lmschem

A coordinate-free circuit schematic engine. Authors describe a circuit as a port graph — components, electrical nets, relational placement — and a deterministic geometry engine places bodies, routes every wire, enforces hard drawing rules, and renders a PNG through lcapy/circuitikz. The same IR always yields the same drawing.

The complete authoring language reference is lmschem/rules.md — it is the only document an author needs, and it ships as package data.

Architecture (four layers)

PortGraph IR ──► geometry engine ──► renderer adapter ──► PNG
(portgraph/)     (geometry/solve,     (geometry/adapters/   
                  router, report)      lcapy.py)            

1. Port-graph IR (portgraph/ir.py) — Pydantic models: components, nets, relational place anchors, rails/snaps, route hints, buses, decorations, blocks. No coordinates, ever.
2. Geometry engine (geometry/) — deterministic placement (trunk-and-branch routing, channel assignment, facing-aware alignment), hard-rule checks with structured, fix-hinted issues, an ASCII occupancy grid, and solver-placed text (net labels, bus captions, decoration labels) backed by measured ink metrics.
3. Renderer adapter (geometry/adapters/lcapy.py) — lowers the solved layout to an lcapy netlist with injected node positions (lcapy's placer is bypassed entirely) and verifies a real PNG came back. Emission is inventory-driven: bipoles, amps, chips/shapes, and the generic node-grammar path (node_pinnames slot order) cover every drawable class.
4. Incremental edits (edit.py) — targeted IR edit operations with a flow-style YAML emitter that keeps files in the shape authors write.

Entry point:

from lmschem.author import author_circuit
result = author_circuit(ir_yaml_string, "out_dir")
# result: ok / report / grid / issues / image

The inventory contract

lmschem/inventory/*.yaml is the sole runtime authority for everything lcapy-specific: legal pin tokens, drawn subsets, pin aliases, netlist-node slot order (node_pinnames), measured body extents and pin coordinates, text metrics, decoration glyph ink, options, and the derived semantic vocabulary. Runtime never introspects lcapy.

The inventory is generated, never hand-edited (except components.yaml, schemcpts.yaml, options.yaml, placer.yaml, which are curated). Regenerate after any lcapy upgrade:

python -m lmschem.gen_inventory

This re-measures glyphs, rewrites the generated YAMLs, re-derives the semantic vocabulary (semantic_types.yaml), and re-splices the generated component reference into rules.md between its markers. lcapy is pinned exactly in pyproject.toml because the measurements are against that version.

Coverage contract: every drawable inventory class either solves and emits (guarded by tests/test_vocabulary_coverage.py, one param per class, no rendering) or sits on gen_inventory.EXCLUDED_CLASSES with a written reason (K, A, W, MISC, sV, sI). A new lcapy class becomes a failing test param until supported or excluded.

Tests

python -m pytest tests -q

Most tests render real PNGs through pdflatex + ghostscript — a TeX distribution and ghostscript must be on PATH. The solve+emit coverage guard and the geometry tests run without rendering.

lcapy quirks this engine works around (do not "simplify" these away)

• LatexRunner.run never checks the pdflatex result and to_png_ghostscript never checks its output — the adapter verifies the PNG file itself and raises a structured error.
• The opts parser splits on commas, so every label/value rides braced (l={...}); unbraced LaTeX with commas kills the compile silently.
• A netlist with zero numeric nodes (all-unwired chips) crashes lcapy's canvas sizing — the adapter raises "nothing in the drawing is wired" first.
• Uisoamp's class declares a phantom ocm pin its own draw rejects; gen_inventory._pins_for reads property-pin dicts from the class's own __dict__ only.
• lcapy injection scales by node_spacing (2.0); gutters never move terminal segments.

Relationship to coursesmith

lmschem is standalone and speaks bare IR. Coursesmith integrates it through a thin adapter layer on its side (generator spec/executor, MCP edit tools, discovery): the generator-spec envelope, analysis requests, and example specs are coursesmith concepts and live there, not here.