Why accessors and not a Proxy

zone.ceiling_height is a real property. It is not a Proxy trap, and the difference is not academic.

The obvious translation

The Python library resolves zone.ceiling_height through __getattr__: the attribute does not exist, Python calls a hook, the hook consults the schema. It works, it is concise, and the mechanical translation to JavaScript is a Proxy.

That translation was rejected. A Proxy gives up two things at once:

Editors cannot see through it. A proxy's properties do not exist until something asks for them, so there is nothing for TypeScript to describe and nothing to complete. Every field name becomes a string typed from memory, and every typo becomes a runtime undefined. That is precisely the failure mode of the Python original, reproduced on purpose.

V8 cannot optimize it. Property access on an ordinary object is an inline cache hit. Property access through a proxy is a call into a trap, every time. On a 290,000-object corpus that difference is the difference between the parser being usable in a browser and not.

What is done instead

Each object type gets one prototype, built once by ObjectShape, carrying Object.defineProperty accessors for every field in the schema. Every instance of that type shares it. Reads are ordinary monomorphic property lookups, and the generated .d.ts interfaces describe them statically.

Shapes are keyed by the schema definition object, not by type name. Because the bundle is content-addressed, Zone in 9.4.0 and Zone in 26.1.0 are the same frozen definition, so they share one shape and one prototype. A document holding objects from two EnergyPlus versions stays monomorphic without anyone arranging for it.

Why the setter matters

The reference graph is live. Renaming an object rewrites every field elsewhere that pointed at the old name:

const zone = doc.require('Zone', 'SPACE1-1');
zone.name = 'Open Office';
surface.zone_name; // 'Open Office'

Nothing was called to make that happen. There is no update(), no rebuildIndex(), no invalidation step for a caller to forget. The write went through the accessor's setter, and the setter is where the graph edge is moved.

That is the real argument for defining accessors rather than storing plain data properties. A plain property would be faster still and would leave the graph stale on every write, which is a correctness problem rather than a performance one. Extensible groups are covered too: ZoneList, Branch, and the supply and return paths carry all of their references inside repeat groups, so a graph that ignored those would make rename() quietly produce a broken model.