Zero-width DOM selection bridges must normalize both directions

What happened

The first slate-dom-v2 zero-width fix only patched the write path:

• Slate point or range
• to DOM point or range

That avoided out-of-bounds native offsets when the zero-width sentinel text was missing.

It still left the read path wrong.

If the browser handed back a native offset of 1 inside a zero-width marker, toSlatePoint returned Slate offset 1 for an empty leaf.

That is garbage.

What fixed it

The bridge needed symmetric ownership:

• toDOMRange clamps zero-width native offsets to actual DOM reality
• toSlatePoint maps zero-width native offset 1 back to Slate offset 0

The key rule is simple:

• zero-width sentinel behavior is DOM implementation detail
• it must not leak into Slate offsets

Reusable rule

For DOM selection bridges around zero-width leaves:

• normalize write-path offsets
• normalize read-path offsets
• test the round-trip explicitly

If only one direction is normalized, the bridge still leaks browser sentinel semantics into the editor model.