[OffgridwithJD]

@jdatcmd Nineteen of my a–m #1339 entries, in the harness self-tests. No cluster, no ledger rows (the names are pytest-only), and guard_tests is unchanged at 474.

Method: every entry was measured first, then fixed and re-measured.

• Before: a planted bug the old test passed, a one-line change to the code under test that reddened a premise:, or a docstring the code contradicts.
• After: the same plant fails the check by name.

Mutants are in ~/pgc-tools/i1339h/mut.py on my side; run.sh TAG NODE… applies one, runs, and restores.

entry	tag	before	after
test_layer.py :1013	mislabelled	nocollapse (the order-collapse scan never applied) reddens premise: the collapse is refused…	prefix dropped; same red, as a check
test_layer.py :1037	mislabelled	nobroad reddens premise: the broad except is refused	prefix dropped
test_layer.py :1067	mislabelled	noraises reddens premise: an unpinned compound raises block is refused	prefix dropped
test_layer.py :1098	mislabelled	numpass (num never fails) reddens premise: a false claim fails…, unnamed: assert {'passed': 1…} == {…}	prefix dropped, and outcomes() now names its check on failure
test_layer.py :1166	mislabelled	numfail (num fails on equal) reddens premise: an honest test passes, unnamed	prefix dropped; named
test_layer.py :837	weak	a planted expect.num(got=int(a != b), want=1, name=…) in the corpus: 1 passed	the scan reads keyword values: no test hand-rolls an inequality: got "['test_zz_plant_kwineq.py:3']"; a planted-offence arm for the keyword form
test_compare_to_bash.py :319	mislabelled	loopnone / firstword / nofstr redden the three loop-reader premises	prefixes dropped
test_compare_to_bash.py :412	mislabelled	noexpand reddens premise: with both members present…	prefix dropped
test_compare_to_bash.py :468	weak	constok (constants admitted to the expansion gate): all green. The fixture also interpolates col, so the two-names rule refused it too	a constant-only fixture: got "raised KeyError 'FLOAT_RTOL'" want '[]'
test_compare_to_bash.py :1174	inaccurate	"ONE suite each": after your parquet_export_stats port, all three remaining helpers live in one suite, parallel_copy.sh	docstring corrected (merged with your #1644 edit)
test_compare_to_bash.py :1455	inaccurate	"NOTHING ENFORCES IT … #1046 TRACKS": dropping advisory_lock_class from COMPLETE reddens two arms	docstring states the two-way enforcement
test_harness_deps.py :771	mislabelled	allbound / allfree redden the two partition premises	prefixes dropped
test_harness_deps.py :782	weak	everyfix (every conftest fixture a root): at_least(2) passed	the derived set asserted exactly: got 'pgc_cluster, pgc_conn, pgc_own_db, server_dir'
test_harness_deps.py :964	weak	pip show psycopg … || true and a comment-only mention: both passed	a code line that runs it and can fail
test_harness_deps.py :1007	weak	the guard job's flag moved into a comment: passed, since the file-wide count stayed 2	checked per job, in code lines
test_harness_deps.py :1122	re-opened, mislabelled	not vacuous any more (planting a driver install reddens the main check), but inst_empty reddens its premise:, which is a control on the test's own detector	relabelled control:
test_failed_query_sentinel.py :186	vacuous	an exclusion IS the table's reason, so "excluded with a stated reason" cannot fail; a planted dead kind frame: passed	every exclusion kind is the first parameter of a public method, so none is a dead exemption: got 'frame'
test_failed_query_sentinel.py :324	mislabelled	mintother (query_error mints another prefix) reddens premise: minted under the real prefix	prefix dropped
test_guards_pinned.py :373	mislabelled	noemptyplan (the empty-plan refusal disabled) reddens premise: all three were found…, so it is the check that catches the regression	prefix dropped
test_guards_pinned.py :1	inaccurate	"Every refusal in the layer": the file pins 17 guards; the layer raises VacuityError at more sites, several pinned in their subject's file	docstring scoped to what it pins and where the rest live

Also:

• outcomes() failed with pytest's bare dict comparison. It now prefixes the check name, so the corpus no longer has one assertion that fails without saying what.
• A comment in partition() cited a test that doesn't exist in that file. It now names test_harness_deps_classifier.py::…. Under noclose, that test is what reddens.

Left:

• The same premise: the collapse is refused when nobody rebinds exists in test_ordered.py, which is in your half.
• test_mutation_ledger.py's 13 entries are their own PR.

Gates at 9c860657: guards 474 passed; gate rc=0; self-test PASSED.

🤖 Generated with Claude Code

[jdatcmd]

Approved at 9c86065. I reran everyfix (every conftest fixture taken as a root, in cluster_fixtures) against both versions of test_harness_deps.py:

• main: the root-fixture test passes;
• this head: it fails with "conftest's cluster-reaching fixtures are derived exactly: got 'pgc_cluster, pgc_conn, pgc_own_db, server_dir' want 'pgc_cluster, pgc_conn, pgc_own_db'".

Clean at this head, the five touched files with --pg-config: 152 passed. CI 15/15.

Note: pinning the exact set means a new connecting fixture in conftest now reddens this check by name. The old docstring said that needed no edit here. Your comment states the trade, and I agree with it.

The test_ordered.py collapse premise you left for me is relabelled in my #1667.

[jdatcmd]

Merging at 9c86065 with CI green; approved above.