[OffgridwithJD]

@jdatcmd This fixes my eleven open #1339 entries in the four docs tests (batch B4 of my a–m triage). Each was shown green under its defect, then fixed and shown red by name. No cluster is needed; all four files are in NO_CLUSTER.

entry	defect, planted on main	main	this branch
test_docs_join_clustering.py:38 [mislabelled]	the star-schema heading renamed	premise: the star-schema join heading is present red: a docs change reddens a premise (#1236)	the file and heading checks lose premise:; the same plant reds the star-schema join heading is present
test_docs_join_clustering.py:41 [weak]	"cluster" in one sentence, "join key" in another; or the phrase only in an HTML comment	2 passed, both plants	and it names the join key as the clustering column: got 0, wanted at least 1, both plants
test_docs_join_clustering.py:52 [weak]	best-practices clusters "on the timestamp", "join key" still mentioned elsewhere	passed	best-practices names clustering on the join key: got 0
test_docs_stripe_floor.py:90 [vacuous]	FSST removed from the advice section only	passed (the compression section names FSST)	and names what lowering past it costs: 'fsst' is absent
test_docs_cross_database_query.py:19 [weak]	the forbidden query in a ```sh psql -c block; under ```postgresql; spelled PG_DATABASE/PG_EXTENSION	passed, all three	got 1 want 0, all three
test_docs_cover_the_corpus.py:55 [weak]	test_whole_row's own TESTS.md entry removed (its name is a prefix of test_whole_row_var_on_off's)	passed	got 'test_whole_row' want 'none'
:208 [weak]	the real arm's duplicate rule mutated to count(n) > 2	the fixture stayed green (it had its own copy)	one helper, _duplicated_names; the fixture reds
:278 [vacuous]	the sweep's glob narrowed to test_d*.py	each file counted once passed (dict keys against themselves)	compared with an independent directory listing; names every missed file
:320 [vacuous]	the no-totals arm stops reading the document	the fixture stayed green (it compared two pinned tuples)	it drives the arm's predicate, totals_line_refused: got 'none' want 'refused: 9 tests in 4 files'
:407 [weak]	README's "28 refused" made "128 refused"; a second, contradicting "30 refused" added	passed, both	got '128' want '28'; got '28, 30' want '28'
:1172 [weak]	the real empty-section reader counts blank lines as a body	the control stayed green (it had its own copy)	one reader, _empty_sections; control … got '' want '999. planted.py'

• Clean: 46/46 across the four files, and the guard set passes.
• Unchanged: no test function is renamed, so TESTS.md and guard_tests don't move. None of the changed check names has a ledger or lossy_arms.tsv row.
• Speed: undocumented() now tokenizes the document once (whole words, plus x.py names), so the corpus file stays around 3 s.
• Follow-up, not in this PR: test/docs_style.sh's arm of the same name still reads only ```sql blocks, case-sensitively. test: docs_style gains its pytest twin, and a sentence no reader sees no longer passes (#1341) #1653 rewrites that file, so I'll bring the shell arm level there rather than conflict with it.

The plants and harness are in ~/pgc-tools/i1339d/ on my side.

🤖 Generated with Claude Code

[jdatcmd]

Approved at 9c90c4d. I reran your :407 plant (README's "28 refused" made "128 refused" in test/pytest/README.md): main's test_docs_cover_the_corpus.py 40 passed; this head 1 failed, "README.md quotes the number of modes actually named as refused: got '128' want '28'". Clean, the four files: 46 passed in 2.93s. CI 15/15. Agreed that the shell arm's ```sql-only reading belongs with #1653.

[jdatcmd]

Merging at 9c90c4d with CI green; approved above.