Phase 2b: wire mkdir + unlink through the VFS to runtime.fs

The engine gained mkdir/unlink in Phase 2a; this exposes them as first-class
filesystem operations so programs can use them.

- vfs protocol: two new path-based operations, mkdir and unlink (appended, so
  existing opcodes/offsets are unchanged).
- VFS router: a forwardPath helper relays a path-based op under a mount to its
  backend; the mkdir/unlink cases forward to the mounted filesystem (the flat
  ramfs refuses them — it has no directories).
- fat server: mkdir -> engine.createDirectory, unlink -> engine.removeFile, each
  resolving the parent via a shared splitParent helper (also used by open-create).
- runtime.fs: makeDirectory(path) and remove(path).
- fat-test now exercises the whole path — mkdir /mnt/usb/TESTDIR, create + write +
  read a file inside it, then remove it — behind a new `fat-mutations` QEMU case.

Verified: zig build, zig build test, zig build check-fat-image, and a sequential
QEMU sweep — fat-mount, fat-mutations (mkdir/write/read/unlink through the mount),
vfs, vfs-client-death, log-flush, orderly-shutdown, initial-ramdisk, smoke — green.

system/services/vfs/vfs.zig

@@ -161,6 +161,22 @@ fn forwardRequest(out: []u8, backend: ipc.Handle, request: protocol.Request, pay
     return copy;
 }
 
+/// Forward a path-based operation (mkdir, unlink) under a mount to its backend and
+/// relay the reply. No handle is created — these operate by path and return only a
+/// status.
+fn forwardPath(out: []u8, backend: ipc.Handle, operation: protocol.Operation, relative: []const u8) usize {
+    const request = protocol.Request{ .operation = operation, .node = 0, .offset = 0, .len = @intCast(relative.len), .flags = 0 };
+    var message: [protocol.message_maximum]u8 = undefined;
+    @memcpy(message[0..protocol.request_size], std.mem.asBytes(&request));
+    const rel = relative[0..@min(relative.len, protocol.maximum_payload)];
+    @memcpy(message[protocol.request_size..][0..rel.len], rel);
+    var reply: [protocol.message_maximum]u8 = undefined;
+    const n = ipc.call(backend, message[0 .. protocol.request_size + rel.len], &reply) catch return fail(out);
+    const copy = @min(n, out.len);
+    @memcpy(out[0..copy], reply[0..copy]);
+    return copy;
+}
+
 /// Best-effort close of a backend node (used when a dead client's forwarding
 /// handles are swept — the backend must not leak the vfs's opens).
 fn forwardClose(backend: ipc.Handle, backend_node: u64) void {
@@ -305,6 +321,13 @@ fn handle(message: []const u8, out: []u8, sender: u32, capability: ?ipc.Handle)
             }
             return writeReply(out, .{ .status = 0 }, &.{});
         },
+        .mkdir, .unlink => {
+            const name = payload[0..@min(payload.len, request.len)];
+            if (longestMount(name)) |m| return forwardPath(out, m.backend, request.operation, m.relative);
+            // Only a mounted backend has real directories; the flat ramfs cannot
+            // create or remove them (and a bare-name path is not a mount target).
+            return fail(out);
+        },
     }
 }