feat: fold sift bench into the binary and read safetensors

[maxgfr]

The last two open items. TODO.md is now 14 done, 3 open.

sift bench

The sift-bench crate is gone — probe and models had become sift engines and sift ls. The harness now drives Ollama as well as LM Studio.

Ollama is deliberately not a copy of the LM Studio driver. Its native API reports eval_count and eval_duration, so its figure excludes prompt processing and HTTP overhead; the OpenAI shape LM Studio speaks reports no decode duration at all, so that path is wall clock. Each run records which clock produced it rather than presenting two different measurements as one.

--engine auto refuses to choose when both servers are up: silently preferring one would label a measurement with the wrong engine.

Runs append to ~/.sift/bench.jsonl. Two hand-set efficiency factors decide every tok/s figure sift prints, one still a guess — a measurement kept in a terminal scrollback cannot replace either.

Safetensors

fit, route and inspect read it. Each tensor's size is the difference of its data_offsets, never dtype × element count — same reasoning as reading real headers instead of shipping a catalog: an FP8 or MXFP4 variant this code has never heard of is sized exactly right by arithmetic that cannot go stale.

The header carries no architecture. GGUF keeps block_count and head_count_kv in its own header; a safetensors repo keeps them in config.json under different names. Rather than branch inside the KV arithmetic — where a mistake is most expensive — both formats normalise into ArchFacts, and ModelShape owns lightweight entries instead of borrowing GGUF tensors.

Three things that would have been quietly wrong:

• route dispatches on format. mlx-lm reads safetensors, llama.cpp does not. Assuming GGUF recommended an engine that cannot open the file.
• The download glob carries the extension. A .gguf pattern on a safetensors set matches nothing, and the user finds out only after the command runs.
• MoE experts are counted by scanning names, not probing indices. A shard can hold experts 40–63 of a layer and nothing else; probing from zero stops at the first gap and reports a fraction of the model.

Verified on allenai/OLMoE-1B-7B-0924-Instruct:

  quant               size     fits    bpw    GB/token  est tok/s
  model           13.84 GB    tight  16.00       2.564         13  moe, 3 parts
  download   huggingface-cli download … --include "model-*.safetensors"

16.00 bpw is BF16 exactly — the arithmetic checks out against a known quantity.

130 tests, clippy clean on macOS, Linux and Windows targets.

[github-actions]

🎉 This PR is included in version 0.3.0 🎉

The release is available on GitHub release

Your semantic-release bot 📦🚀