[OffgridwithJD]

@jdatcmd The pytest twin of logical_decoding_cdc_recipe.sh (#754), every check by the shell's name, one check per test. Like the shell, it extracts the CDC recipe from docs/user-guide.md at run time and executes it, on a private cluster at wal_level=logical (pick_port(27); #1640 uses 26).

This will be restated into my stack (on top of #1641, after logical_subscriber), so its counts below are over 9d4c9416 and will move.

Holes, measured on PG17 against main's shell

hole	main's shell	this head (shell = twin)
The rollback arm read only after ROLLBACK. If the rolled-back INSERT never ran, "captures nothing" and "no row in the columnar table" both pass. Harness mutation: the INSERT made a duplicate key, so the block stops at it.	35/35	red by name: "inside the transaction the row was captured and is in the columnar table, before the rollback"
"and it saw the same values the table holds" used check on two readings. Harness mutation: server stopped before it.	that check PASSES ("" = "")	red: check_text, "a side is empty"

Also:

• the pgc_skip name is now a literal (it embedded pg_config --pkglibdir), and the twin records the same name;
• the seven recipe-contains premises are unrolled into literal names the grader can read (same names).

Mutations (PG17, shell and twin, SAME reds)

mutation	reds
nodelete: pgcolumnar_tuple_delete marks nothing	1: the row really is gone
updnulls: the update writes its new version with every column null	1: recorded as UPDATE with the NEW value (the AFTER trigger reads the stored row)
updkeepold: the update leaves the old version live (the new version then hits the primary key)	3
nodeltrig: the guide's trigger no longer fires on DELETE	2: the DELETE capture and the decoded DELETE
updasins: the guide's capture records every non-DELETE as INSERT	4

No mutation reddens a premise. I tried a "replaying the feed reproduces the table" check and dropped it: AFTER row triggers re-read the stored row, so the feed and the table never diverged without another check already going red.

Runs and counts

• Runs: head 36/36 in the shell and 37/37 in the twin (the 37th is the skip-name mirror) on each of PG15–19.
• Census: 37 rows, 7 dated and 30 never. Never 3793 → 3823, dated 2073 → 2080, 5903 rows. suites_not_covered 113 → 112, cluster_tests 2454 → 2491, TESTS.md §256.
• Gates: guards 467 passed, collection 2491, ledger gate rc=0, self-test PASSED.

🤖 Generated with Claude Code

[jdatcmd]

Blocking, and the same defect as #1640's first head: this commit adds a ledger row named after the suite's skip message.

logical_decoding_cdc_recipe  logical_decoding_cdc_recipe  test_decoding.so is not installed; build contrib/test_decoding  15;16;17;18;19  never  -

CI has test_decoding, so the suite never emits that name, and the ledger's orphan scan fails the suites jobs, as it did on #1640. To fix: drop the row and take one from the never count in the budget. #1640's 9109eb2 is the model.

The rest holds. On PG17 at main a1b04ee, with the rolled-back INSERT made a duplicate key (id 1) in a temporary copy of each harness:

run	result
main's shell	35 passed + 0 failed: the hole
this head's shell	1 failed: "inside the transaction the row was captured and is in the columnar table, before the rollback: got [ERROR: duplicate key value violates unique constraint "events_pkey" …"
this head's twin	the same name, refused as a failed query

Once the row is gone, I'll approve a head whose other files are unchanged.

[OffgridwithJD]

@jdatcmd Fixed, at 9d5b0adb, as #1640 was.

• The row: check_ledger.tsv no longer holds the pgc_skip name "test_decoding.so is not installed; build contrib/test_decoding" (grep -c = 0). The budget paragraph now says "It enters with its 36 checks …" and "29 arrive never", and gives the reason: it's a skip, not a check, CI has test_decoding and never emits it.
• The census: the stated never count equals the tsv's never count, and the stated rows equal the tsv's rows, at this head and at every head above it (table on Missing tests: 87 confirmed behaviours no shell or pytest arm exercises (#1339 audit) #1341).
• The suite: test/logical_decoding_cdc_recipe.sh, test_logical_decoding_cdc_recipe.py, the compare list and unported_suites.txt are identical line for line to what you reviewed.

🤖 Generated with Claude Code

[jdatcmd]

Approved at 4d3ca68: the skip-name row is gone, and against 9d5b0ad (which I verified) every content file is unchanged; census agrees with the tsv.