feat(rustd): afd_db — pools by role, and the migrator that shares Zig's ledger

§2 of M176, plus the lane §2 could not be tested without.

**The lane.** M175 §6 deleted `make/test-integration.mk` and kept
`make/test-infra.mk` — the compose half. `make test-integration-rustd` is
built on what survived: same compose services, same reset, same
`KEEP_TEST_STATE=1` escape hatch. Integration tests are `#[ignore]`d, so the
unit lane still compiles and lints all five of them and runs none, which is
what keeps live Postgres off the fast path. The lane fails on a suite that
reports zero passing tests, because a selection matching nothing exits 0 and
reads as a pass — the Zig lane learned that one expensively.

**The migrator.** sqlx is the driver and nothing more: `migrate` and `macros`
are off, because the ledger is shared with the Zig daemon and is therefore a
data format. Same advisory key, same bounded `pg_try_advisory_lock` poll, same
`audit.schema_migrations` shape, same reap. What is not the format is written
as Rust: versions derive from filenames during constant evaluation rather than
being restated beside them, the reap list is a bind instead of a rendered
`IN (…)` and its stack-buffer arithmetic, an unterminated migration cannot
become a statement list at all, and the lock owns the session that holds it.

**Two findings the tests made, not the plan.**

`PgPoolOptions::connect_with` cannot say why it failed — it retries until the
acquire timeout and reports `PoolTimedOut`, the same error a busy pool gives.
Against a dead port that read as "pool exhausted", which is the page that
sends an operator to raise a limit while the database is down. Boot now opens
one connection directly (its error is the real one) before building the pool,
and acquire reads the pool's census: below its ceiling and still timing out
means the datastore, at its ceiling means capacity.

`PgConnectOptions::from_str` accepts `mysql://host/db` and reads it as host
`host`, database `db`. `parseUrl` refuses anything but the two Postgres
schemes; so does this now.

Also: `test_zig_lanes_absent` matched `make test-integration` as a prefix and
failed the new lane for existing. It now matches whole target names, with the
boundary rule itself under test.

Gates: lint-all ✓ · test-unit-all ✓ (277 rustd, 5 ignored) ·
test-integration-rustd ✓ (5 against live Postgres) · harness-verify 9/9 ·
check-version 0.26.2.