select-install: backup round-trip suites need seaweedfs/bucket/backup-controller in their PackageSource closure #3508

Prerequisite for the hack/select-install.sh / test-minimal work in #2847 (Track A, whose own issue #3277 is closed).

packages/core/platform/sources/clickhouse-application.yaml declares dependsOn of only cozystack.clickhouse-operator + cozystack.cozystack-engine. But the chainsaw backup round-trip in that suite (clickhouse-2-backup-roundtrip) additionally needs seaweedfs, the bucket app and the backup/backupstrategy controllers: it provisions an apps.cozystack.io/Bucket, reads its COSI BucketInfo, copies the seaweedfs CA, and drives BackupJob / RestoreJob.

The same is true of mariadb-2-backup-roundtrip and postgres-2-backup-roundtrip — this is not clickhouse-specific.

Why nothing is broken today: CI installs the full platform, so select-install.sh never actually narrows the install; the closure it computes is unused. select-install.sh --validate only checks that every suite maps to a PackageSource, not that the mapped source's closure covers what the suite exercises — so it stays green.

Why it matters later: the moment test-minimal installs only a suite's computed closure, these three round-trips get a cluster with no object store and no backup controllers, and fail on a missing Bucket / CRD rather than on anything they mean to test. That failure will look like a broken test, not a missing dependency edge.

Options (a decision for whoever lands test-minimal, which is why this is an issue and not a PR):

• add the backup-flow packages to dependsOn of the three *-application sources — simple, but overstates the dependency for everyone who does not run the round-trip (the plain clickhouse / mariadb / postgres Tests do not need them);
• model per-suite rather than per-source dependencies, so a suite dir can declare extra packages — the "modelling multi-package suites is deferred to the test-minimal work" note already in hack/select-install.sh;
• or teach --validate to fail when a suite's closure omits something the suite demonstrably uses, so the gap is caught rather than documented.

Surfaced by IvanHunters while reviewing #3385 and explicitly noted there as pre-existing and not introduced by that PR.