v1.0.5

• Performance: optimized encode/decode for order-1, order-2 (and order-3) context models.
  • ContextModel and ContextState:
    • Added _LazyContextModel base class that allocates one Fenwick tree per
      context lazily, on first use, instead of eagerly creating a tree for
      every possible context.
    • init/reset now cost is proportional to the contexts actually used
      rather than to the full context space (size² for order-2, more for
      order-3), eliminating millions of wasted operations per call.
    • Recycled trees are reused via an internal pool to avoid re-allocation
      across encode/decode runs.
    • ContextModelOrder1/Order2/Order3 now extend _LazyContextModel and
      flatten their context space into a single slot table; output is
      byte-identical to the previous implementation.
  • Fenwick:
    • Maintains a parallel direct-frequency array, so single-symbol frequency
      lookups (freqOf) are O(1). This removes one of the two cumulative tree
      walks per symbol in encode.
    • Added findWithLow, which decodes a symbol and returns its lower
      cumulative bound (sum(symbol - 1)) in a single descent, removing two
      tree walks per symbol in decode.
    • init now builds the all-ones tree directly in O(n) (each node i holds
      i & -i) instead of O(n log n), the dominant cost when many contexts are
      visited (e.g. high-order models on high-entropy data: ~4–5x faster).
    • rescale and toFrequencyList now use the direct-frequency array
      (O(n) instead of O(n log n)).
• Added benchmark/arithmetic_coder_benchmark.dart to track compression ratio
  and encode/decode throughput across orders and datasets for future metrics.
• Tests:
  • Added test/fenwick_test.dart covering the direct-frequency array,
    freqOf, findWithLow, the O(n) init, rescale/rebuild/reset sync,
    and the sum/cumulative invariants (including through many rescales).
  • Added test/context_model_test.dart covering the factory, totalSize,
    contextIndex mapping, lazy allocation, order-3 bucketing, and pool reuse
    across reset/init.
  • Extended test/arithmetic_coder_test.dart with cached-model reuse across
    sequential calls, deterministic re-encoding, and rescale-triggering inputs.