[Lojhan]

JSON result guards allocate callbacks while recursively checking arrays, tuples and unions. Replace those callbacks with loops that retain the previous short-circuit order, captured lengths and absent-index handling. The preceding finite JSON validation pass and all shape, depth and value-count checks remain unchanged.

A separate SQLite allocation run on Apple M3 Pro / Node 24.11.1 sampled 26.5% fewer allocated bytes for compiled JSON queries and 30.1% fewer for parsed-policy SQL-file queries (5,000 calls each). Matched ordinary timing trials do not establish a latency improvement: compiled JSON is 25.93 → 25.79 µs while raw/Reylo controls drift slightly slower. No claim of closing the Reylo latency gap.

The evidence directory preserves raw timings, compressed V8 heap profiles, source/patch identities, methodology and a derived summary. Only the JSON validator changes production behavior; there is no new API or dependency.

Validation: full pnpm verify passed, including package tests and coverage, TypeScript builds, packed-consumer/editor contracts, documentation build, and performance/soak budgets. GitHub CI and CodeQL passed, including packed real-database consumers and packaged editor tests on Linux, macOS and Windows. The first source-quality attempt failed the unchanged editor schema-churn p99 budget (454.88 ms against 120 ms); one clean retry of the same source passed at 12.71 ms. No code or budget changes were made to obtain that retry result. CI evidence: https://github.com/Lojhan/typed-sql/actions/runs/40793109552/attempts/2.