feat(cli): tab completion for bash, zsh and fish

[jvsena42]

loopky had no tab completion. It does now, for bash, zsh and fish.

eval "$(loopky completion bash)"                                 # in ~/.bashrc
loopky completion zsh > "${fpath[1]}/_loopky"
loopky completion fish > ~/.config/fish/completions/loopky.fish

Commands, subcommands, every flag, and the values of the flags with a closed set — --env,
--separator, the shell name. import completes filenames; so do --from-file and --qr-out.
The .deb and the Homebrew formula install all three for you.

Generated from one table, not checked in as three scripts

A completion cannot fail loudly. It offers a flag the parser refuses, or stays silent about one
that works, and the user shrugs and types it out — so three files that were right the day they
were written go wrong at the first flag added afterwards, with nothing reporting it.

CommandSurface.kt describes the commands, their options and what each option accepts; the three
generators read that and nothing else. CompletionTest holds it against Args.SWITCHES and
against the usage block in both directions, so a flag that nothing describes fails the build. The
.deb and the formula generate their copies by running the binary they are packaging, for the
same reason.

Four things not to undo

• Nothing it offers touches the network. Completing a deck id after deck show would put a
  homeserver round trip on a keypress — a tab that hangs for a second on a good network and
  forever on the hourly-expired session (Every homeserver write dies on /session with a raw transport error, and publish reports it as "check your connection" #165), in the one place you cannot interrupt without
  losing the line. Deck ids are Operand.Opaque: zsh and fish name the word and offer nothing. A
  test asserts no script contains an address.
• completion runs before Koin, beside update. Generating a static string must not depend
  on libpubkycore loading, so a host outside the shipped matrix can still install completions.
• The update check is off for this verb. The script's documented home is a line in .bashrc,
  which puts this command in the path of every new shell; the check is awaited before exit, and a
  terminal that opens a second slower once a day is a bug nobody would trace back here.
• Each script walks the line rather than counting words. loopky --env staging deck create is
  legal, and "the first one or two words" reads staging as the command. fish's own
  __fish_seen_subcommand_from has the same flaw one layer up — it matches a word anywhere on the
  line, an option's value included — so the fish script computes the command itself.

--yes and --force stay parseable and are deliberately not offered: nothing here prompts, so
advertising them would describe a confirmation step that does not exist. UNOFFERED_SWITCHES
records that as a decision the test can tell apart from an omission.

The installer only mentions the command. Enabling completions means writing to a shell rc file or
a system directory, and an installer that edits ~/.bashrc behind you is one you cannot cleanly
undo — on a box where the whole install is "copy one file", that is the wrong trade.

Verified by driving the shells, not by building

Full table in journeys/RESULTS.md. bash locally, zsh and fish in alpine:3.20 containers, since
neither is installed on this machine. Every case passed, including the two that motivated the walk:
loopky --env staging deck <TAB> and loopky card add abc --front x --ba<TAB>.

Driving zsh found a bug a green build could not. The script ends with _loopky "$@", which is
correct when zsh autoloads it out of $fpath — and wrong for the eval "$(loopky completion zsh)" install the same file documents, where it runs _loopky outside any completion context and
prints "command not found" for _describe on every new shell. It now branches on funcstack[1]
and registers with compdef on the sourced path; both documented install lines work as written.

CI gains a step that runs bash -n / zsh -n / fish -n over the three scripts and sources the
bash one — the one class of failure a Kotlin test cannot see, and the one that lands in a user's
.bashrc as a syntax complaint about a file nobody wrote.

./gradlew :cli:test :detektAll green. Reasoning in docs/Architecture.md §13.13; usage in
cli/README.md. No app UI is touched, so no journeys/*.xml applies.

🤖 Generated with Claude Code

https://claude.ai/code/session_0116NZ92Pe7aGQpqtiRkP4Ye