The default order of account.account joins an aggregate over the ancestor paths of every account. A company created in a test only exists in the test transaction, so the planner assumes it has a single account and reruns that aggregate for each of its accounts, including the dead rows left by earlier test classes. On runbot this turned 5s
L10n class setups into minutes and got buckets killed.

Analyzing the account tables after the chart load gives the planner the real counts, as test_balance_sheet_balanced already does. Order changes alone are not enough, and rewriting the query slows down production searches. The journal default account lookup only needs a code length, so it now uses the id order.

| Benchmark                                  | without  | with     |
|--------------------------------------------|----------|----------|
| Peru TestSequence, 25k dead rows (local)   | 120-132s | 4.3-7.2s |
| L10n build, slow account path queries      | 94       | 0        |
| L10n build, TestPosAR setup                | 145s     | 15s      |
| L10n build, TestCIIFR setup                | 109s     | 8s       |