Why does RS-232 need a stop bit?

The thing to remember is that RS232 is an asynchronous protocol. There is no clock signal associated with it.

Figure 1. Receiver sampling points. Source: Sangoma.

The start bit is used to trigger the read cycle in the receiver. The receiver synchronises itself on the start bit and then waits 1.5 cycles to start sampling bits. Thereafter the bits are sampled at the baud rate. This initial delay means that even with a 5% clock error the receiver should still be within the bit timing for the last bit.

Since the start bit - shown low in Figure 1. - is identified by a falling edge then it must be preceded by a high and this is what the stop bit ensures. The alternative would be two start bits and no stop bits but it wouldn't change the total message length.

The linked article has some other points worth noting.

RS-232 doesn't require it; some RS-232 devices do. In particular, serial/RS-232 interfaces on computers are often RS-232 with UART (Universal Asychronous Receiver/Transmitter) which supports only asynchronous transmission.

Back in its heyday, RS-232 was commonly used for networking protocols like 'bisync' (BSC), SNA/SDLC, X.25/LAPB, and DECnet/HDLC, which used synchronous transmission of a 'frame' or 'block', typically up to several hundred octets, continuous (no start or stop bits) from a beginning marker to an ending marker. The latter three used bit stuffing (transparent to software at either end) partly to ensure enough transitions to maintain bit-level synchronization regardless of data. Both UART (async only) and USART (sync and async) chips were available, but the former were cheaper and more commonly used.

By the 1990s most if not all synchronous uses of RS-232 were superseded by local Ethernet (and later Ethernet-emulating 802.11) or Token-Ring (now mostly forgotten but then a serious competitor) and remote T-1 ISDN or Frame Relay, while some connections that were naturally or conventionally async (such as cheapish dot-matrix printers) remained, so computer designers used a cheaper async-only serial interface (or in recent years none at all).