feat(cli): --json output + updated skills that use jq

[tato123]

Summary

Adds a global `--json` flag so skills (and any jq-driven consumer) can extract fields directly instead of parsing ASCII. Also rewrites the two skills that consume output (`amos-focus`, `amos-next`) to use `jq` and craft natural-language summaries rather than piping raw CLI text back to the user.

JSON shapes

• `amos milestones --json` → `{focus, milestones: [{title, open, ready, done, focused}]}`
• `amos next|blocked|orphans --json` → `{focus, count, nodes: [{name, issue_number, description, blocked_by, blocks, related_to, milestone, facts, labels_local}]}`
• `amos focus --json` → `{focus, action?}`
• `amos validate --json` → `{ok, node_count, issues: [{kind, ...}]}`, non-zero exit on errors
• `amos show --json` → full node with body, facts, every relationship field

Human-readable output paths are unchanged when the flag is absent — existing tooling keeps working.

Why

Skills had to scrape lines like `- @github:org/repo#347 — Research: DMA-BUF...` to find issue numbers and descriptions. Brittle and slow. Now:

```bash
amos next --json | jq -r '.nodes[] | "(.name)\t(.description)"'
```

Skill updates

• `amos-focus` now reads/writes via `--json` + jq; prescribes natural-language confirmations instead of raw echo.
• `amos-next` protocol rewritten to consume JSON at every step and compose English announcements, test results, and task completion reports. Rule Make amos runnable from the project root with no arguments #7 added: "Present data to the user in natural English, never raw JSON."

Test plan

• 57 unit tests passing
• `amos milestones --json` + `jq` against streamlib (9 milestones parse cleanly)
• `amos next --json` + `jq` against streamlib focus (3 ready items parse cleanly)

🤖 Generated with Claude Code