From the datasheet section 22.7.7 linked below, the number of clock cycles for a single bit is 1 + (TS1[3:0] + 1) + (TS2[2:0] + 1). Using the previous values of 12 and 3 for TS1 and TS2 respectively, this lead to a bit length of 18 clock cycles. With a BRP value of 15 for a prescaler of 16 to achieve a bitrate of 125kbit/s, the bit length would be 1/32 us * (18) * (16), leading to a bit length of 9 us instead of the targeted 8 us. With the values of 11 and 2 for TS1 and TS2 instead, the bit length is 1/32 us * 16 * 16 which corresponds with the correct bit length of 8us, while still having an 80% sample point.

https://www.keil.com/dd/docs/datashts/st/stm32f10xxx.pdf