fix(install): keep a declined editor on record, and redraw the harness prompts

[alifsayalee]

Fixes apimatic/contextmatic-crawler#39.

Two changes, both in the install path.

1. A declined editor stays on record

Re-installing a plugin and declining an editor it was already installed into narrowed the recorded targets without removing the copy. installed then reported that editor clean while the editor went on loading the plugin — and since update replays the recorded targets, that copy was never refreshed again.

Install only ever adds, so the record is now the union of what is on disk: what this run installed, plus the editors an earlier run installed into that this one skipped.

manifest.upsert is deliberately left alone as a plain "make this row say exactly this" primitive — uninstall relies on that to narrow a row on purpose, so merging inside upsert would have traded this bug for its mirror image.

The closing summary names the skipped editor:

  ✓   Installed into: Claude Code, Cursor
      Already installed: VS Code

This takes the second of the two fixes #39 offered ("keep the declined target on record and report it accurately"), not the first. n does not delete the copy, and I'd argue it shouldn't: the common reasons to re-run install are upgrading via the familiar verb and adding a newly-installed editor. In both you aren't trying to remove anything — so making a decline destructive would turn an absent-minded n into the loss of a working install. uninstall <plugin> --targets vscode already does that job correctly (verified: files removed, chat.pluginLocations unregistered, record narrowed).

2. The harness questions read as one flow

The ? sat in a six-column gutter four spaces from its own question, the [Y/n] hint stayed on screen after the decision, and nothing tied the questions to each other or to the result line.

Before / after:

  ?   Install into Claude Code? [Y/n] y        ◆  Install into Claude Code? (Y/n)
  ?   Install into Cursor? [Y/n] n             │  Yes
  ?   Install into VS Code? [Y/n] y            ◆  Install into Cursor? (Y/n)
      Installing into: Claude Code, VS         │  No
      Code                                     ◆  Install into VS Code? (Y/n)
                                               │  Yes
                                               │
                                               └  Installing into: Claude Code, VS Code

The keystroke goes, the question and its hint stay — the hint is what tells you which way a bare Enter would have gone. Rewriting that row needs the cursor, so it runs only on a TTY and only when the row cannot have wrapped; past the terminal width the arithmetic would clear the wrong line and take real output with it. Both fallbacks leave the row untouched.

Input is unchanged: y/n/yes/no in any case, bare Enter takes the default.

Glyphs come from code points with ASCII stand-ins (* | +), so a cp437/cp1252 console gets those rather than mojibake — the rule log.js already applies to its check mark.

Only the interactive branch changed. --targets, -y, a non-interactive shell and an injected confirm all print exactly what they did before, which is why no existing test needed editing.

Verification

• 155/155 tests, 8 new (3 for the record fix, 5 for the prompt).
• Windows 11 — real keystrokes at the prompt, Already installed: VS Code, record kept all three.
• macOS 14.5 / Node 22.14 / x86_64 — verified twice, via --targets and via the real prompt code path. Record kept all three, VS Code copy intact at 74 files, settings.json entry still present.

Not covered by any test: how a real terminal interleaves the row rewrite with readline's own echo. That was eyeballed manually rather than asserted.