Restructure CLAUDE.md into skills; add Sphinx docs on GitHub Pages

[taw181]

This branch has two parts.

---

Part 2 — Sphinx documentation + GitHub Pages (commit 21f20f2)

Adds a Sphinx build and deploys it to GitHub Pages.

• API reference auto-generated from docstrings by sphinx-apidoc at build time (not checked in), covering servers/, configuration/, coordinators/, loggers/, analysis/, drivers/.
• Narrative skeleton: index, getting-started, architecture (Markdown via MyST) — deliberately short.
• Furo theme; hardware/GPU imports mocked so the build runs anywhere.

.github/workflows/docs.yml: build on every push/PR (fails CI if the docs don't build); deploy main to https://coldmatter.github.io/pytweezer/ (Pages source already set to "GitHub Actions").

poetry install --with docs
cd docs && poetry run make html

Interaction with Part 1

Part 1 said docs/ was being retired and its deletion was left to @twalker27. That is now superseded: the 7 design notes move to docs/notes/ (byte-identical, not rendered) and docs/ becomes the Sphinx source root.

Excluded from the API reference (rationale in docs/Makefile)

• GUI/, bin/ — out of scope for now
• servers/model_sync.py — reads CONFIG["Servers"]["Model Sync"] at import, not always present (looks like a latent bug)
• drivers/thorcam.py, drivers/tweezermonitorcam.py — module-level pll.par[...] = ... the import mock can't satisfy

Minor source fixes (warning-free build)

• projections.py: bullet-list indentation in the module docstring
• device_status.py: one-line docstring on the status_received signal

CI note

The docs job runs poetry install --with docs, installing the full dependency tree (autodoc imports the real modules). If a heavy wheel fails in CI, a leaner docs-only install can be carved out.

---

Part 1 — Restructure CLAUDE.md into on-demand skills (commit b5a657e)

Why

CLAUDE.md was 267 lines, ~170 of them architecture prose loaded into context on every agent run — most of it relevant to maybe one task in twenty. Its own opening section says situational material belongs in a skill; this makes it follow its own advice.

What changed

CLAUDE.md: 267 → 129 lines. Keeps what applies to every run (what this is, commands, docstring/comment/TODO/git conventions) plus a short architecture summary that routes to skills. Adds the get_logger convention, which is broad enough to stay always-on.

Three new orientation skills, covering understanding and modifying the frameworks — the existing add-* skills already cover creating new instances:

Skill	Covers
pytweezer-architecture	server/client split, CONFIG categories, get_config() vs paths.py, launch protocol and dual-mode main(), messaging fabric, logging
pytweezer-device-framework	get_device() addressing, simulation, composites and coordinators, what blocks what over RPC, running calls concurrently
pytweezer-gui-internals	TabbedGUI vs the legacy BWidget stack, teardown, dock-based tabs, the three status sources, Device Status service

Skills are self-contained. The conventions, constraints and rationale that can't be read off the code now live in the skills — including the one-command InfluxDB setup, which moves into add-logger. The docs/*.md pointers left in source, tests and tools are rewritten to stand on their own; the SLM preload/hardware-trigger optimisation is preserved in the rearrangement.py docstring.

Correction: the old "two-PC role split" framing is wrong — it's one server PC and arbitrarily many client PCs, each running whatever devices are attached to it. Fixed throughout.

add-logger trimmed 279 → 217 lines: dropped the walkthrough of base.py behaviour that's readable off the code, kept every gotcha.

Verification

ruff check and ruff format --check clean; 126 tests pass. Source changes are docstrings and comments only.

🤖 Generated with Claude Code