docs: add Cua-S1 4B 0.2 inference contract

[twu3202]

Purpose

Part of #10. This records the pinned upstream revisions, the inference contract, the /v1/systemone request mapping and the comparison tolerances for Cua-S1 4B 0.2, before any worker or engine code lands. It adds src/models/cua_s1/README.md and a row for that directory in the README layout table.

The first target is the text adapter on CUDA, loaded directly through Hugging Face Transformers and PEFT. The multimodal adapter, score and noul questions, and Metal are listed as not covered yet.

Three choices in the mapping are worth a look:

• confidence is the normalized entropy 1 - H(p) / ln(n), the same as the LAYA worker reports, so both models behind the frontend use one definition. The Jev API does not publish its formula, and upstream's chooser reports p_max.
• Questions the adapters cannot answer (score, noul, more than 26 options) return 422 instead of being mapped onto letters.
• Ties go to the earliest option. Upstream's chooser returns an error on a tie instead.

The second commit defines the error body ({"detail": ...}, as the LAYA worker returns) and splits rejections into 400 for bodies that are not usable JSON, including repeated keys, and 422 for requests the model cannot answer. This keeps the text worker and the multimodal worker in #12 consistent.

Test Plan

I checked the factual statements against the pinned revisions in upstream's four-b environment (Python 3.12, torch 2.14.0, Transformers 5.17.0, PEFT 0.21.0):

• Downloaded both model revisions and verified every file with upstream's ci/fetch_pinned_weights.py --verify-only.
• Read the adapter safetensors headers for target modules, layers, rank, alpha and dtype.
• Loaded the tokenizer and chat template to check the letter token ids, special tokens, the prompt suffix, and the prompt length for upstream's positive fixture.
• Loaded FourBModel with the text adapter on CPU and scored upstream's two fixtures through upstream's chooser, once with the default template and once with enable_thinking=False.
• Scored the same two fixtures with FourBModel on CUDA (one RTX 6000 Ada, torch 2.14.0+cu130) in bfloat16.

System1-Omni Version / Commit: 7aa2c30, based on 66f689f.

Test Result

• All 20 files match the lock: 9.34 GB for the base model and 187 MB for the adapters.
• The text adapter attaches 128 LoRA modules (21.2M parameters) to Qwen3_5ForCausalLM, which matches the layers listed in the README.
• Letters A to Z are single tokens with ids 32 to 57. The positive fixture prompt is 218 tokens and ends with <|im_start|>assistant\n<think>\n.
• The fixture probabilities are close to upstream's published results. Upstream measured on MPS, so these numbers are context rather than a pass target:

Fixture (expected option)	CPU bfloat16	CPU float32	CUDA bfloat16	Upstream, MPS bfloat16	Upstream, MPS float16
Positive (submit-form)	0.976	0.975	0.973	0.973	0.975
Negative (abstain)	0.425	0.431	0.435	0.455	0.430

• With enable_thinking=False, the expected option's probability shifts by 0.007 to 0.031 (float32: positive 0.985, negative 0.455), so the suffix belongs in the contract.
• The full CUDA comparison over the fixed input set comes with the text worker PR.

[hsliuustc0106]

I suggest not to add third level folder name here and revise it as src/models/ for model implementation

[twu3202]

Done in a59d5ea: the layout table now has one src/models/ row for model implementations, in place of the Laya and Cua-S1 rows. #13 and #19 include the change.

[hsliuustc0106]

Reviewed the inference contract against the pinned upstream source and model metadata. No actionable findings. Also checked the latest README layout-table update at a59d5ea. Validation was source-based; model inference and GPU comparisons were not rerun.