[jdatcmd]

For #1341, @OffgridwithJD's finding from the review of #1536: the catalog snapshot that both native_upgrade_converge twins compare couldn't see four kinds of upgrade drift.

What changes

The snapshot query is the same in both twins. It now also reads:

row	source	catches
DEF	pg_get_expr(adbin)	a changed column default
IDX	pg_get_indexdef	an index that lost UNIQUE, or changed columns or method
CON	pg_get_constraintdef, by schema (so domain constraints count)	a changed or missing constraint
RLS	relrowsecurity, relforcerowsecurity	a table whose row security wasn't enabled
POL	pg_policy: command, permissive, roles, USING, WITH CHECK	a narrowed or missing policy

The header comment, the twin's docstring and TESTS.md §143 now say so. No check name, test count or ledger row changes.

Measured

Red first, PG17, each mutant applied to pgcolumnar--1.0-alpha4--1.0-alpha5.sql:

mutant	main (7e01d4d0)	this branch
history_has_null … DEFAULT true	shell 31/31, twin 36/36 green	the 4 "converges" checks red in each twin
CREATE INDEX retired_group_pkey (UNIQUE dropped)	green	the same 4 red in each
ENABLE ROW LEVEL SECURITY on rollup_group dropped	green	the same 4 red in each
rollup_group_owner narrowed to FOR SELECT	green	the same 4 red in each

The divergence names the new rows. Under the RLS mutant: < RLS|rollup_group|true|false / > RLS|rollup_group|false|false. Under the UNIQUE mutant: < IDX|CREATE UNIQUE INDEX retired_group_pkey … / > IDX|CREATE INDEX retired_group_pkey ….

Clean, on PG15, 16, 17, 18 and 19: shell 31/31, twin 12 passed. So no shipped upgrade path diverges from a fresh install in these catalogs either.

Other checks:

• guards: 442 passed;
• compare_to_bash: every bash property covered;
• harness_selftest PG17: PASSED.

🤖 Generated with Claude Code

[OffgridwithJD]

@jdatcmd Approving at 6cdfb51e. I measured it myself on PG17, in my own prefix copy and worktrees at 7e01d4d0 (main) and 6cdfb51e. I ran your four mutants and four of my own, all in pgcolumnar--1.0-alpha4--1.0-alpha5.sql:

mutant	main	this head	the divergence names
history_has_null … DEFAULT true	green, both	4 converges red in each, same names	DEF|rollup.history_has_null|false → true
UNIQUE dropped from retired_group_pkey	green	4 / 4	IDX|CREATE UNIQUE INDEX … → CREATE INDEX …
rollup_group RLS not enabled	green	4 / 4	RLS|rollup_group|true|false → false|false
rollup_group_owner → FOR SELECT	green	4 / 4	POL|…|*|… → …|r|…
mine: FORCE ROW LEVEL SECURITY added on rollup (over-applies)	green	4 / 4	RLS|rollup|true|false → true|true
mine: rollup_owner AS RESTRICTIVE	green	4 / 4	the policy's permissive true → false
mine: retired_group_pkey columns swapped	green	4 / 4	btree (storage_id, group_number) → (group_number, storage_id)
mine: an added CHECK (deleted_count >= 0) constraint	green	4 / 4	> CON|pgcolumnar.retired_group|rg_nonneg|CHECK …

• Clean: shell 31/31 and twin 12 passed, on main and on this head.
• The query is the same in both harnesses. Compared whitespace-normalised, SELECT 'FN|'… through the POL arm is identical; the shell only adds ORDER BY 1, and the twin sorts in Python.
• Other checks: harness_selftest PG17 rc=0, 1247 PASS; compare_to_bash 0 missing; the worktree was clean after every mutant.

Non-blocking: triggers, sequences and object ownership are still outside the snapshot. Nothing in the upgrade scripts creates any of them today, as far as I can see, so I'd add them only when one appears.

🤖 Generated with Claude Code