had_install was computed AFTER the pre-0.7.0 migration block, so it observed
a versions/ directory that the migration had just created one line earlier. A
flat /opt/box therefore read as "already installed", host/setup-host.sh was
skipped, and the host kept every artifact the old release left behind while
box --version reported the new one — silent, and self-concealing. Computing
it BEFORE the migration asks the honest question: a tree that needs migrating
has by definition never been converged by this version's setup-host.
The accepted consequence: an unattended (BOX_YES=1) upgrade on a flat-tree
host now runs setup-host, which the #66 note cautions about. setup-host
converges and is idempotent, and a release whose host half is silently
missing is the worse failure.
Once setup-host can run at all, a second defect in the same block becomes
reachable: it went through $DEST/current, but the #66 guard holds the default
under existing boxes, so on such a host current still names the OLD version —
converging the host with the previous release's host scripts. It now runs the
installed version's own tree.
Separately, the migration named itself but not the lifecycle: the old tree
becomes a first-class 'box versions' entry the operator never installed and
cannot tell apart from a deliberate rollback target. The migration line now
names both ways out, and the closing summary re-states it, since the original
line scrolls past ~250 lines before the install ends.
test/cli.sh gains an inst_setup helper (no BOX_SKIP_SETUP_HOST) and a stub
host/setup-host.sh on the fabricated upgrade source, so the flat-tree fixture
proves end to end and fully offline that host setup runs — and that a
genuinely versioned tree still skips it.
Refs #115, #117
