the same volume (a fresh pid, a
-    // fresh range). The reap-and-rebuild the device manager proved, one layer up.
+    // A filesystem died. The exit reason drives the decision, exactly as the
+    // device manager supervises drivers: a clean exit meant to stop; a fault
+    // restarts with backoff until a fast crash loop gives up. The old range is
+    // reclaimed by the driver on the same death; the respawn confines afresh.
     if (got.isChildExit()) {
         const dead = got.childProcessId();
-        if (volume) |*v| {
-            if (v.filesystem_pid == dead) {
-                v.filesystem_pid = 0;
-                std.log.info("filesystem for volume {d} died; respawning", .{v.id});
-                spawnFilesystem(v);
-            }
+        const v = &(volume orelse return);
+        if (v.filesystem_pid != dead) return;
+        v.filesystem_pid = 0;
+        const reason = process.exitReason(dead) orelse .fault;
+        if (reason == .exited) {
+            std.log.info("filesystem for volume {d} exited cleanly; not restarting", .{v.id});
+            return;
         }
+        const alive = time.clock() -| fs_spawn_ns;
+        fs_restarts = if (alive < fast_death_ns) fs_restarts + 1 else 1;
+        if (fs_restarts >= crash_loop_cap) {
+            fs_failed = true;
+            std.log.info("filesystem for volume {d} is failing repeatedly; giving up", .{v.id});
+            return;
+        }
+        std.log.info("filesystem for volume {d} died ({s}); restarting", .{ v.id, @tagName(reason) });
+        armRestart();
     }
 }