docs: recreate the harness without the CLI via the docsite llms.txt

[doidor]

What

Documents how to recreate the entire AgentRig harness without the CLI by pointing any coding agent
at the docsite's machine-readable llms.txt. This is the
concrete payoff of AgentRig's no-lock-in design — the package is never a runtime dependency.

Supersedes the earlier idea (first revision of this PR) of shipping a dedicated bootstrap-harness
skill. That skill was redundant with the docs: the docsite already contains the recreation
knowledge, and llms.txt already makes it agent-consumable. This revision reverts the skill
entirely and instead makes the docs the single, authoritative home for the recreate-by-hand path.

Changes (docs + README + site config)

• docs/getting-started.md — new section "Recreate the harness without the CLI (no lock-in)":
  what llms.txt gives an agent (the 12 principles as a complete per-principle artifact inventory,
  the surface projection map, the install-completeness audit), an explicit pointer to
  knowledge/templates/ for exact
  copy-paste file contents, and a ready-to-use prompt.
• docs/index.md — the "No lock-in" card now links to that walkthrough.
• README.md — concise "No lock-in — recreate the harness without the CLI" section.
• markbook.config.ts — siteUrl set to the canonical custom domain https://tudorpopa.com/agentrig
  so generated canonical / og:url / sitemap (and every docsite link in the prose) agree with where
  the site is actually served.
• .changeset/recreate-from-llms.md — patch (README ships in the package).

No behavior changes, no installed-artifact changes, no knowledgeVersion bump.

Did I verify the docs are actually sufficient to recreate the harness?

Yes — gap analysis:

• Conceptual completeness: strong. principles.md is a full per-principle inventory (12/12, each
  naming its artifact + path); agent-surfaces.md is the exact projection map; evals.md documents
  the deterministic audit the agent can self-check against.
• The one gap: the docs didn't point at the copyable canonical templates for exact file
  contents (inlining them verbatim would duplicate and drift from knowledge/templates/). Fixed by
  making knowledge/templates/ the explicit "exact bytes" source in the new section. llms.txt
  (architecture + inventory) + knowledge/templates/ (verbatim files) — both in the same public
  repo llms.txt indexes — together are sufficient.

Verification

• node dist/cli.js eval --static . — Install Completeness 100% (33/33), Quality Probes 100% (6/6) (back to the main baseline; the skill's extra check is gone)
• npm run build — clean · npm test — 66/66 pass
• npm run docs:build — clean; the new section's anchor slug
  (recreate-the-harness-without-the-cli-no-lock-in) matches the cross-links, the content flows into
  the regenerated llms.txt, and generated canonical/og:url/llms all use tudorpopa.com (zero
  doidor.github.io left in site/)
• Branch rebased to a single linear commit on origin/main

---

Also in this PR: bumps the docs toolchain @doidor/markbook + @doidor/markbook-core 0.2.0 → 0.6.0 (devDependencies). All changes are additive with safe defaults — no breaking config changes; markbook.config.ts is unchanged. 0.6.0 collapses the homepage canonical to the site root, so the generated docsite canonical is https://tudorpopa.com/agentrig/. docs:build + the full suite pass on the upgrade.