Update claude and codex skills. Fix corresponding sections in README etc

[ScSteffen]

Summary

Consolidates the agent-skill set into four workflow-oriented skills maintained in parallel for both Claude Code and Codex, and removes the now-dead visualization config (baseline/output) that no bundled renderer consumes. Docs (CLAUDE.md, README.md) and example configs are brought in line with the current src/apeiron/ package layout.

Motivation & Context

The old skills were fine-grained, per-task scaffolds (new-detector,
new-updater, new-config, visualize, lint-check, …) that no longer matched how people actually use the framework, and several referenced a visualization entry point (src/visualize.py) and VisualizationCfg fields that the package no longer ships. This PR replaces them with task-level workflows and prunes the stale config/doc references so the repo describes what actually exists.

Approach

• Skills, reworked into 4 workflows (each maintained for both .claude/skills/
  and .codex/skills/):
  • install-apeiron — add Apeiron as a dependency to another project.
  • explore-examples — run a bundled MNIST/CIFAR example.
  • custom-experiment — scaffold harness + data utils + TOML for your own
    data/model and run it.
  • integrate-apeiron — bolt drift detection / CL onto an existing training loop.
  • Removed the old granular Claude skills (debug-experiment, explain,
    lint-check, new-config, new-detector, new-harness, new-updater,
    run-experiment, visualize).
• Config cleanup: VisualizationCfg keeps only input (the CSV path run
  metrics are written to); the unused baseline and output fields are dropped.
  Example TOMLs (mnist, cifar10_vgg11, cifar10_vit) are updated to match.
• Docs: CLAUDE.md and README.md updated to the src/apeiron/ package
  paths, MLflow/WandB logging backends, the EnsembleDetector
NotImplementedError caveat, and a new "Agent Skills" README section. Removed
  references to the dropped src/visualize.py entry point.

Screenshots / Logs (optional)

N/A — docs/config/tooling change.

API / CLI Changes

• VisualizationCfg — removed fields baseline: float and output: str; only
  input: str remains (default "output/cl_only.csv").

Breaking Changes

• TOML configs that set [visualization] baseline = … or [visualization] output = …
  must drop those keys; only input is still recognized. (The bundled example
  configs are already updated.)

Performance (optional)

N/A.

Security & Privacy

• No secrets committed
• Input validation added where needed (N/A — no new input paths)

Dependencies

• None added or removed. poetry.lock content-hash/generator header updated only.

Testing Plan

• Unit tests — tests/test_config.py::test_visualization_cfg updated to assert
  the remaining input default.
• Integration tests
• e2e / smoke test
• Manual steps: poetry run pytest, poetry run ruff check .,
  poetry run mypy .

Documentation

• Docstrings updated
• User docs / README updated
• CHANGELOG entry

Checklist

• Code formatted (Ruff) → ruff format --check
• Lint passes (Ruff) → ruff check .
• Types pass (mypy/pyright) → mypy src
• Tests pass (pytest) → pytest -q
• Backward compatibility considered (see Breaking Changes)
• Adequate comments for tricky parts
• CI green

Risk & Rollback Plan

Low risk — docs, skills, and a small config-field removal. Rollback by reverting
the PR.

Notes for Reviewers

• Start with src/apeiron/config/configuration.py + tests/test_config.py for
  the one behavioral change.
• The bulk of the diff is skill markdown moving from many granular skills to four
  workflow skills, mirrored across .claude/skills/ and .codex/skills/.
• Note the README's reminder to keep the two skill trees in sync.

[anagainaru]

Looks great, I'll test them in a little bit and come back to approve. This is a good base on which to add the driver, drift detection and CL skills.

[anagainaru]

All the skills except the custom-experiment went well. I had some issues creating a good harness with the custom experiment. I think once we have a better documentation for the model harness this might go better. Let's merge this for now.