I prefer polling the VBL rather than using an NMI handler for simple experiments on the NES but have observed $2002.7 not being set sometimes. The following code demonstrates this by toggling the DMC's DAC for every VBL ,which makes missed VBLs audibly obvious:

      lda   #0          ; init PPU
      sta   $2000
      sta   $2001
      
      lda   #63
loop: bit   $2002
      bpl   loop
      eor   #63         ; toggle DAC
      sta   $4011
      jmp   loop

If I use LDA instead of BIT (with appropriate saving and restoring of A), the results are the same. If I insert NOPs into the loop, the VBL is periodically missed. My current solution is to use a minimal NMI handler that just sets a "NMI occurred" flag and returns, but this is inelegant. Is there some subtlety to properly polling $2002.7?