fix(cli): <plugin> config set writes the value instead of printing

[ValentinFigue]

Found by the documentation audit in #8, and deliberately left out of it because it needs a behaviour change and a test rather than a doc edit.

The bug

config)  cmd_config show "$(_mf "$_PLUGIN" config_section …)" "$@" ;;

Everything after config went to show as trailing arguments, which ignores them. So:

$ temper config set auto_nudge_lines 300
[temper]
  enabled              true      default
  auto_nudge_lines     200       default        ← unchanged
$ echo $?
0

It printed the config and exited 0 having set nothing. cairn config reset local did the same. A silent success, in the CLI of a tool whose whole argument is that reporting success while doing nothing is the dangerous failure.

Worth noting the sibling *) case does reject unknown subcommands — only config swallowed.

The fix

Forward the subcommand, and namespace the key by looking up the section the manifest schema declares for it. That lookup is the point of having a shim: you should not need to know that trailers lives in [git] rather than [cairn].

temper config set auto_nudge_lines 300     → [temper] auto_nudge_lines
cairn  config set trailers Signed-off-by   → [git] trailers      (resolved from the schema)
cairn  config set pr.base develop          → [cairn] pr.base     (dots intact)

_schema_all already emits section|key|plugin, so _plugin_section_for is a lookup rather than a second place to encode which key lives where.

Three details that decide whether this is correct, each with an assertion:

• Prefixing is what makes dotted keys work. _split_key splits on the first dot, so a bare pr.base would resolve to a section called pr. There is a test that no [pr] section is invented.
• An already-qualified key is passed through. Otherwise temper config set temper.auto_nudge_lines 300 writes temper.temper.auto_nudge_lines.
• doctor, path and edit are not section-scoped, so they forward unchanged.

Everything else comes free — trust re-hashing, global/local scope parsing, the unknown-key warning — because it lives in cmd_config, which is now actually reached.

An unrecognised subcommand exits 1 and names the real ones.

Also

The trailers doc string suggested Co-Authored-By. That string is written as a comment into every config file the tool generates, so it was quietly recommending the one trailer this project's own rules forbid. Now Signed-off-by.

Verification

16 new assertions in tests/test_cli.sh, the regression first:

✓ temper config set exits 0
✓ …and actually writes the value          ← the regression
✓ …naming the resolved section.key
✓ a key declared under [git] lands in [git]
✓ a dotted key keeps its dots
✓ a dotted key does not invent a [pr] section
✓ an already-qualified key is not prefixed twice
✓ get reads the value back
✓ unset removes the key
✓ the global scope argument still works
✓ bare temper config still shows
✓ an unknown config subcommand exits 1
✓ set with no key exits 1

726 assertions across 10 files pass.

Docs updated in the same change: the three plugin READMEs said "the plugin shim's config only shows", which was true when written and is not now, and the CLI reference gains the equivalence.