Add canonical pipeline benchmark and optimize hot paths

[schani]

Summary

• add a whole-pipeline benchmark covering JSON parsing, schema processing, inference and transformations, TypeScript/Rust generation, and maximum observed JS heap
• add a canonical real-world set: USGS earthquakes, GitHub REST OpenAPI, NVD CVE 2024, HL7 FHIR R5, and Kestra 0.19
• expose per-pass timing through an optional onTiming callback
• document cache/download behavior and benchmark usage
• optimize the hot paths found by profiling the canonical suite
• add focused unit coverage for timing, JSON/YAML parsing, naming dependencies and forbidden names, and accumulated string cases

Root causes and optimizations

The Rust renderer's naming pass repeatedly rebuilt identical forbidden-name unions and rescanned namespaces that were already complete. This was especially expensive because Rust checks its global keyword namespace for properties, enum cases, and union members.

Large inferred JSON inputs also accumulated millions of redundant empty attribute entries and allocated a StringTypes object/map/set per observed string. Large JSON schemas were being parsed through the YAML lexer even when they were strict JSON.

This change:

• caches each namespace's forbidden-name union
• tracks only unfinished naming namespaces
• skips redundant empty attribute entries and batches string-case counts
• tries JSON.parse before retaining YAML as a compatibility fallback

Canonical results

Node 24.6.0, macOS arm64, three measured runs and one warmup per case:

Case	Before	After	Change	Max heap before	Max heap after
USGS / TypeScript	250.9 ms	189.5 ms	-24.5%	146.5 MiB	130.1 MiB
USGS / Rust	257.2 ms	186.4 ms	-27.5%	148.4 MiB	121.1 MiB
GitHub OpenAPI / TypeScript	1.614 s	1.176 s	-27.1%	203.1 MiB	197.3 MiB
GitHub OpenAPI / Rust	6.557 s	1.209 s	-81.6%	204.8 MiB	195.8 MiB
NVD / TypeScript	5.558 s	3.450 s	-37.9%	2674.2 MiB	1825.2 MiB
NVD / Rust	5.402 s	3.464 s	-35.9%	2673.7 MiB	1831.6 MiB
FHIR / TypeScript	2.238 s	1.519 s	-32.1%	499.3 MiB	204.7 MiB
FHIR / Rust	4.872 s	1.715 s	-64.8%	492.1 MiB	245.8 MiB
Kestra / TypeScript	6.178 s	4.850 s	-21.5%	633.1 MiB	530.3 MiB
Kestra / Rust	6.538 s	4.931 s	-24.6%	635.0 MiB	530.9 MiB

The sum of the ten medians is 42.5% lower. Every case is faster and every measured maximum heap value is lower.

All ten generated canonical outputs were compared against an unoptimized build and matched byte-for-byte by SHA-256.

Validation

• npm run build --workspace quicktype-core
• npm run test:unit — 20 files, 136 tests
• npm run lint
• QUICKTEST=true FIXTURE=typescript npm run test:fixtures — 70 tests
• full canonical benchmark, three measured runs plus one warmup
• byte-for-byte SHA-256 comparison for all five inputs in TypeScript and Rust

The external Rust fixture compiler was not run because cargo is unavailable in this environment; Rust generation is covered by unit tests, the canonical runs, and output identity checks.