Diagnostic experiment: restore the commit d8df5d4 polling decode
algorithm/thresholds unchanged, while P3.6 ISR now only increments an
edge counter (no decode state machine). Successful decodes print the
edge-count delta over the decode window to compare against the
EV1527 theoretical 50-edges-per-frame, to validate/refute the
edge-interrupt glitch-noise hypothesis.
TH0/TL0 torn-read protection didn't cover the case where the hardware
counter already wrapped but Timer0_Isr hasn't run yet to bump
timer0_ovf_count, making the timestamp appear to jump backward by one
overflow period (~32.768ms). Read TF0 under a brief EA guard and treat
a pending-but-unserviced overflow as ovf+1 without clearing it.
