feat(skills): skill quickstart + agentforge-cli (conversational ops surface)

[mandarnilange]

Summary

Two related additions for first-time and day-2 users:

1. Skill Quickstart — docs/skill-quickstart.md

A 4-step, fill-in-the-blanks markdown for users new to the skills. No
YAML, no scaffolding by hand — just:

1. Install the skill (npm install @mandarnilange/agentforge-core,
   npx skills add mandarnilange/agentforge)
2. Open your AI coding agent (Claude Code / Cursor / Codex / Aider /
   Continue.dev — same flow for all)
3. Tell the skill what you want — four sample prompts (Classic SDLC,
   Content generation, Code review, Something new)
4. Run the pipeline the skill built — either type the commands yourself
   or let `agentforge-cli` do it for you

Day-2 follow-up prompts shown inline so users see the full surface,
not just day 1. Pitched at the `@mandarnilange/agentforge-core` install
path since that's the primary user audience.

2. New skill — `agentforge-cli` (0.1.0)

Conversational interface to the AgentForge CLI. Maps natural-language
asks to the right command, executes via the host agent's shell, and
summarises the output instead of dumping raw tables.

User says	Skill runs
"Run my pipeline for project X"	` run --project X --input ...`
"Approve the architect gate"	` gate approve `
"Show me current pipeline runs"	` get pipelines` (read-only)
"Show me logs for run Y"	` logs Y`
"Apply this YAML to the cluster"	` apply -f ...`
"Start the dashboard"	` dashboard` (backgrounded)

Discipline matches the rest of the family:

• Read-first / confirm-before-mutate. Read-only commands
  (`list`, `info`, `templates list`, `get *`, `describe node`,
  `logs`, `dashboard` start) execute immediately. State-mutating or
  cost-incurring commands (`run`, `run --continue`, `exec`, `gate
  {approve,reject,revise}`, `apply`, `node start`) require explicit
  confirmation via `AskUserQuestion` or equivalent, with cost ceiling
  surfaced when visible from agent budgets.
• Never invent flags. Every flag in `references/command-map.md` is
  sourced from `packages/{core,platform}/src/cli/commands/` verbatim.
• Long-running commands backgrounded. `dashboard` and `node
  start` don't block the conversation.
• Human summaries by default. 3–5 line "what state is in, what to do
  next" instead of raw tables.

Skill family is now four:

Skill	Audience	Role
`agentforge-workflow`	End user	Design (writes `.agentforge/`)
`agentforge-cli`	Operator	Operate (run, monitor, manage) — new
`agentforge-debug`	Operator	Triage (stuck/failing runs)
`agentforge-template-author`	Contributor	Contribute upstream

Same install command picks them all up:

```bash
npx skills add mandarnilange/agentforge
```

Versioning

Per the bump policy in `skills/CHANGELOG.md`, only `agentforge-cli` is
new (0.1.0). The other three skills had zero changes to their
`SKILL.md` or `references/` directories on this branch — repo-level
docs (`README.md`, top-level `skills/CHANGELOG.md`, quickstart) don't
trigger a per-skill bump.

Skill	Version	Reason
`agentforge-workflow`	0.2.0 (unchanged)	Skill payload unchanged
`agentforge-cli`	0.1.0 (new)	Initial release
`agentforge-debug`	0.2.0 (unchanged)	Skill payload unchanged
`agentforge-template-author`	0.2.0 (unchanged)	Skill payload unchanged

Test plan

• `npm run skills:validate` — all four skills frontmatter-valid;
  CI Publish Skills workflow should turn green.
• `npx biome check .` clean.
• All 11 doc links in the README's Learn more section resolve
  (the new `docs/skill-quickstart.md` is added there too).
• Manual validation in a fresh coding-agent session:
  • Quickstart flow end-to-end: install → open editor → paste a
    sample brief → confirm `agentforge-workflow` fires and emits
    `.agentforge/`.
  • `agentforge-cli` happy path: "Run my pipeline for project
    X" → confirm prompt with cost ceiling → execution → summary.
  • `agentforge-cli` confirmation hold: respond "no" to a
    `run` confirmation → no command runs.
  • Disambiguation: "approve the gate" with two pending gates →
    skill lists both and asks which.