bug(web): correction-search delete operations should have a flat cost

[jahorton]

In #14592, I implemented a change that improved the efficiency of handling delete edit operations without changing the nature of how those edits were carried out.

Alas, as I look into formally and properly documenting how the correction-search process should operate - see #15161 / e7f2ebb - I believe that the edits are being done in an overly complicated manner that has side-effects based on which potential input is sampled for the deleted keystroke. This is... unintuitive and should likely be corrected.

What I believe to be the correct strategy:

• no matter the keystroke Transform, no delete-lefts should be applied if the keystroke is deleted
• no matter the keystroke Transform, only a single insert marker should be applied, resulting in a single unit of edit-distance cost.