👻 Ghost Serializer 1.3.0 — Unified JSON + YAML + Proto3 JSON runtime

This isn't an incremental bump from 1.2.7 — it's a rewrite of what "Ghost" ships as an artifact, plus the biggest decode-throughput jump in the project's history.

Highlights

• 🔀 One runtime, three wire formats. JSON, YAML, and Proto3 JSON now live in a single ghost-serialization artifact. ghost-yaml, ghost-yaml-ktor, and ghost-protobuf are gone.
• ⚡ Decode throughput up to 1.2 GB/s. A rewritten decode hot path (in-order field prediction + SWAR scanning) pushed every channel well past 1.2.7 — String decode crosses 1 GB/s for the first time.
• 🌐 Kotlin/Wasm everywhere. ghost-api, ghost-serialization, ghost-ktor, and the new Ghost Playground now ship wasmJs targets alongside JVM/Android/iOS.
• 🕹️ New Ghost Playground — Compose Multiplatform (JVM + Wasm) site with a live JSON/YAML/Proto Studio and a browser speed test vs kotlinx.serialization and Moshi.
• 🐛 Two silent data-corruption bugs fixed in the streaming reader and the WebFlux NDJSON decoder.

⚡ Performance: crossing 1 GB/s

Twitter macro fixture (631,514 bytes). Same 3 decode channels:

Decode	1.2.7	1.3.0	Δ 1.2.7→1.3.0	vs KSER	vs Moshi
String	0.803 GB/s	1.229 GB/s · 514 µs/op	+53%	+71%	+229%
Bytes	0.698 GB/s	1.052 GB/s · 600 µs/op	+51%	+148%	+283%
Streaming	0.304 GB/s	0.529 GB/s · 1194 µs/op	+74%	+174%	+24%

(1.2.7 GB/s derived from its own published ops/s × the 631,514-byte Twitter fixture — same fixture in both versions; 1.2.7 didn't track GB/s directly.)

Three profile-driven changes got us here, each A/B-verified and kept only when it won:

1. In-order field prediction — machine-generated JSON lists fields in declaration order, so the reader checks the predicted next field first instead of hashing every key. Biggest single win.
2. SWAR whitespace + key comparison — 8 bytes at a time (VarHandle wide load on JVM, scalar SWAR elsewhere).
3. SWAR value-string scan with deferred pool hash — branch-free 8-byte scan for quotes/backslash/control/non-ASCII; the string-pool hash is now only computed for strings short enough to actually pool.

Encode got targeted wins too (ASCII-prefix fast path +10%, BMP Unicode fast path +13% on the string channel). Full multi-engine tables: benchmarks.md.

🔀 Unified runtime: JSON + YAML + Proto3 JSON

1.2.7 shipped three separate artifacts. 1.3.0 merges them into ghost-serialization:

// Before (1.2.x)
implementation("com.ghostserializer:ghost-yaml:1.2.x")
implementation("com.ghostserializer:ghost-yaml-ktor:1.2.x")
implementation("com.ghostserializer:ghost-protobuf:1.2.x")

// After — one runtime artifact
implementation("com.ghostserializer:ghost-serialization:1.3.0")
implementation("com.ghostserializer:ghost-ktor:1.3.0") // JSON + YAML + Proto adapters

Or apply id("com.ghostserializer.ghost") version "1.3.0" and let the Gradle plugin auto-inject adapters.

Full diff: v1.2.7...v1.3.0