Revise gidx_merge behavior with unknown

[komzpa]

Spin off #4139

In gserialized_gist_nd there are two conflicting clauses:

• gidx_unknown is a gidx with 0 dimensions, and considered to include nothing;

• an unshared dimension is considered to be "true" for any spatial relation.

Also volume/edge of unknown gidx is considered 0 when it could be infinity which may lead to better trees.

Evaluation whether we need to special case gidx_unknown is needed

[pramsey]

We do need it, because otherwise there's no way to handle indexing empty geometries (correctly). In fact I think (have to check) our 2d index still chuffs on empty because it lacks a good way to flag an empty geometry.

[komzpa]

I think for WKB we already do POINT EMPTY as POINT (NaN, NaN). If we apply same thinking here, GIDX(NaN NaN, NaN NaN) will be a box for such a thing, and it will give the relation we want with all the functions:

• gidx_merge(GIDX(NaN NaN, NaN NaN), GIDX(1 2, 3 4)) → GIDX(1 2, 3 4)
• overlaps(GIDX(NaN NaN, NaN NaN), GIDX(1 2, 3 4)) → False
• overlaps(GIDX(NaN NaN, NaN NaN), GIDX(NaN NaN, NaN NaN)) → False
• intersects(GIDX(NaN NaN, NaN NaN), GIDX(1 2, 3 4)) → False
• intersects((GIDX(NaN NaN, NaN NaN), GIDX(NaN NaN, NaN NaN))) → False

And there will be no need to special case the undefined/empty across all functions. Any invariant that breaks this way?

[komzpa]

Milestone renamed

[pramsey]

No, I don't think that works, since then you end up with EMPTY in the index that are basically non-findable. An empty key is not consistent with any node, so an index search always returns no results. The index needs a place to put empty things…

[robe]

Are we still doing anything with this or shall we close as wontfix?