df11pack v0.1.0

First release.

Compression

• DFloat11 output byte-identical to the official compressor (pip dfloat11
  0.5.0), as whole files: on real Qwen3-8B, all 39 files match the official output
  and the published DFloat11/Qwen3-8B-DF11 release, in 40 s and 6.6 GB of RAM
  against 37 min and 27.3 GB.
• Four layouts: transformers, diffusers, diffusers single-file, ComfyUI single-file.
• 35 built-in model definitions: every official DFloat11 release with a
  dfloat11_config, and every model in ComfyUI-DFloat11-Extended. Custom
  definitions by TOML path or DF11PACK_ARCH_DIR.
• Parallel chunked encoder; memory-aware worker count (respects container limits);
  --ram, --workers; sequential reads for spinning disks.
• Atomic output: a failed or interrupted run never leaves a half-written file.
• Refuses what it cannot do exactly, rather than guessing: non-BF16 sources, and the
  rare ambiguous tie in the official 32-bit code-length limiter.

Verification

• --safe: every unit decoded and checked against the source before it is written.
• df11pack verify: structure without the source; sampled or full decode with it;
  exit codes 0 / 1 / 2 for pass / fail / could not check.

Opt-in extras (clearly marked, never the default)

• --luts=correct: zero-fills leftover LUT entries; proven unreachable by the
  official CUDA kernel.
• --hashes: per-tensor SHA-256 in the header; verify then catches corruption
  without the source.
• --index idx8: experimental 8-bit block index with an escape table, 0.66%
  smaller on Qwen3-8B. Not DFloat11; no current loader reads it.

Known limits

• Image and video models have not yet been loaded from df11pack output in ComfyUI or
  diffusers on a GPU (files are byte-identical to what those loaders read).
• Gated repositories (Gemma-3, SD3.5, FLUX.1-dev diffusers, Llama) were not checked
  against real checkpoints.
• verify --level full decodes on one core (~10 minutes for a 12B model).
• The legacy DFloat11 0.1.0 format (pickled releases) is not supported.