BUG #14623: pg_trgm doesn't correctly process some regexp with negative lookahead

The following bug has been logged on the website:

Bug reference: 14623
Logged by: Jeff Janes
Email address: jeff(dot)janes(at)gmail(dot)com
PostgreSQL version: 9.6.2
Operating system: Linux
Description:

I'm reporting this bug based on
https://dba.stackexchange.com/questions/170669/postgres-function-index-doesnt-work-correctly-with-regular-expressions/170897#170897

This simplified case is easy to reproduce back to at least 9.4, and is still
present in 10dev HEAD. Matches are missed when using the trgm index, but
not when doing the full table scan.

create table foobar (x text);
insert into foobar values ('eldrazi scion'),('eldrazi scio');
create extension pg_trgm ;
create index on foobar using gin (x gin_trgm_ops);
select * from foobar where x ~ 'eldrazi (?!scion)'; -- returns 1 row
set enable_seqscan TO off;
select * from foobar where x ~ 'eldrazi (?!scion)'; -- returns 0 rows

Building a gist rather than gin index exhibits the same bug.

The trigrams seem to be extracted correctly, but the graph stuffed into
extra_data is not correct. Looking at /tmp/packed.dot, there are no arrays
pointing to the successful termination state s1. Instead, I get lead to a
dead-end state s7.

If I change the regexp to 'eldrazi (?!s)', then bug goes away, and
/tmp/packed.dot shows the correct graph pointing to s1.