v1.0.0

First release: a fast, pure Dart Brotli (RFC 7932) encoder and decoder, with
no dependencies.

pub.dev: https://pub.dev/packages/brotli_fast

dependencies:
  brotli_fast: ^1.0.0

• API: brotliEncode / brotliDecode, the dart:convert codec brotli
  (BrotliCodec, BrotliEncoder, BrotliDecoder) with chunked stream
  conversion, BrotliDictionary for shared dictionaries, and
  BrotliException with typed BrotliErrorCodes.
• Encoder: qualities 0–11 and window sizes 10–24 bits, plus the
  large-window format (25–30). Qualities 0–5 use a fast tier; 6–11 add
  context modeling, block splitting, distance-parameter search and, at 10–11,
  cost-guided parsing.
• Decoder: one-shot and streaming, with a decompression-bomb limit,
  compound and serialized shared dictionaries, and large-window streams on
  opt-in.
• Platforms: VM, AOT, Flutter and the web (dart2js and dart2wasm). The
  encoder produces byte-identical output on all of them.
• Interoperability: CI cross-checks against the reference brotli CLI
  in both directions.
• License: Apache 2.0. The bundled google/brotli dictionary, tables and
  test corpus keep their MIT notice in
  NOTICE.

Inspired by brotli_neo (MIT), as a
full rewrite. Compared with it:

• Fixes four decoder bugs that rejected or silently misdecoded valid streams
  on the VM and WASM:
  • long literal runs and long copies (10–24 extra bits) were read without
    refilling the bit window;
  • a meta-block that exactly filled the ring buffer wrapped its position
    before the ring could grow;
  • with small windows (10–17 bits), the one-shot fast path mistook distances
    past the wrapped ring position for dictionary references, and read copy
    sources that wrapped the ring end from past the window;
  • concurrent decoders (two streams through brotli.decoder, or a stream
    and a one-shot decode) shared one pooled ring buffer and corrupted each
    other.
• Streamed decoding now rejects trailing data after the end of the stream,
  like one-shot decoding, however the input is chunked (it was ignored when
  it arrived in a later chunk).
• Fixes an encoder crash (RangeError) at qualities 2–5 on random-looking
  input.
• The fast tier's hash is now exact under dart2js, so web output matches the
  VM.
• Faster encoding with byte-identical output. Against brotli_neo 0.2.0 on
  the same machine (AOT): text q5 1.7× (858 vs 502 MB/s), text q9 2.2×
  (93 vs 43 MB/s), binary q0–q5 +24–41%. See
  benchmark/HISTORY.md.
• Web-motivated code has per-platform implementations, so the VM, AOT and
  dart2wasm no longer pay for dart2js limits: unaligned reads and 8-byte
  match-length comparison use ByteData there, while dart2js keeps its
  32-bit-safe paths.