fix(tests): run the maintenance-routine e2e against the current routine body

[nicoloboschi]

The flake

test_maintenance_routines.py::test_routines_callable_from_non_public_schema fails
intermittently with asyncpg.exceptions.DeadlockDetectedError. Not new and not tied to any
one branch — most recently on main's own run
32142821174 (2026-08-18),
and again today on an unrelated PR.

Process 4732 waits for AccessShareLock on relation 20240; blocked by process 4645.
Process 4645 waits for AccessExclusiveLock on relation 20256; blocked by process 4732.

Why it kept happening

The test installs the routines into a non-public schema and then calls one — but it loaded
the bodies from b6d2f8a4c1e7, which has been superseded. #2638's install gating is what the
test is about, and that has been carried forward unchanged. The bodies have not:

c8b4e2a71f95 ("maintenance routines skip locked schemas") gave each per-schema query a
lock_timeout and lock_not_available / deadlock_detected skip arms, for this exact
deadlock — its own commit message cites this test. The older body has no such guard: it waits
indefinitely for AccessShareLock on every schema it scans while already holding
AccessShareLock on the ones it scanned earlier. Under xdist, where peer workers create and drop
schemas continuously, a peer's DROP closes the cycle and PostgreSQL kills one side.

So the fix landed, and never reached the test that reports the bug.

Measured directly — victim schema held under ACCESS EXCLUSIVE from a second connection,
routine called from a non-public copy with an 8s ceiling:

OLD body (b6d2f8a4c1e7, what the test installed): BLOCKED
NEW body (current chain head definition):         returned

The fix

• Resolve the migration that owns the current bodies from the revision chain — via Alembic's
  walk_revisions(), so the DAG's merge revisions (tuple down_revision) resolve correctly —
  instead of hard-coding a path that silently rots the next time the bodies are replaced.
• Add test_current_routine_bodies_keep_the_concurrent_ddl_guard: whichever migration owns the
  bodies must still carry the skip arms. Fast, no DB.
• _load_schema_local_migration stays pinned to b6d2f8a4c1e7 for the two tests that assert
  what that migration's install gating emits — those don't execute anything.

test_banks_needing_consolidation_skips_schema_locked_by_ddl already proves the installed
public copy behaves correctly; this just makes the non-public end-to-end test use the same body.

Verification

tests/test_maintenance_routines.py: 24 passed. ruff check / ruff format --check /
ty check hindsight_api/ clean. Tests-only change.