A rattle is not a tilt

A rattle is a real change

Carry the block, or tap the table, and the balls bounce. Each time they lift off the contacts the switch really is open, for a moment. The first read prints those moments; a counter would count them; an alarm would go off.

Press Play. The block sits with its left edge down, switch closed, and is knocked; the top lane is the balls, lifting off six times in 45 ms. The timing is a picture, not a measurement: nobody has put this switch on a scope. The middle lane is the part that is worked out, from the parts on the board.

What the capacitor hides

When the balls lift off, nothing is driving SIGNAL any more, but the 100 nF is still full. It empties through the 1 kΩ and the LED quickly down to about 2 V, where the LED stops conducting, then through the 10 kΩ alone, with a time constant of 1 ms. Your pin reads LOW only once SIGNAL falls under its own threshold:

• ESP32, ESP32-S3: under 0.825 V, about 1.1 ms after the balls lift.
• Pico: under 0.8 V, about the same.
• Uno: under 1.5 V, about 0.55 ms, because an Uno calls LOW at a higher voltage.

A gap shorter than that never shows. Switch the figure to the Uno and two more of the six get through. It works in one direction only: when the balls land, they fill the capacitor through the contacts in a microsecond, so a brief closing from a level block is not hidden at all.

What the sketch has to do

Anything longer than a millisecond is the sketch's problem. The fix is the one every contact needs: believe a new level only once it has held for a while. A rattle lasts milliseconds, a real tilt lasts as long as the block stays tipped, so 50 ms sits comfortably between them. The four-direction block's book uses the same 50 ms, as a count of five scans; counting tips does it with a clock.