Missing points when bounds are scoped to the extent of points

I've noticed this (inconsistently) on several occasions: when the bounds that are used to create the bbox sent to Supercluster are created by iterating over them and extending the bounds to each point's geometry, sometimes points get left out. I have a working CodeSandbox to reproduce this: https://codesandbox.io/s/supercluster-points-8kqttt.

In this example, the points are:

const features = [
  {
    type: "Feature",
    properties: {
      isUser: true,
      userId: "5"
    },
    geometry: {
      type: "Point",
      coordinates: [-123.211605, 43.972615]
    }
  },
  {
    type: "Feature",
    properties: {
      isUser: true,
      userId: "3"
    },
    geometry: {
      type: "Point",
      coordinates: [-123.245515, 43.9150233333333]
    }
  },
  {
    type: "Feature",
    properties: {
      isUser: true,
      userId: "16"
    },
    geometry: {
      type: "Point",
      coordinates: [-123.192528333333, 44.0307166666667]
    }
  },
  {
    type: "Feature",
    properties: {
      state: "old",
      userId: "8"
    },
    geometry: {
      type: "Point",
      coordinates: [-123.194573333333, 44.0107]
    }
  }
];

The bbox that gets passed to supercluster ends up being [-123.245515, 43.9150233333333, -123.192528333333, 44.0307166666667]. The point that gets left out (ie not contained within a cluster, but not returned as an individual point either) has the coordinates [-123.245515, 43.9150233333333], which represents the southwest corner of the bbox. It would make more sense to me if both that and the point with userId: "16" were both left out, since point 16 represents the northeast corner of the bbox, but point 16 is consistently returned.

This is definitely an edge case (literally, on the edge of the bounding box, get it...), but I'm curious to understand what's going on.

Start with the linked CodeSandbox reproduction and trace how Supercluster handles the bbox whose southwest and northeast corners are point coordinates. Inspect the bbox query and clustering behavior around boundary points, then add a regression check showing that the reproduced points are consistently accounted for.