ColliderSet

The grid is updated for these colliders and for no others, which is the whole point: a consumer whose world streams was rebuilding a set at every region crossing, at about 2.1 µs a box, and paying a whole frame for it.

Slots come off the free list before the arrays grow, so a set that loads a ring of regions around a player settles at the ring's high-water mark instead of climbing with distance travelled. Live indices are untouched by this: a slot already in use stays where it is.

Copy and offset instead of recompute. One Float32Array.set moves the packed bounds, the shapes are copied by reference, and each of the source's buckets is appended to this set's with a constant added to every index. Nothing is floored, no shape bounds are recomputed, and the result queries identically to a one-shot build of the union.

It appends past everything and skips the free list, because a constant offset is what makes the bucket remap a copy: source slot i has to land at base + i and nothing may disturb that.

Which is exactly why the source's holes come across as holes. A source that has had remove called on it — every streamed set — cannot be compacted on the way in, because skipping its holes would make the offset a function of position. Copying them while calling them live would be worse: those slots carry the bytes of colliders that were deleted, and a set reporting them would have resurrected geometry a consumer believed was gone. So liveness is copied with the bytes, the holes land on this set's free list, and the next add fills them. Requiring a dense source instead was the other option and was rejected: it would refuse the only input this was asked for.

What a churned source does cost is its dead slots as bytes copied and space held until the free list drains them. Bounded, visible in bytes(), and not a leak.

A copy and not a move. The source is unchanged and stays usable, which is what lets a caller keep a prototype region and stamp it more than once.

join(a, b, c) is this in a loop, which is why there is no second method for it.

Refuses the base group and a group it does not hold; a streamer double-freeing a region is a bug, and a silent no-op hides it until the memory it was meant to release turns up in a budget.

data is public and the stride is not. A reader outside this file that decodes data[i * 6 + 1] by hand keeps working right up until the packing changes, and then reports confident nonsense rather than failing.

Because a consumer can count its own bytes and cannot see the engine's, which makes a streaming budget half-measurable: the question "does the ring of regions around the player fit on a phone" has no answer without this.

O(n) over slots and over the index, and it allocates a Set to count each distinct shape once. Not a per-frame call.

The one defect in this structure that has no symptom of its own, which is why it gets a check of its own. A remove that misses a bucket entry leaves the slot marked dead — so the fingerprint skips it and is right, and nothing walks it by index — while the stale entry goes on naming it: query returns it, and the sweep reads its stale bounds and collides with geometry that is gone. No measurement of frame time, query cost or hash can see that.

O(total entries) and not something to call in a frame. For tests and for the streaming check.