[jdatcmd]

Fixes #1420. A harness-only change; no product code or suite changes.

The defect

pgc_conn and the cursors it hands out are proxies (_WatchedConnection, _WatchedCursor in pgc_vacuity.py). They forward reads through __getattr__ but had no __setattr__. So pgc_conn.autocommit = False created an attribute on the proxy and left the real connection in autocommit. Every statement committed, rollback() undid nothing, and nothing said so. I hit this writing #1419.

The fix

Both proxies get a __setattr__. Names starting with _ are the proxy's own state (_conn, _cur, _nodeid, _pgc_write) and are set on the proxy. Any other name raises AttributeError. The message names the attribute and says to use BEGIN/ROLLBACK, or a connection of the test's own.

Raise, not forward: #1420 left this choice open. I took "raise" because a test that wants a transaction should say so in SQL, as the existing twins do. Forwarding is a one-line change if you prefer it, @OffgridwithJD.

Evidence

Red first, against the unchanged proxies:

• test_a_setting_on_the_watched_connection_is_refused_not_dropped: got 'set', want 'refused'.
• test_a_setting_on_a_watched_cursor_is_refused_not_dropped: got 'set', want 'refused'.
• test_setting_autocommit_on_the_connection_the_tests_use_is_refused (the wiring arm, on the real pgc_conn): got 'set', want 'refused'.

Nothing in the corpus depends on the old behaviour. An AST scan of all 114 files under test/pytest finds no assignment of a public attribute on either proxy. A grep couldn't answer this, because it matched SQL text inside strings.

Green, on PG17:

• Guard set: 435 passed with --pgc-expect-tests 435.
• Cluster half: 567 passed with --pgc-expect-tests 567, rc=0. It runs serially, as CI does. Under -n 4 the same tests pass, but the run exits 67: each worker checks the decline registry (expected_unrunnable.txt) against only the tests it ran, and test_format_downgrade declines under xdist. That happens with or without this change.

Counts and docs:

• guard_tests goes 433 -> 435 and cluster_tests 566 -> 567, both by collection.
• TESTS.md §7 and §22 gain a row for each new test.
• No ledger change: the ledger covers suites, and these are harness guard tests.

🤖 Generated with Claude Code

[OffgridwithJD]

Approved at 28f2e41. Raise rather than forward is the right call: a test that wants a transaction should say so in SQL.

• The fix: both proxies get a __setattr__. Names starting with _ set the proxy's own state, which covers everything their constructors set; any other name raises AttributeError with a message that says what to do instead.
• Red, one proxy at a time (PG18, __pycache__ cleared before each arm, the runtime checked for which class still had __setattr__):
  • Without the connection's: 2 failed, 1 passed. Both test_a_setting_on_the_watched_connection_is_refused_not_dropped and the wiring arm on the real pgc_conn fail with setting autocommit on the watched connection is refused, not kept on the wrapper: got 'set' want 'refused'. The cursor arm stays green.
  • Without the cursor's: only the cursor arm fails, setting an attribute on a watched cursor is refused, not kept on the wrapper: got 'set' want 'refused'.
  • Restored: 3 passed.
• Nothing else depends on the old behaviour: a grep for assignments to the psycopg connection and cursor attributes (autocommit, row_factory, isolation_level, read_only, arraysize, prepare_threshold and so on) across test/pytest finds only the new tests and the fake _Conn's own __init__. That agrees with your AST scan.
• Counts: collection gives guard 435 and cluster 567, matching expected_tests.txt.
• CI: 11 of 15 checks green at the time of review, 4 still running.

🤖 Generated with Claude Code