[OffgridwithJD]

Stacked on #1465 (merge after it). Two commits: 189cf7bc (the twin and the shell fixes) and 93d297d8 (the ledger seed). This is the second suite of #1341 item 70.

What was wrong

Every arm in the suite could already go red, but three went red for the wrong reason, so neither the log nor the ledger could say which arm a mutation actually broke.

• Cumulative snapshots. Each cleanup arm compared the catalogs with the suite's starting $base, so one leak reddened every arm after it. With the cleanup gated on the columnar AM, the old shell reddened five arms, among them "dropping a heap table that was never columnar leaves the catalogs alone", on aac_t's leftover rows. Each arm now compares with the snapshot taken just before its own table, and the same mutation reddens exactly the two arms it breaks.
• REFRESH inherited CREATE's failure. "and REFRESH MATERIALIZED VIEW does too" refreshed the CREATE arm's matview, and under the gate mutations it went red with relation "rwm" does not exist, the cascade the suite's own comment warns about. It now refreshes a matview of its own, made WITH NO DATA (a premise, green under every mutation), and goes red on the hook's own error.
• A premise a product change reddens. "premise: and took pgcolumnar.options with it" went red with the member skip removed, because DROP EXTENSION aborted. It's an arm now (test: 87 arms have survived every mutation aimed at their neighbours: vacuous, or named for something they do not measure #1236).

Mutations of the drop hook (PG18), new shell and twin

mutation	reds (shell)	twin's first red
cleanup gated on the columnar AM (#304's shape)	the converted-table arm, the pre-extension options row	same
gate on the schema, not pgcolumnar.options	the 7 after-DROP-EXTENSION paths	DROP after DROP EXTENSION
no gate	those 7, plus the 8 no-extension paths	one of each
skip the whole schema, not the members	the in-schema table	same
no member skip	DROP EXTENSION, "and took pgcolumnar.options with it"	DROP EXTENSION
no declaration delete	the converted-table arm	same
no options delete	4 cleanup arms, each against its own snapshot	the converted-table arm

The twin

test_alter_am_cleanup.py carries all 47 checks by the shell's name (compare_to_bash: 47 literal, 0 missing). It's four tests: conversion cleanup, no extension, pre-extension table, and after DROP EXTENSION. Each runs in a database of its own that the fixture drops. A statement's failure comes back as rc=1 ERROR: <message>, as in the shell.

Runs

Head is 47/47 in both twins on each of 15, 16, 17, 18 and 19.

Ledger

• The suite had no rows. All 47 checks enter, 21 dated by the seven mutations.
• Twenty-six arrive never:
  • the sixteen premises;
  • the conversion and round-trip reads;
  • the CREATE TABLE arms and the control;
  • the never-columnar heap arm;
  • the pre-extension DROP's exit code;
  • the two TRUNCATE arms. TRUNCATE drops no relation through the hook, so no gate mutation reaches them.
• suites_not_covered 210 → 209. The census is 3900 / 2844 / 1056.
• pgc_ledger.py gate returns rc 0.

Counts and guards

• cluster_tests 683 → 687, and pytest collects 687.
• COMPLETE gains alter_am_cleanup, and TESTS.md gains §154.
• The NO_CLUSTER guards pass (438), harness_selftest.sh passes, and docs_style.sh passes.

🤖 Generated with Claude Code

[jdatcmd]

Approving 93d297d.

Measured on PG17, at 93d297d: alter_am_cleanup.sh 47/47, and test_alter_am_cleanup.py 47 checks over 4 tests. The tsv is 3900 / 2844 / 1056, as described.

Your per-arm snapshot claim, with the first mutant run myself: the drop hook returns early, after the storage-tree delete, unless the relation is still columnar (#304's shape).

shell	arms red
main's	5: the converted-table arm, the round-trip, "dropping a heap table that was never columnar", the in-schema table, and the pre-extension options row. Each of the three middle ones shows aac_t's leftover 0/0/0/1/1.
this PR's	2: the converted-table arm and the pre-extension options row, exactly the two it breaks

I took the other six mutants from your table.

[jdatcmd]

Re-approving 62c80a4, the restate onto main db56168. I compared every commit's patch with the approved stack using git range-diff. The test and source commits are unchanged, and the ledger-seed commits changed only by re-anchoring on #1470. The census at this head matches its budget paragraph, and the counts step consistently. At #1467's head: the guards pass (438), 687 collected (matching the file), and harness_selftest is 1245/1245.