docs: add docker-agent v1 skill (SKILL.md + references)

[aheritier]

🤖 Automated implementer agent — this comment was posted by the implementer bot from Docker Agentic Platform, not by a human developer

Adds the v1 docker-agent Agent Skill bundle: SKILL.md + references/{cli,config,examples}.md under skill/docker-agent/, so an agent in any other project or harness can install it to author docker-agent configs and operate the docker-agent CLI as an external tool — including delegating to another agent (this repo's own contributor skills are unaffected: .agents/skills/{bump-config-version,bump-go-dependencies,triage-prs} still live where they always did).

Reworked after review feedback: the skill initially lived at .agents/skills/docker-agent/ — this repo's own dev-skill namespace, self-discovered only when docker-agent runs from inside its own source tree. That read as "how to develop docker-agent" instead of "how to use it", and missed the actual goal: an end user in some other project installing this so their agent can delegate to / operate docker-agent (e.g. docker-agent run <config> --exec to run a sub-agent, serve mcp/serve a2a to expose one for delegation). Fixed by:

• Relocating the source-of-truth to skill/docker-agent/ at the repo root, with the non-installed editorial policy (AGENTS.md) moved one level up to skill/AGENTS.md — so skill/docker-agent/ is exactly the installable payload, nothing to exclude.
• Rewriting every recipe repo-agnostic: no repo-relative paths (examples/*.yaml, docs/*, pkg/*) remain in installed content — converted to public URLs or self-contained inline YAML an installed reader can copy directly with no repo files on disk.
• Adding a prominent "Delegating to another agent" section to SKILL.md (the concrete use case that motivated the correction) with a full worked flow in references/examples.md.

Verified, not guessed, throughout both passes: every command, flag, hidden flag, and mutual-exclusion claim checked against the built binary's --help output and cmd/root/*.go; every config claim against agent-schema.json; every public URL against the docs' own canonical: frontmatter; the delegation example's run --json NDJSON shape and serve mcp --http requirement empirically confirmed by running the built binary against real cassette fixtures (two bugs caught this way in review and fixed before this push — see review history below).

Frontmatter safety: description is a single physical line (~344 chars, front-loaded with delegation/trigger keywords) — pkg/skills/frontmatter.go is a line-based parser that silently mangles a YAML-folded multi-line description. SKILL.md is ~160 lines, well under the ~500-line/~5k-token budget.

Also included: a one-line .gitignore fix — the existing docker-agent ignore pattern was unanchored and matched the skill's own directory at any depth (its neighbours, /cagent, /docker-mcp-*, are all anchored); anchored it to /docker-agent.

Went through two full review rounds with the reviewer sub-agent at the original location (fixed a mischaracterized __askpass description and a fallback/title_model grouping error) and two more after the relocation/reframing (fixed a missing --http flag and a broken jq extraction example in the delegation flow, both caught by actually running the built binary). Final verdict: approve.

Testing: task build, task test (green except the pre-existing, unrelated pkg/rag/treesitter CGO/gcc failure, confirmed to reproduce identically on a clean origin/main checkout), task lint (0 issues). Verified install-simulation: copying skill/docker-agent/ into an unrelated scratch directory's .agents/skills/ and running pkg/skills.Load from there discovers the skill with a correctly-parsed name and description.