Fix nushell completion path; adopt shared install helpers

[l1a]

Description

The bug: scripts/install_completions.py wrote nushell completions to
$XDG_CONFIG_HOME/nushell/autoload. On Windows $nu.user-autoload-dirs is exactly
%APPDATA%\nushell\autoload — one entry — and nushell there never reads the XDG path, whatever
XDG_CONFIG_HOME says. So the helper wrote a real file to a directory nothing consults, printed
that path, and delivered nothing. Measured in rusticprofile (its 0.2.14), where the same defect
explained shell aliases missing for months while present in the dotfiles the whole time.

Two more in the same helper. It logged a generation failure to stderr, continued, and then ran
print("Installed completions for retch:") plus the full path list unconditionally — success
reported over work not done. And nothing checked whether zsh would ever load the file; it now asks
an interactive zsh, because a non-interactive one sources neither .zshrc nor anything it
includes and so returns the built-in fpath, getting the answer wrong just as confidently.

This repo's mechanism was right, and is now the standard

v0.6.16 moved these recipes to Python so they run natively on Windows without Git's usr\bin.
rusticprofile first proposed replacing them with plain-sh recipes because it held the
correctness fixes — which would have regressed that portability work in the name of consistency.
Each repo had solved half the problem. The shared standard keeps retch's mechanism and
rusticprofile's correctness.

scripts/install_completions.py, scripts/install_man.py and templates/justfile-common.just are
now vendored byte-identically across retch, rusticprofile and etr (TEMPLATE_VERSION = 2).
Project facts (BINS, MAN_PAGES) sit in a PROJECT header above the marked block, because etr
ships two binaries and a block with a hardcoded name cannot be copied.

just standard-check runs the helpers' --self-test; just check depends on it. Not a text
diff — three separate repos cannot diff each other's files, and a diff would pass happily on a repo
that never adopted the standard.

Also: install-tag VERSION

Installs a released tag with all three artefacts from that tag: binary via cargo install --git --tag, completions from the installed binary so they cannot disagree with its CLI, man page from
the tag via git show. It deliberately does not depend on install-man/install-completions, which
work from the checkout — reusing them would pair a tag's binary with the worktree's man page.
Prompted by a fleet host found running a current binary beside a man page eleven releases old.

Type of change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• This change requires a documentation update

How Has This Been Tested?

• just pr green: fmt, clippy --workspace and --features graphics, 69 + 114 + 15 tests,
  man regen at retch 0.6.20, audit advisory-only.
• The self-tests were watched failing, and the first is literally this bug: reverting the
  nushell line reports windows nushell dir: expected …AppData\Roaming… got …/.config/….
  Dropping a shell and treating an empty XDG variable as a location also fail.
• Exit codes checked without a pipe — an earlier check read $? from head and reported
  exit 0 on a failing self-test.
• A generation failure now aborts with no Installed line (verified with a bad binary name).
• Recipe inventory vs HEAD: exactly two additions (install-tag, standard-check), none
  removed. just --list parses.
• The vendored helpers are byte-identical (sha256) to rusticprofile's copies.

Backup safety

n/a — tooling only; no repository or filesystem writes beyond the completion/man install paths.

Checklist:

• Style, self-review, comments
• Documentation updated — wiki pushed before this PR per AGENTS §4.8
  (Development-Setup.md, 7fc584f..03796ba)
• just man re-run after the bump and committed
• No new warnings
• Tests added that prove the fix — the self-tests are the test, watched failing
• Existing tests pass
• No failure mode introduced can degrade silently — the helper raises rather than reporting
  success, and standard-check fails the build on a violation