-- The recheck context of a BM25 scan survives the rescan of the scan: a correlated subquery runs the scan once for each outer row, and a rescan resets
-- the context of the scan, which deletes the contexts made inside it. The recheck context stood in it, and the next recheck (the phrase of the second row)
-- worked in a context that was gone, which another context took the place of (the stem memo of the backend, reset by a recheck: a crash in the first
-- statement of a backend, test/t/bm25_cold_stems.sh).
CREATE EXTENSION pg_sift;
SET client_min_messages = warning;
CREATE TABLE pd (id int PRIMARY KEY, body text);
INSERT INTO pd
SELECT i, s[1 + i % 5] || ' ' || s[1 + (i * 3) % 5]
FROM generate_series(1, 300) i, (SELECT ARRAY['der Hund und Katze', 'Katze die Straße', 'Kinder laufen über die Straßen', 'das Wetter war gestern schlecht', 'Haus Häuser schön'] s) x;
CREATE INDEX pd_idx ON pd USING sift (body) WITH (stemmer = 'german');
SET enable_seqscan = off;
SET max_parallel_workers_per_gather = 0;

-- one row of the outer query for each phrase and word: every one runs the scan again
SELECT q, (SELECT count(*) FROM pd WHERE body @@> q) AS n
FROM unnest(ARRAY['"Hund und Katze"', '"Katze die Straße"', 'Haus', '"Kinder laufen"', '"Hund und Katze"']) q;
-- the answers are the heap's (to_tsvector, the same stems and stop words)
SELECT q, (SELECT count(*) FROM pd WHERE body @@> q) = (SELECT count(*) FROM pd WHERE to_tsvector('german', body) @@ phraseto_tsquery('german', btrim(q, '"'))) AS same
FROM unnest(ARRAY['"Hund und Katze"', '"Katze die Straße"', '"Kinder laufen"']) q;
DROP TABLE pd;
DROP EXTENSION pg_sift;