feat: ZSTDMT-style parallel compress/decompress + wasm-simd128-mt payload

[polaz]

Summary

Add ZSTDMT-style parallelism: split the input into independent jobs so both
compression and decompression can run across multiple threads, and ship a
thread-capable wasm payload (wasm-simd128-mt) so the same parallelism is
available in the browser.

Core: parallel compress + decompress via job splitting

Mirror libzstd's ZSTDMT model. The unit of parallelism is an independent
job, not a single zstd block:

• Compress — partition the input into jobs, compress each on its own
  context/thread, concatenate the results. (Native track: feat(encoding): Rust-side rayon-based multi-threaded compression (multi-frame concat) #19 rayon
  multi-frame compression.)
• Decompress — only independent jobs/frames can decode in parallel.
  A single frame decodes strictly sequentially (every block depends on the
  window built by the preceding blocks), so parallel decode requires the jobs
  to be self-contained (concatenated independent frames, or explicit
  rsyncable / window-reset boundaries). Cross-frame parallel decode + the
  intra-frame block scheduling question is the native track perf(decoding): parallel block decompression for multi-block frames #72.

So the deliverable is: a job-splitting layer (compress side) that produces a
stream whose job boundaries are independently decodable, plus a parallel driver
on both sides. Within one job/frame, decode stays sequential by construction.

wasm delivery: wasm-simd128-mt payload

A third wasm payload, built with the WebAssembly threads proposal, that runs the
parallel compress/decompress across Web Workers — shipped alongside the existing
simd and scalar payloads.

1. Build with -C target-feature=+atomics,+bulk-memory,+mutable-globals and
   -Z build-std (shared-memory std).
2. Bootstrap a worker pool (e.g. wasm-bindgen-rayon): the main module hands
   each worker the compiled module + shared memory.
3. Runtime gate in the npm loader: select mt only when crossOriginIsolated
   is true (SharedArrayBuffer available); otherwise fall back to simd /
   scalar. Document the required COOP (same-origin) + COEP
   (require-corp) response headers for consumers.
4. Package the third payload directory (mt/) and extend capability detection
   • lazy load.

Dependencies

• Native multi-threaded compression: feat(encoding): Rust-side rayon-based multi-threaded compression (multi-frame concat) #19 (rayon multi-frame / job splitting).
• Native parallel decompression: perf(decoding): parallel block decompression for multi-block frames #72.
• Part of Roadmap: structured-zstd feature parity with C zstd #28 (feature parity roadmap).

The wasm-mt payload is the browser delivery of the same job-splitting parallel
codec; the native tracks (#19, #72) provide the parallel driver this packages.

Acceptance criteria

• Job-split compression produces a valid stream whose jobs are independently
  decodable; parallel compress shows a multi-core speedup, output decodes in any
  compliant zstd decoder.
• Parallel decompression across independent jobs/frames shows a multi-core
  speedup; single-frame decode is unchanged (documented sequential limit, not a
  regression).
• wasm-simd128-mt builds and loads under a cross-origin-isolated host; falls
  back cleanly to simd / scalar when SharedArrayBuffer is unavailable.
• Round-trip + cross-decoder parity tests pass on all payloads.
• Consumer docs cover the COOP/COEP requirement.

Estimate

The wasm-mt payload + loader: 3d 4h (build-std + worker bootstrap + capability
detection + COOP/COEP docs + packaging + parity tests). The parallel
compress/decompress engine itself is the native tracks (#19, #72) and is
estimated there.

[coderabbitai]

🔗 Related PRs

#43 - test(bench): expand benchmark parity matrix [merged]
#213 - perf(decode): SIMD-16 fast path for short offsets {1, 2, 4} [merged]
#227 - perf(decode): 8-slot software prefetch pipeline for sequence execution (#208) [merged]
#367 - feat(wasm): streaming compress ZstdCompressStream + align decoder window cap to zstd default [merged]
#372 - ci(bench): continuous wasm dashboard shard vs @bokuweb/zstd-wasm [merged]

---

📝 Issue Planner

_{Check the box below or use the @coderabbitai plan command to generate an implementation plan and prompts that you can use with your favorite coding assistant.}

• Create Plan

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!