Re:

On Tue, Sep 24, 2019 at 1:55 PM Vladimir Ryabtsev <greatvovan(at)gmail(dot)com> wrote:
> I inspected both values to the best of my abilities and found them equal byte-to byte.

I would try running contrib/amcheck, which should be able to confirm
index corruption, and give you a specific complaint. You may then be
able to inspect the exact index page with the problem using
contrib/pageinspect. Something like this ought to do it on Postgres
10:

CREATE EXTENSION IF NOT EXISTS amcheck
SELECT bt_index_check('vocabulary_phrase_key');

If that doesn't show any errors, then perhaps try this:

SELECT bt_index_parent_check('vocabulary_phrase_key');

Let us know what you see. We may be able to offer additional guidance,
depending on how that goes.

--
Peter Geoghegan