v0.8.2

v0.8.2

Performance

• Removed the natural 3-byte LZ77 parse candidate and its writer pass from
  q4-q8; q3 and below keep it. In the interleaved same-time comparison against
  v0.8.1, aggregate q4-q8 encode time decreased by 20.6%: native 64 KiB and
  128 KiB rows by 16.5%-25.5%, wasm-gc rows by 16.3%-23.6%. Encoded size was
  unchanged on every measured row, including Silesia 2 MiB. Candidates are
  selected by exact cost, so dropping one can only leave output unchanged or
  larger; no measured input regressed.
• Cached the identity and mixed static-dictionary encode indexes for the
  process lifetime and materialized each entry's transformed output bytes into
  a flat pool, replacing per-byte transform re-derivation during match
  verification. In the same-time comparison, q10 encode time decreased by
  16.4% (native) and 13.5% (wasm-gc) on 64 KiB and by 10.5% and 7.8% on
  128 KiB; q9 64 KiB decreased by 10.8% and 11.9%; q11 rows decreased by 0.8%
  to 5.2%. Encoded output is byte-identical. Once a dictionary candidate has
  run, the caches retain about 2.4 MB (mixed) and 7.9 MB (identity) for the
  process lifetime; nothing is retained before that.

Tests

• Added white-box coverage verifying that the materialized dictionary output
  pool matches the per-byte transform machinery it replaced, and that both
  index caches return the same instance on repeat calls.

v0.8.1

v0.8.1

Changed

• Migrated the module manifest from moon.mod.json to moon.mod and updated
  the codebase for the current MoonBit formatter and checker without changing
  the public API.
• Enabled the bounded optimal-parse pass for q11 chunks up to 1 MiB. On the
  measured Silesia inputs, q11 output decreased from 21,302 to 20,904 bytes at
  64 KiB and from 264,341 to 256,952 bytes at 1 MiB, with higher encode time.

Performance

• Reduced q2 encode time by 11.3% on native and 13.4% on wasm-gc for the
  measured 64 KiB input, with encoded size unchanged.
• Adjusted recent-distance probing for q4-q8. Across the measured 64 KiB and
  128 KiB rows, aggregate encode time decreased by 10.6% and aggregate encoded
  size increased by 0.116%.

Tools and docs

• Updated the release benchmark report and checked-in benchmark data.
• Fixed the release validation script for Nushell 0.114 strict return-type
  checking by returning elapsed milliseconds as an integer.

v0.8.0

v0.8.0

Changed

• Changed Brotli q0/q1 encoding from stored-only meta-blocks to compressed
  low-quality streams.
• Tuned q0/q1 low-quality search and writer selection for a practical
  speed/size balance while keeping the measured q2 guard rows unchanged.

Performance

• On the Silesia 64 KiB and 128 KiB benchmark slices, q0/q1 encoded sizes are
  now within +2.38% of Google Brotli in the measured rows.
• In the same target-perf runs, q0 native encode min time was 3.95 ms versus
  Google Brotli 4.21 ms on 64 KiB; q1 native encode min time was 4.85 ms
  versus Google Brotli 4.30 ms on 64 KiB.

v0.7.1

v0.7.1

Performance

• Widened the Brotli decoder bit-reader accumulator to 64 bits so a 4-byte
  refill can be appended without discarding remaining buffered bits. Same-time
  decode-compare validation passed across the checked wasm-gc and native
  rows, with an aggregate decode improvement of about 1.1% in that run.

Compatibility

• Updated internal checked-error handling and tests for moonc v0.10.0
  warnings without changing the public API.

Tools and docs

• Removed the native benchmark cc-o0 compiler workaround now that default
  native release builds complete reliably with the current MoonBit toolchain.
• Changed the benchmark report and decode comparison defaults to
  --repeats 20, and documented that MoonBit timings amortize one moon run
  process startup across repeats.
• Regenerated docs/brotli_release_report.md and the checked-in current
  benchmark JSONL files with the updated harness.

v0.7.0

v0.7.0

Changed

• UnbrotliStream now decodes incrementally: it emits output through ondata
  as complete Brotli meta-blocks become available (intermediate calls report
  is_final=false), instead of buffering all input and emitting once on the
  final push. Consumers that assumed a single ondata call with the whole
  result must now concatenate the emitted chunks.
• UnbrotliStream now ignores UnbrotliOptions.out; streaming output is
  delivered only through ondata. One-shot unbrotli_sync still honors out.
• UnbrotliStream::push now rejects non-empty input pushed after the stream has
  finished.

Added

• BrotliDistanceRing::clone in the common package, used by the streaming
  decoder to speculatively decode a meta-block and roll back on short input.

Tests

• Added UnbrotliStream coverage: output equivalence against one-shot
  unbrotli_sync across every input split point, byte-at-a-time decoding,
  max_input_size enforcement, post-finish rejection, the empty-input final
  marker, and truncated-input errors.
• Added white-box coverage for common, decode, and encode helpers.

Tools and docs

• Documented the UnbrotliStream streaming trade-offs in the README: output is
  retained for the stream lifetime (memory is not window-bounded), each push
  may re-decode the in-progress meta-block, and out is ignored.
• Reworked the coverage-analysis script (tools/analyze-coverage.nu) and
  COVERAGE_ANALYSIS.md.

v0.6.0

v0.6.0

Performance

• Adjusted Brotli encoder candidate gating for quality levels 3 through 11.
  The current selection keeps intermediate LZ77 candidates for q4-q8, narrows
  four-byte high-quality and mixed-dictionary candidates to q10/q11, and avoids
  several q3/q9 candidate passes that were recorded as rejected trials.
• In the regenerated just bench report, q3 encode min timings improve by
  1.78x-2.49x versus v0.5.0 on the Silesia 64 KiB and 128 KiB inputs. The q3
  output size increases by 3.3% on 64 KiB and 4.2% on 128 KiB.
• The same report shows q9 encode min timings improving by 1.57x-3.48x. The
  64 KiB q9 output is unchanged; the 128 KiB q9 output grows from 38,926 to
  40,207 bytes, changing its Google-size delta from -1.46% to +1.78%.
• q10 and q11 encode min timings improve by 1.33x-1.54x with unchanged output
  size in the measured rows. q2 also improves by 1.10x-1.23x with unchanged
  output size.
• q0 and q1 output sizes are unchanged, with measured min timings 2.7%-13.8%
  slower. q4 through q8 output sizes are unchanged, with measured min timings
  ranging from 10.6% faster to 7.6% slower depending on quality, target, and
  input.
• Added small decoder hot-path changes for command info lookup, short-distance
  handling, explicit distance decoding, and validated Huffman tree-group access.
  The regenerated decode report shows native min timings 5.7%-7.4% faster than
  v0.5.0; wasm-gc rows range from 3.6% faster to 0.3% slower.

Tools and docs

• Moved Brotli tooling from tools/brotli/ to direct tools/ subdirectories
  and updated just recipes and documentation paths.
• Added same-time benchmark harnesses for comparing encode and decode behavior
  against a baseline git ref.
• Updated the size verification script to report size as a filesize value and
  to select native .exe artifacts explicitly.
• Documented recent js, wasm-gc, and native size-verification results in
  the README.

Tests

• Added white-box coverage for context-map validation, command-info table
  lookup, validated short-distance decoding, explicit distance formulas, and
  encoder candidate gating.

v0.5.0

v0.5.0

Performance

• Improved Brotli q4-q8 encode time by pruning expensive candidate passes,
  reusing match-chain tables, disabling lazy matching for those qualities, and
  preferring the context8 writer when applicable.
• In the 5591fe6 target-perf report, q4-q8 encode min timings improved
  from v0.3.1 by 2.15x-3.01x on wasm-gc and 3.08x-3.95x on native for the
  Silesia 64 KiB input, and by 2.57x-3.25x on wasm-gc and 3.39x-4.88x on
  native for the Silesia 128 KiB input. The same benchmark rows grew output
  size by 1.77%-2.63%.
• The same report shows q2-q3 encode min timings improving by 1.11x-1.34x and
  q9-q11 by 1.13x-1.77x on the Silesia 64 KiB and 128 KiB inputs, with
  unchanged output sizes in those rows.
• Optimized Brotli encoder prefix lookup paths for command lengths and
  distances, plus context8 literal-tree reuse.
• Added a decoder fast path for compressed meta-blocks with one literal tree
  and one literal block type, avoiding per-literal block tracking.
• Decode benchmark min timings for the 1 MiB Google-produced Silesia streams
  ranged from 1.6% slower to 3.5% faster than v0.3.1, so no material decode
  speedup is claimed for this release.

Tests and docs

• Added white-box coverage for command-prefix tables, distance-prefix mapping,
  hash-chain boundary pre-checks, candidate gating, shared previous-match
  tables, and lazy-match deferral.
• Added roundtrip coverage for literal-heavy streams and lazy-match deferral
  inputs across multiple Brotli qualities.
• Regenerated the 0.5.0 Brotli release benchmark report and current benchmark
  JSONL files.

v0.3.1

Full Changelog: v0.3.0...v0.3.1

v0.3.0

v0.3.0

Added

• Added just bench, backed by tools/brotli/bench/report.nu, to regenerate
  the Brotli release benchmark report and current benchmark JSONL files.
• Added checked-in Brotli benchmark data under docs/current-bench/, including
  Google Brotli 1 MiB comparison streams.

Performance

• Improved Brotli decode hot paths by sizing dynamic output buffers from the
  compressed input length, refilling the bit reader four bytes at a time, and
  skipping zero-width command extra-bit reads.
• Packed decoded Brotli Huffman table entries into FixedArray[Int] and
  updated encoder/decoder users to avoid boxed table-entry loads.
• Optimized compressed meta-block literal decoding with a single-literal-tree
  fast path, a streamlined multi-tree path, and one reused command record per
  meta-block.

Tests and docs

• Added white-box coverage for the decode optimizations, packed Huffman table
  behavior, output capacity sizing, and command record reuse.
• Regenerated docs/brotli_release_report.md from the current release-build
  benchmarks for wasm-gc and native targets against Google Brotli CLI 1.2.0.

v0.2.0

v0.2.0

Breaking

• Renamed Brotli error APIs from FzipError, FzipErrorCode, fzip_err, and
  fzip_error_code_to_int to FbrError, FbrErrorCode, fbr_err, and
  fbr_error_code_to_int.
• Renamed the stream callback wrapper from FlateStreamHandler to
  FbrStreamHandler.

Changed

• Updated package metadata and README examples for version 0.2.0.
• Updated Brotli verification tooling and generated API metadata to use the
  hustcer/fbr package naming consistently.