[jdesrosiers]

Validating a valid 150KB GitHub workflow against its schema with JSE output went from about 44ms to 7ms. An invalid one now takes about 24ms.

• Compute each output location's URI once. The instance location's URI was built three times for every keyword, and each build encodes the pointer. (44ms → 30ms)
• Build output records without hidden classes. Records are keyed by locations and URIs, so nearly every new key gave a plain object a new hidden class. Prototype-less records start in dictionary mode instead. (30ms → 26ms)
• Compute the instance location's URI once per schema. All of a schema's keywords share an instance location, so its URI is now built when the schema is entered rather than after every keyword. (26ms → 22ms)
• Build the normalized output only when it's asked for. The output is only needed when there are errors, but it was built during every evaluation. The plugin now records what each schema's keywords did and builds the output lazily. (21ms → 7ms valid; 23ms → 24ms invalid, since the output is now built after evaluation)

🤖 Generated with Claude Code