The stall-during-erase picture in #3 is the right first split: many MCUs halt the CPU or mask interrupts for a page or sector erase, so an otherwise healthy watchdog reset is expected if the timeout is shorter than the worst-case erase. I would not treat any particular erase time as a general budget. Two further checks still matter before calling it “just feed the dog more often,” and they also speak to the temperature question in #2.
First, confirm what can still run while flash is busy. If the pet lives in a SysTick or other ISR, and the flash controller blocks interrupts or stalls instruction fetch from the same bank being erased, that pet will not run no matter how generously you slice the write. On Cortex-M parts the programming routine usually needs to execute from RAM or from a different bank, and the watchdog source must still be able to tick — an independent internal watchdog clock, a timer in a domain that does not stop, or a hardware feature that pauses the watchdog for the documented flash-busy window. A windowed watchdog is harder: feeding early fails, and a single “disable the WDT around the update” path is how units come back with a half-written image and no protection. Prefer A/B slots or a recovery loader that can stand a reset mid-erase.
Second, a reset at the same moment as a flash write is not automatically the watchdog. Programming current can dip the 3.3 V rail for a few tens of microseconds and trip a brown-out detector, or a reset pin can be glitched; those look like “WDT during flash” in a log that only records the reboot. Check the reset cause register, the supply at the MCU pin during erase, and whether the fail is worse at low temperature, where erase can take longer and a timeout that was fine at room temperature no longer covers the tail. If the cause is really the watchdog, size the timeout to the documented maximum erase at the coldest required ambient, with margin, and keep the pet stall-safe; then verify image integrity after a deliberately interrupted write so a partial bank cannot be booted as if the update had succeeded.