OpenEval 0.3.0

Code and hybrid judging is now supported. Export an ordinary function from judge.ts, use judge.md for LLM criteria, or use both against the same recorded EvalRun. Boolean and fractional criterion scores use equal-weight aggregation. JudgeContext exposes recorded metrics, traces, files, full native history, SDK/schema access, and read-only SQL. The viewer now shows code output, frozen source, and process logs. This pre-1.0 release uses canonical criterion names and results schema 5; older stores require a separate migration.