[reececomo]

Removes per-call heap allocations from encode, decode and peekHeader so that hot-path encode/decode in real-time loops produces less garbage.

Changes

1. encodeInto() reuses one BufferWriter per format via BufferWriter.$reset(). A new encodeIntoLength(data, bytes): number returns the byte count without creating a view. A target that is too small still throws TinybufError.
2. decode() and decodeInto() reuse one BufferReader per format via BufferReader.$reset(). The shared reader is taken out of the format while it is in use, so a decode inside another decode, such as from a field transform, creates its own reader. Encode works the same way.
3. BufferWriter.$writeBytes() copies with Uint8Array.set() and only wraps non-Uint8Array input.
4. peekHeader() reads Uint8Array and DataView input without allocating, and now throws RangeError whenever the input itself has fewer than 2 bytes.
5. Decoded arrays keep their identity when their length changes. The existing test that asserted a new array was updated.

For 1 and 2, the writer and reader keep a DataView over the whole underlying ArrayBuffer and track a base offset and length. Retargeting to another view of the same buffer, such as slots in a pool or subarrays of one packet, reuses the DataView. Reads and writes are bounds-checked against the target view. A DataView passed to decode() is adopted directly.

New tests in src/__tests__/allocations.test.ts cover shared-buffer slots, subarray decoding, view bounds, re-entrancy, error recovery, array reuse, and constructor counts for the zero-allocation paths. The shared reader and writer keep a reference to the last buffer until the next call.

Also adds .github/workflows/publish.yml, copied from the pixijs-* repos. It publishes to npm with provenance through OIDC and creates a GitHub release whenever the package.json version changes on main.

🤖 Generated with Claude Code

[reececomo]

✅ LGTM — Clean allocation-free retargeting with solid bounds checks, re-entrancy fallback and thorough tests - Couple of minor nits