Remove the documentation, now maintained in LightconeResearch/docs

[EiffL]

The documentation has moved to LightconeResearch/docs, which now serves docs.lightconeresearch.org. This PR removes the in-repo copy and everything that existed only to build, publish or check it.

Removed

• docs/ (35 files) and zensical.toml
• .github/workflows/docs-deploy.yml: the mike deploy to gh-pages on release
• .github/workflows/check-docs.yml: the Claude doc-drift review on merged PRs
• the docs dependency group in pyproject.toml (zensical, mike)

CLAUDE.md

• Removed the "docs/ is live again" rule, which required changes to land with their docs/cli/, docs/api/ and user-guide pages.
• Removed the "Documentation versioning (mike)" section.
• Removed the docs build command and the two docs workflows from Development Commands.
• Removed "update the docs pages the layer touches" from the Extending table.
• Reworded two sentences that pointed at the in-repo docs/ (the design-records note, and the uvx troubleshooting entry).

README

The relative docs/user/cluster.md link now points at the docs site root, and the deep /user/getting-started/ link is gone. The live site still serves the old paths, but the docs repo's main uses new ones, so deep links would break on its next deploy.

Already-published docs keep serving

This PR removes only the source. The gh-pages branch and this repo's Pages settings are untouched, so the site already published from them stays up. With docs-deploy.yml gone, releases no longer publish new docs versions there.

Testing

No Python changed. A local pytest -x run passed 511 tests, then stopped at test_compute_slurm.py::test_standard_bootstrap_starts_scheduler_and_worker_on_rank_zero_and_worker_on_rank_one[False]. Nothing here touches what that test runs, but I have not checked it against main, and the rest of the suite did not run.

🤖 Generated with Claude Code

[github-actions]

✅ Eval

Metric	Value
Outputs check	success
Agent run	success
Turns	9
Tool calls	7
Cost	$0.14
Agent wall time	0m47s
Model	claude-sonnet-5-5

lc status

  mode:    direct
  sandbox: landlock (fs: declared, network: allowed)
  crate:   up to date with the outputs

  · current  baseline/best_fit        517dbf8
  · current  baseline/hubble_diagram  517dbf8
  · current  baseline/residuals       517dbf8

3 current

Confusion & pain points (Claude analysis)

Confusion & pain points

• Essentially clean run. There were no errored tool calls, and the agent finished in 9 turns. The friction below is minor and only inferred from the digest.

• The spec stub needed hand-editing, with no lc/astra help. The scaffolded astra.yaml lacked format:, inputs:, decisions: and complete recipe commands on the outputs. The agent patched them with a Python string-replace script run over the YAML. The cause is that format: is required but must be written by hand, which CLAUDE.md already records. Fix: have astra validate or lc status flag a missing format: early, and consider an astra helper for wiring up outputs.

• The agent probed the data format by hand. It ran head, wc -l and a second sed -n 1,15p to learn the column layout of SCPUnion2.1_mu_vs_z.txt, because the header comments weren't in the first view. The cause is that nothing in the spec or data/README.md documents the input's columns. Fix: let the eval or spec prompt describe the columns, or add a description or schema to the declared input.

• A hand-picked pinned astra-tools@0.2.18 via uvx was needed to validate. The agent didn't use astra directly, which suggests the CLI wasn't on PATH. Run this way, the validator version can drift from what lc uses. Fix: have the eval harness put astra on PATH, or make the skill name the right invocation.

• The license key and the cluster lifecycle were manual steps. The agent added license = "CC-BY-4.0" by sed to get the crate, then ran lc compute launch --wait and lc compute down local by hand. It chose the license without being told. Fix: have the task prompt state the license and the compute-lifecycle expectation. The sed insertion also relies on requires-python being on one line.

Full trace: agent-trace artifact on this run.