Fix the updater, which upgraded the wrong directory and then stopped

[Slicit]

Three faults, all in v0.1.0, all found by installing it and running update.sh the way somebody would.

update.sh upgraded the wrong directory. It never told install.sh where it lived, so install.sh fell back to its default of /opt/componium. An install anywhere else upgraded a directory it had nothing to do with, created it if it was not there, and reported success while leaving the install it belongs to on the old image. The default directory is the common case and worked, which is exactly why this survived being tested by hand.

confirm ... || return ends the installer. A bare return carries the exit status of the last command, which is the confirm that just answered no, so configure returned 1 and set -e stopped everything. That is every upgrade that keeps its settings, which is what --yes always does.

[ -n "$previous" ] && PORT="$previous" is the same shape one level down: the statement returns 1 when the test is false, and each of those lines is "keep the old value if there was one". An upgrade of an install that never turned the show profile on stopped there.

A fourth, quieter one: download_bundle set an EXIT trap naming a local variable, so every run that actually downloaded ended with tmp: unbound variable from a trap firing in a scope where the local was gone.

The test that should have existed

hack/check-install.sh installs into a temporary directory and updates it, with no Docker and no network, and asserts what those faults broke:

• the update upgrades its own directory, and creates nothing in /opt
• every answer survives it
• the database password does not change, so an upgrade cannot invalidate the database it is upgrading
• the films are untouched

It runs in check-edits.sh. Both of the first two faults were put back to watch it fail.

The installer had no compiler and no test runner, and was checked by running it. Three of these only happen on the second run.

One more, unrelated and mine

The test added yesterday to hold the README's score to the parser contained a literal markdown fence, which is two adjacent backticks, which is the exact pattern check-edits.sh greps for as evidence that a shell ate a raw string. It broke that check on main for every edit after it. CI never saw it, because CI runs the language suites and the line-ending job and not that script.

🤖 Generated with Claude Code