caldrix v0.2.0

Added

• caldrix pool and caldrix pool --protocol-version: a daemon on a Unix socket that holds loaded models
  and serves them to local clients over a framed protocol (caldrix pool).
  It has run on the CPU lane, and on Metal on one Apple M1 Max at max context 2048 and one request at
  8192; it has not been run on AMD GPUs.
• caldrix models: search, inspect and list models from the downloaded model index
  (caldrix models).
• ollama:<name>[:<tag>] model references, downloaded from the Ollama registry
  (Ollama tags).
• Ollama Q4_K_M and Q8_0 tags, read and sized before any download and run at an f32 plan; an explicit
  --dtype bf16 and --weight-budget are refused for them by name
  (Ollama tags). On Metal, on one Apple M1 Max, three small
  tags and an 8 B Q4_K_M tag were token-exact in f32 and caldrix run of the 8 B tag completed. On the CPU
  lane at x86-64-v3, four small tags were token-exact in f32 at the desk on one machine, a desk reading,
  not a run. On AMD GPUs the hip lane compiles them and holds a quantized model to one request at a
  time (--max-batch lowers from 8 to 1 with a startup line, and an explicit --max-batch 8 is refused
  by name); nothing quantized has run on an AMD device.
• Sampling for caldrix run: --temperature, --top-k, --top-p, --min-p, --repetition-penalty and
  --seed, and the chat's commands for them (Sampling).
• Sampling for caldrix serve: the same six flags set the server's defaults, and /v1 requests accept
  temperature, top_p, top_k, min_p, repetition_penalty and seed per request, where v0.1.2
  refused a non-neutral temperature, top_p, top_k, min_p or repetition_penalty with a 400
  naming the field and accepted seed while decoding greedily
  (Sampling). On the AMD lane with batching on (--max-batch 8,
  its default there) the GPU picks each sampled token, so the same seed can give a different reply than a
  pick on the CPU; --sampler host moves that pick back to the CPU. Every other lane, and
  --max-batch 1, always picks on the CPU.
• --system TEXT|@FILE to set a system prompt, --raw to send the prompt as plain text, and --json for
  caldrix run's output as JSON.
• The qwen3-30b-a3b alias and the Qwen3MoeForCausalLM architecture, verified by its token gates on
  one AMD MI300X with every expert resident, on one Apple M1 Max streamed under a 56 GiB
  --weight-budget, and on the CPU lane streamed under a 20 GiB budget; --max-batch above 1 is refused
  for it (Aliases).
• CALDRIX_OFFLINE, like HF_HUB_OFFLINE but covering every network edge caldrix has: the Hub, the model
  index and the Ollama registry (Environment).
• MistralForCausalLM is accepted by the import; it is not verified on hardware.
• Llama 3 rope scaling, applied from a config.json and from a GGUF's rope_freqs.weight.

Changed

• Generation is no longer greedy-only: it is greedy unless a sampling flag asks otherwise.
• The Metal prefill and decode attention are tiled over the context. On one Apple M1 Max a 6002-token
  prompt ran at max context 7680, and max context 8192 loaded and ran.
• caldrix run opens the chat screen on a terminal; --plain keeps the line-by-line chat.

Known limitations

• Quantized tags have not been run on AMD GPUs; on the CPU lane they have only a desk reading, not a run;
  quantized safetensors checkpoints are refused.
• caldrix pool has not been run on AMD GPUs.
• NVIDIA GPUs are not supported.