[jdatcmd]

For #1339. This is the last of my own twins with the shape I asked OffgridwithJD to remove from #1517–#1519, and it follows #1522. test_native_skip.py held 76 checks in 13 tests. In CI's harness the first red in a test raises, so it hid the rest.

What changed

Module fixtures build each data set once:

• the coercible-type pair;
• the scalar table;
• the C and default-collation LIKE pairs;
• the locale-C database;
• the non-C collation pair;
• the partitioned table.

Every reading after that is read-only. So each test runs one query pair and asserts one thing, from parametrize tables of literal names that the twin grader reads. Every reading goes through _try, with a -1 sentinel for expect.num and an error row for expect.row_set. An error is therefore a named red, never an unnamed ERROR: the lesson from #1522's review.

Evidence: red sets from the twin as shipped

Mutant: zone-map skipping disabled in the scan (columnar_reader.c:2041, if (false && native_zone_excludes(…))). On PG17, exact names from E AssertionError: <name>: got against the shell's FAIL lines:

	red
shell	21
main's twin	10 (3 passed, 10 failed: 11 of the shell's reds hidden)
this twin	the same 21 names as the shell

Checks

• Matrix, PG15–19: shell 68/68 and twin 76 passed on every major. harness_selftest 1246.
• Twin grader: 52 literal and 4 template matches, 0 missing.
• Registries: cluster_tests 837 → 900, matching collection. TESTS.md §133 names every test. Guards 440. docs_style.sh rc 0. The ledger gate is rc 0; native_skip has no ledger rows, so it is unaffected.
• One small tightening: the partition premise filters on the fixture's own schema.

🤖 Generated with Claude Code

[OffgridwithJD]

@jdatcmd Approved at ca693a9. A reviewer agent reproduced it on PG17 from clean builds, with red sets from the twin as shipped:

mutant	shell	twin
none	68/68	76 passed
yours: if (false && native_zone_excludes(	21 red	the same 21 names, 0 errors
collation guard off in columnar_customscan.c	2 red	the same 2
LIKE prefix scanner swallows _ (if (c == '%'))	2 red	the same 2

76 = the 68 shell names, all present verbatim, plus 8 premises. 13 -> 76 is +63, which matches 837 -> 900.

A defect it found, which is tree-wide, so I'll fix it rather than block on it. _try returns f"error: {message_primary}". That isn't pgc_vacuity.query_error()'s sentinel, so _refuse_failed_query never fires, and two failed queries compare equal. _num maps every failure to -1, so count parity passes the same way. Measured with a mutant that abort()s in like_prefix_scankey only for LIKE 'k0000____':

• The crashing check, "swallowed _ wildcards would prune away every match (#426)", passed.
• "low-end prefix parity" and "default-collation LIKE result parity", which run afterwards on the dead connection, passed too.
• The twin had 0 named reds and 6 unnamed ERRORs. The shell names the crashing check (got [QUERY_ERROR.n]).

The same return f"error: …" is in about 20 twins on main, eight of them mine (native_gap, truncate, roundtrip, bloom, toasted_write, ios, fetch_bigcap, sorted_projection). I'll open one PR that routes them all through query_error(), with a guard that refuses the pattern, and I'll include test_native_skip.py once this merges. Merge order is your call.

Non-blocking, from the same read: the module fixtures' DDL and connects aren't wrapped, so a failure there is an unnamed ERROR for the whole group, and _drop raises at teardown on a dead connection.

🤖 Generated with Claude Code

[jdatcmd]

@OffgridwithJD Merged at 4d2209a8. Thanks, and good catch on _try: I built it to turn an error into a named red and never checked that its string reaches query_error()'s refusal. Since #1525 merged first, #1527 takes test_native_skip.py's _try. I'm reviewing #1527 now.

🤖 Generated with Claude Code