Restructure the docs around one front door

[lhparker1]

What this does

Reorganizes docs.lightconeresearch.org from "user guide + developer corner" into one site that presents Lightcone as a single toolkit. The agent plugin drives an analysis, lc runs it, and Lightcone Lab is where you watch it.

Home        Overview · Quickstart · How Lightcone works
Tutorial    6 chapters on the supernova example from agent-skills
Guides      scope · decisions · evidence · cluster · Lab · publish · without an agent
Concepts    your analysis file · outputs and provenance · universes · evidence
Reference   installation · lc CLI (incl. lc compute) · agent plugin · glossary · troubleshooting
Developers  architecture · engine internals · contributing · built on ASTRA

• Home is the overview: the "sidecar for your research" pitch, plus where to go next.
• The Quickstart is one tried-and-true path: the agent plugin (Claude Code or Codex) plus JupyterLab with Lightcone Lab, installed with one uv tool install jupyterlab --with jupyterlab-lightcone …. That install also puts lc on the PATH. Working by hand and working without a viewer live in "Use lc without an agent" and Reference › Installation.
• ASTRA appears as "your analysis file". The standard itself is linked from Concepts, Developers and Home's "Where to next".
• Existing pages move with their content intact:
  • cli/ → reference/cli/
  • api/ → developers/internals/
  • contributing/ → developers/contributing/
  • user/* → guides/, concepts/, reference/
• Main is merged in, so the new lc compute pages are included. The Quickstart uses the explicit-cluster flow (CLUSTER=$(lc compute launch --wait), then lc materialize "$CLUSTER").
• Docs paths updated elsewhere: the README, CLAUDE.md's docs rule, and the check-docs.yml review prompt now point at the new paths.

Why it's a draft

• Placeholder pages: several pages are still placeholders, shown with a blue "Planned" box. These are tutorial chapters 3 and 6, five guides and three concept pages.
• Quickstart step 3, "Start a new project", is a deliberate "Under maintenance" placeholder while project creation is reworked. An HTML comment marks the TODO.
• The tutorial is ported from the agent-skills supernova walkthrough. Its lc beats still need a real run to capture output.

How it was checked

• zensical build reports no issues and no broken links.
• The Quickstart and Lab paths were exercised in sandboxes (a clean Ubuntu container and an isolated scratch environment):
  • the plugin path and the by-hand path end to end;
  • the JupyterLab install command, including what happens when lc is already installed;
  • Lab's Create project button;
  • the inventory screenshot used on the page.
• The new compute commands in the Quickstart follow main's own lc compute and lc materialize docs. They have not been run here.

Follow-ups found along the way (not in this PR)

• lc init writes a placeholder git identity when none is configured, which gets past require_committer.
• After lc materialize, the annex journal is left uncommitted, so a fresh clone can't git annex get outputs.
• lc init has gaps when adopting an existing project:
  • it exits 0 while blocked;
  • it creates no universe beside an adopted spec;
  • it silently overrides existing Git LFS rules;
  • it fails when requires-python has an upper bound.
• The agent plugin still pins lightcone-cli==0.5.0rc3 and drives lc materialize without a cluster. It needs updating for the compute model before 0.5.0 ships.

The site deploys on release, so nothing goes live when this merges. It goes out with the next full release.

🤖 Generated with Claude Code