[OffgridwithJD]

@jdatcmd The last suite on item 74's list, stacked on #1497 (#1489 → #1493 → #1496 → #1497 → this). This PR's own commits are the top two.

The twin, name for name

All 32 checks by the shell's name, in three tests. Under each of nine mutations of the directory walk and glob expansion (columnar_parquet_reader.c), the twin reddens exactly the shell's set:

mutation	reds, shell = twin
an empty directory resolves to nothing	2
an unmatched glob resolves to nothing	1
an unstat-able path is dropped	2: both dangling-symlink arms
a FIFO is a candidate	1, and the read blocks
_/. names are walked	2
a symlinked directory is descended	2
the files are unsorted	1
no depth bound	1
every extension is read	22

On the "any error counts" arms: I checked whether removing a refusal let a downstream error keep them green, as on #1497. It doesn't here: with the empty-directory refusal gone, "empty directory errors" is red.

Blocking reads. The shell bounds the FIFO and symlink-loop reads with timeout around psql. The twin runs them on their own connection, in a thread with a deadline. A cancel is no substitute, because a backend blocked in open(2) on a FIFO resumes the open after a signal (SA_RESTART). On a timeout the twin opens the FIFO for writing, which releases the reader, so the mutant run finishes and the cluster tears down. The FIFO arm is its own test, so it can decline under the shell's check_skip name, "the FIFO case", where mkfifo is unavailable.

Five-major runs at head: 32/32 for the shell and 32/32 for the twin, on each of 15–19.

Ledger

Census 4310 / 3010 never / 1300 dated (on #1497's 4278 / 3006 / 1272), with suites_not_covered 195 → 194. 28 of the 32 checks are dated. The gate gives rc 0, and orphan-scan finds 0. 4 arrive never: the three glob arms (subset, directory match skipped, only directories) and the schema mismatch.

Collection gives cluster_tests 737 -> 740, and TESTS.md gets §171. The grader is 34/34 by name, and the guards pass 438/438.

That clears item 74's list. Next I'll take the #1339 custom-scan audit entries I posted, unless you've started on them.

🤖 Generated with Claude Code

[jdatcmd]

Approve, at 40ebcfc (its own commits 1871e47 and 40ebcfc, on #1497). This time I ran both twins: pyarrow is now in my pytest env.

My own mutant, different from your nine: the glob path drops its candidate filter (if (true || pq_path_is_candidate(...)) after glob()), so a directory that matches is read as a file. The directory walk keeps its filter. PG17:

• shell: glob skips a directory match, reads the files: got [QUERY_ERROR.1] want [225b83ef…];
• pytest: the same arm, red: … is not a regular file ….

That row arrives never in this PR, and this mutant dates it, if you want to fold it in.

A finding, not blocking. "glob matching only directories errors" stayed green under the same mutant. Its error came from reading the directory (is not a regular file), not from the refusal it names, pattern … matched no regular files. It's the "any error counts" shape you checked for on the empty-directory arm. Requiring matched no regular files (or its SQLSTATE) would make it see the refusal. Since it's one of the four that arrive never, the follow-up can date it too.

[jdatcmd]

@OffgridwithJD, on your next step: I haven't started on the #1339 custom-scan entries you posted, so they're yours. So we don't overlap, here's where I am:

• Open from me: test: differential's float sum/avg arm asks for two numbers, and its text-bloom C arm probes a value that is there (#1339) #1490 (differential, to be re-anchored after your stack) and test: check_records proves the count is the record stream, and covers every recording method (#1339) #1495 (check_records, needs your re-review at 62d02f3).
• Next for me: the native_parquet_pushdown row-group-edge follow-up from test: native_parquet_pushdown gets its pytest twin (#1341) #1496, and the harness-only pytest audit: 260 confirmed checks that cannot fail, prove something else, or carry the wrong label #1339 entries (test_loop_coverage_premise, test_mutation_ledger, test_suite_accounting, test_raises_sqlstate, test_assertion_carries_its_measurement).

If you want any of those, say so and I'll leave it.

[jdatcmd]

@OffgridwithJD, merging #1498 now: approved at 40ebcfc, 15/15 green, CLEAN.