fix: --agent-help omits the skills subcommand group and metadata.command labels say 'plugin skills'

[christso]

Priority: high (correctness, not feature) · Surfaced during a side-by-side review against gh skill in cli/cli.

Problem

Two related defects in the agent-help surface:

1. src/cli/agent-help.ts::allCommands does not include the skills metas (`skillsListMeta`, `skillsAddMeta`, `skillsRemoveMeta`). These metas are defined in `src/cli/metadata/plugin-skills.ts` and imported by `src/cli/commands/plugin-skills.ts` for help text, but never wired into the agent-help tree. An AI agent calling `allagents --agent-help` cannot discover that `skills list / add / remove` exist.
2. Each skills meta's `command` field is labeled `'plugin skills list'` / `'plugin skills remove'` / `'plugin skills add'` — but the runtime command path is `skills list` / `skills remove` / `skills add` because `skillsCmd` is registered as a top-level group in `src/cli/index.ts` (not under `pluginCmd`). Even if defect (1) were fixed, the meta's `command` field would direct an agent to the wrong command path.

This is also visible in normal JSON output today, where the envelope's `command` field is wrong:

```json
{
"success": true,
"command": "plugin skills list", ← wrong; actual path is "skills list"
"data": { "scope": "all", "skills": [] }
}
```

Current behavior

```bash
$ allagents --agent-help | jq '.commands | map(.command)'
[
"workspace init",
"update",
"workspace status",
"plugin install",
"plugin uninstall",
"plugin marketplace list",
"plugin marketplace add",
"plugin marketplace remove",
"plugin marketplace update",
"plugin marketplace browse",
"plugin list",
"plugin validate",
"self update"
]

$ allagents --agent-help skills
Unknown command: skills

$ allagents --agent-help "skills list"
Unknown command: skills list

$ allagents --json skills list # runs, but command field is misleading
{
"success": true,
"command": "plugin skills list",
"data": { "scope": "all", "skills": [] }
}
```

Expected behavior

```bash
$ allagents --agent-help | jq '.commands | map(.command)' | grep skills
"skills list"
"skills add"
"skills remove"

$ allagents --agent-help "skills list"
{
"command": "skills list",
"description": "List all skills from installed plugins",
"when_to_use": "To see available skills and their enabled/disabled status",
...
}

$ allagents --json skills list
{
"success": true,
"command": "skills list", ← matches the actual path
"data": { "scope": "all", "skills": [] }
}
```

Verification gate (must pass before closing)

```bash
set -euo pipefail

bun run build

(1) agent-help tree includes all three skills subcommands

allagents --agent-help | jq -e '.commands | map(.command) | contains(["skills list", "skills add", "skills remove"])'

(2) Per-command agent-help works for each skills subcommand

allagents --agent-help "skills list" | jq -e '.command == "skills list"'
allagents --agent-help "skills add" | jq -e '.command == "skills add"'
allagents --agent-help "skills remove" | jq -e '.command == "skills remove"'

(3) The --json envelope's "command" field matches the actual command path

allagents --json skills list | jq -e '.command == "skills list"'

(4) No meta still labeled "plugin skills *"

! grep -RIn "'plugin skills " src/cli/metadata/plugin-skills.ts
```

All four checks must pass.

Implementation notes

• `src/cli/agent-help.ts`: import `skillsListMeta`, `skillsAddMeta`, `skillsRemoveMeta` from `./metadata/plugin-skills.js` and add them to `allCommands`.
• `src/cli/metadata/plugin-skills.ts`: change every `command:` value from `'plugin skills X'` to `'skills X'`. Update any `examples:` entries that read `allagents plugin skills X` to `allagents skills X`.
• `src/cli/commands/plugin-skills.ts`: the JSON envelope uses `command: 'plugin skills list'` etc. inside the handler bodies — update those to match.
• Add a unit test that asserts every `AgentCommandMeta.command` resolves to a real command path in the cmd-ts tree. This guards against future drift between metadata and runtime registration.

Refs

• Comparison context: `gh skill` (cli/cli) has a single source of truth for command metadata because each cobra command's `Use` field is the canonical name; allagents has separate metadata that has drifted from runtime registration.
• Companion wiki page: `concepts/allagents-vs-gh-skill.md` in EntityProcess/ai-research-wiki.

[christso]

Fixed by #380 (skill rename), which superseded this fix: the skills metas were wired into agent-help's allCommands and the command: fields + JSON envelopes were corrected from 'plugin skills X' straight to the new singular form 'skill X' rather than the intermediate 'skills X' proposed here. Net effect on the original gate's intent — metadata.command matches the runtime path and --agent-help discovers all three subcommands — is the same, just under the singular naming.

Rebasing the original fix/371-agent-help branch onto main produces an empty diff, so there is no PR to re-open.