[jdesrosiers]

The language server's evaluation plugin adds a lot to every validation. With a 150KB GitHub workflow, it added about 260ms on top of hyperjump's own ~50ms. These changes cut the plugin's own overhead from about 260ms to about 90ms and stop sending diagnostics while someone is typing.

• Index incomplete locations by their parent. The object keywords looked at every incomplete location in the document each time they were evaluated, and every array has one. Now they only look at the ones in the object they apply to.
• Combine oneOf completions without a Pact pipeline. A pipeline ran for every location under every oneOf. Now one pass groups the sets by location.
• Mark conditional annotations while appending them. Every anyOf/oneOf alternative copied its subschema's whole annotation record just to mark annotations at one location as value annotations.
• Store subschema tracking state on the context. SubschemaTracker set a WeakMap entry for every keyword. Its state is now a property under a symbol that's unique to the tracker, so trackers still don't share state. This relies on hyperjump creating a fresh context for each keyword, which the WeakMap version also relied on.
• Filter annotations by kind with a loop. Deprecated diagnostics asks for annotations at every node, and the Pact pipeline dominated that walk.
• Use the pointers walkNodes builds for document colors. getPointerForNode searches arrays for the node's index, so colors were quadratic in array size.
• Wait for edits to pause before sending diagnostics. Diagnostics for an edit are sent once the document hasn't changed for 500ms, like VS Code's JSON language server. A document is validated without waiting when it's opened. A change or close cancels a send that's waiting or in progress. The delay is a buildServer option so tests don't wait for it.

Measured with a generated 154KB GitHub workflow against the SchemaStore schema (medians):

	Before	After
Validation	~310 ms	~144 ms
Annotations at every node (deprecated diagnostics, colors)	~26 ms	~3 ms
hyperjump alone, JSE output	~53 ms	~53 ms

Most of what's left is copying completions and annotations up through each schema, which keeps the plugin simple. Recording a tree during evaluation and flattening it when requested took validation to ~200 ms before the last two improvements, but it was left out because it makes the plugin harder to follow. The JSE output plugin is a large part of the remaining time. It computes Instance.uri three times per keyword, and computing it once saves about 20 ms. That fix belongs in @hyperjump/json-schema-errors.

Testing

The debounce has new tests for opening a document, changes in quick succession, and closing before the delay is over. The other changes aren't intended to change behavior. All 540 existing tests pass.

🤖 Generated with Claude Code