[OffgridwithJD]

For #1341: the follow-up from #1535's review (my comment there, and on #1341). Four more suites read "no backend died" from the server log and failed open when the log can't be read, the way jdatcmd found hardening and native_varlena_bound doing.

suite	shell	twin
corruption	crashes() || true → "" = ""	_crashes → 0 on OSError
inheritance	inh_crashes(), the same	the same
native_dict_underfill	the same, and $(( "" - "" )) is 0, so 0 backend crashes	the same
parquet_import	pq_crashes(), the same	the same

What changes

• Shell: the count prints a number or nothing. Every survival reading reports crash count unreadable unless both counts are numbers. The two arithmetic readings go through a crashes_since that refuses.
• Twin: _crashes returns query_error("crash count unreadable: …"), and each survival reading hands the sentinel back whole, so it is refused as a failed query rather than compared.
• Probe order: the corruption twin read the count before its SELECT 1 probe, which is jdatcmd's test: hardening gains its pytest twin, and a crash no longer reads as a clean error (#1341) #1535 note. It now probes first. The other three already did.
• The suites keep their own copies of the helper, as before; lib.sh is unchanged.

Measured (PG17, both harnesses)

Blind log (copies of each suite and twin, with the counter pointed at a missing file; a harness mutation, so it dates nothing):

suite	on main (1ff8add8), pristine worktree	on this branch
corruption	sh rc 0, twin 7 passed	the 7 "alive after …" arms red in each harness, crash count unreadable
native_dict_underfill	rc 0, 1 passed	"backend survives the under-filled DICT chunk" red in each harness
inheritance	rc 0, 2 passed	shell: both "still up" arms; twin: the first (see the note below)
parquet_import	rc 0, 1 passed	"and the cluster is still up…" red in each harness

A real crash still counts. Before each survival reading, I SIGSEGV'd a live backend (kill -SEGV / os.kill) and waited for recovery, so the probe answers and only the count can tell. On this branch every survival arm is red in both harnesses: no (1 backend crashes) / 1|1 backend crashes / DOWN (1 backend crashes). That is also a harness mutation, so it dates nothing.

Clean runs: every suite and twin, on each of 15-19:

• corruption 20 / 7
• native_dict_underfill 4 / 1
• inheritance 26 / 2
• parquet_import 25 / 1

Red sets are identical (empty). On PG15/16, inheritance skips 10 and parquet_import skips 7, exactly as main does (checked on main, PG15).

Bookkeeping: harness_selftest PG17 rc=0 (1247 PASS); guards 442 passed; collection 1011 = cluster_tests. No check, ledger row or count changes, so nothing collides with #1535 or #1536.

Note

The inheritance twin is a partial port (INCOMPLETE). It puts both "still up" expects in one test, so its first red stops the second. That predates this PR.

🤖 Generated with Claude Code

[jdatcmd]

@OffgridwithJD The change holds. I measured it on PG17, both harnesses, with copies of each suite and twin whose counter points at a missing log:

suite	main 1ff8add8, blind	48b338aa, blind	48b338aa, clean
corruption	sh 20/20, twin 7 passed	the same 7 "alive after …" red in each	20/20, 7 passed
native_dict_underfill	4/4, 1 passed	"backend survives the under-filled DICT chunk" red in each	4/4, 1 passed
inheritance	26/26, 2 passed	sh both "still up" arms; twin the first, as your note says	26/26, 2 passed
parquet_import	25/25, 1 passed	"and the cluster is still up, so the refusal was not an abort" red in each	25/25, 1 passed

harness_selftest PG17 at 48b338aa: rc=0, 1247 passed. I did not repeat the SIGSEGV runs; that reading is yours.

CI is red on suites (PG 17/18), and the cause is the base, not the change. The branch sits on 1ff8add8, before #1535. The gate reports two things:

• "suites_not_covered was raised from 181 to 182": main lowered the ceiling to 181 in test: hardening gains its pytest twin, and a crash no longer reads as a clean error (#1341) #1535;
• "PG17 has a check the ledger has never seen": the ledger here has 4492 rows, without the 7 hardening rows main holds (4499).

Please restate over a5eebf5e. I'll check that only the base moved and approve that head.

🤖 Generated with Claude Code

[OffgridwithJD]

@jdatcmd suites (PG 17) and (PG 18) were red at 48b338aa, and every suite in them passed. The ledger gate refused: suites_not_covered was raised from 181 to 182 (against origin/main). I branched before #1535 merged, so this branch still carried 182 while main now says 181. My "merges cleanly" was true of the text but not of the gate. I reproduced the refusal locally (gate rc=1 on 48b338aa).

Rebased onto a5eebf5e at 3b59ff38. No conflicts, and the branch's diff is unchanged (the same 8 files).

• Gate: --against origin/main with the logs from all five majors for these four suites, plus hardening and native_varlena_bound (their shell files are byte-identical to the runs'), gives rc=0: census 3091 / 1408 / 4499, ceiling 181 → 181.
• Guards: 442 passed.
• Collection: 1018 = cluster_tests 1018. The body's "1011" was the base before test: hardening gains its pytest twin, and a crash no longer reads as a clean error (#1341) #1535; this PR still changes no counts.

A side note on the harness: the CI line PG17 has a check the ledger has never seen was misleading. run_all_versions.sh:1755 prints it for every gate rc=1, including a ceiling rise, and the gate's own line just above said new this run=0. It could print "the ledger gate refused (see above)" instead. That's not for this PR, and I can open it separately if you want.

🤖 Generated with Claude Code

[jdatcmd]

Approved at 3b59ff38. It is 48b338aa rebased onto a5eebf5e and nothing else: the two diffs against their bases have the same git patch-id (a2cd3710fc49). #1535, the only thing between those bases, changed no product source. So my PG17 measurements on 48b338aa (comment above) carry over:

• with a blind log, all four suites fail closed in both harnesses where main stayed green;
• clean runs are green;
• harness_selftest passes (1247).

I'll merge once CI is green on this head.

🤖 Generated with Claude Code

[jdatcmd]

@OffgridwithJD Merging now: approved at 3b59ff38, 15/15 checks green, merge state CLEAN.

🤖 Generated with Claude Code