feat(json): add standard numeric conversion and owned encoder results

[sanohiro]

Owned JSON records containing finite f32/f64 fields now use the same structural codec as other supported leaves. Selected numbers decode directly to their target width with nearest/ties-even rounding; rounded overflow rejects, while grammar-valid unselected large numbers remain skippable.

Both json.encode and json.encode_bounded return Result<string, Error>. Success transfers an independent output buffer; selected NaN/infinity and invalid byte limits return Error.Invalid without partial output. The compiler uses one encoder operation, reserves the source across limit evaluation, validates the complete MIR JSON sequence, and carries owned float graphs through V3/interface 12 and cache invalidation. Existing callers and English/Japanese contracts are synchronized. R65 and align-llm adoption remain separate.

Validation covers independent numeric/descriptor goldens, host rounding modes, ownership/cleanup and native admission, every body emitter, whole/per-unit routes, cache changes and compiled ABI export parity. Local nine-pair measurements found no reproducible ordinary decode regression; the final mixed numeric encoder is faster than both baseline raw encoding and baseline plus the explicit owned copy. Detailed measurements and artifact identities are in bench/json_numeric/r63-results.md.