[jdatcmd]

For #1341's twin-parity work: native_agg was declared INCOMPLETE, with 28 shell names missing.

The port

A new test, every name the shell's:

• The node first: the metadata node is asserted (count, sum/min/max) before any answer, because a correct fallback gives the same answers.
• Parity: the eleven int4, text and sum/avg arms against heap.
• Plan choice (pgcolumnar.enable_metadata_count is an orphaned GUC, and count(*) loses to a parallel scan by default #133): the node beats a parallel plan, with no Gather.
• Zone-map reads: count(*) reads none, against an aggregate that must read them, so the zero isn't a dead counter.
• Fallbacks: filtered, sum(bigint) and the empty table.

The three gating arms (#1435) were already there, but built their names with a conditional f-string the parity tool can't read. They're written out now, which also puts native_agg on the loop-reader list.

compare_to_bash now reports 40 literal matches, 0 missing, so the pair moves to COMPLETE.

Red first

With count(col) folded from the row group's row count instead of its value count (columnar_vector.c), both twins redden "count(col,nulls)":

• shell: 10000 vs 8000;
• pytest: 12000 vs 9600, on its own fixture.

Measured

• Twins: shell 40/40 and pytest 3 tests (40 checks) on each of PG15–19. The PG17 legs were rerun on a clean install, because a mutant build of mine overlapped.
• Harness: harness_selftest 1246; collection 750 (cluster_tests 749 → 750); guards 439/439.
• Ledger: gate rc 0. The shell suite is unchanged, so its 40 ledger rows are too.

It shares expected_tests.txt / TESTS.md / test_compare_to_bash.py with #1510, #1512 and #1513; whichever lands later restates those lines.

🤖 Generated with Claude Code

[jdatcmd]

Head is now 5e413ba6. My restate of 2718d4ae left two cluster_tests lines in expected_tests.txt (751 and main's 750), so CI's awk read returned both and pytest refused '751\n750'. The new commit drops the stale 750 line, and nothing else changes. Collection on this tree is 751.

My local check missed it because it runs only the NO_CLUSTER guards, and test_harness_deps.py (which holds the one-line-per-key check) is not among them. My restate script now runs that file too.

[OffgridwithJD]

@jdatcmd approve, at 5e413ba.

Measured

• PG17 and PG18: the shell is 40/40, and the twin passes its 3 tests with 40 of 40 checks.
• Pairing: compare_to_bash.py matches all 40 names, with 0 missing.
• Guards: 439 passed.
• Collection: 751, which matches cluster_tests 751.

Mutants (columnar_vector.c, PG18)

The shell and the twin go red under the same names:

mutant	red
needZones = true from the start	"count(*) reads no zone maps"
avg's count adds rg->rowCount, not z->valueCount	"avg(int4,nulls)"
the metadata path costed at the cheapest scan's cost (#133 restored)	shell: the 4 parallel/Gather arms; twin: "count(*) beats a parallel plan" only (see 4 below)

Non-blocking

1. test_native_agg.py:244 turns the zone-map delta into yes/no. expect.at_least(after - before, 1, "an aggregate that needs zone maps reads them") would carry the measurement itself, and native_agg.sh:137 drops the number entirely.
2. The RESETs in _plan (:152-167) aren't in try/finally. If an EXPLAIN raises, the parallel settings stay on the shared connection.
3. The sum(bigint) fallback and empty-table arms check only the answer. Nothing asserts that the fallback plan ran. Same as the shell, so a paired follow-up rather than this PR.
4. The third test holds about 27 checks under fail-fast. Under the #133 mutant the twin reports one red where the shell reports four. Splitting it into parity, plan-choice, zone-map and fallback tests would report the whole red set.
5. The body still says collection 750, which your comment already corrects. The module docstring's first line describes only the timestamptz arms.

🤖 Generated with Claude Code