[Rohit7sali]

Support custom-code backbones (Nandi, Lumma) and training on from decision checkpoints

Adds inference and fine-tuning for FrontiersMind's models:

Model	Inference	Fine-tuning
Lumma-fev-0.1B-v2 (custom Nandi code)	yes, trust_remote_code=True / --trust-remote-code	trains on from it
Lumma-fev-4B-v2, Lumma-fev-27B-v2 (Qwen3.5)	yes (already loaded; now documented)	trains on from them (LoRA / QLoRA)
Nandi-Mini-150M, Lumma-0.6B-Base (custom code)	n/a	yes
Other Hub decoders with their own modeling code	yes	yes, model.trust_remote_code: true

Changes

• trust_remote_code end to end: new --trust-remote-code flag on eval / predict / serve; LoRA merge passes the flag along; exports copy the model's .py files and auto_map, so checkpoints load on their own. Loading without the flag fails with a message that names it.
• Nandi/Lumma plain prompt (data.prompt.plain: auto | decisionformer | lumma): <bos> + prompt + \n\nAnswer:, tokenized with the fast tokenizer so the custom encode() cannot add a second BOS. auto picks lumma for Nandi/Lumma model types; the choice is saved as plain_prompt in decision_config.json. Other models keep the existing prompt byte for byte.
• Factorized embeddings: the code readout starts from the LM head through lm_head_proj, so it reproduces the base model's code-token logits at step 0.
• Training on from a decision checkpoint: when model.backbone has a decision_config.json, its readout seeds the code_readout head, and its prompt, attention mode and codes carry over unless set in the config. Adapter-only checkpoints are refused with a clear message.
• Answer-code detection uses the fast tokenizer directly (same ids for standard tokenizers; robust to custom encode overrides).
• Recipes: nandi_mini, lumma_base (LoRA), lumma_fev_continue.
• Docs: README sections on custom-code backbones and on training on from a decision model.

Tests

• tests/test_custom_models.py: 18 offline tests on a tiny custom-code model built like Nandi/Lumma (factorized embedding, lm_head_proj, BOS-prepending tokenizer). They cover the exact prompt tokens, readout = LM-head logits, train → export → reload, LoRA merge, the refusal without the flag, training on from a 4B/27B-v2-style and a 0.1B-v2-style checkpoint (identical outputs at step 0), the CLI flag and the recipes.
• tests/test_hub_models.py: real checkpoints, opt in with DECISIONFORMER_HUB_TESTS=1 (DECISIONFORMER_HUB_LARGE=1 adds 4B/27B-v2 on a GPU). On an RTX 3060: 8 passed, 1 skipped (the lumma-fev package comparison, skipped when the package is not installed; it passed in a local run), 27B not run. Includes Lumma-fev-0.1B-v2 matching an independent re-implementation of the lumma-fev runtime, Nandi / Lumma base fine-tuning, and Lumma-fev-4B-v2 load + LoRA training.
• CLI smoke runs on GPU: train nandi_mini → eval --trust-remote-code (reproduces the training-time accuracy exactly) → serve --trust-remote-code answers /v1/systemone. train local_smollm2 → eval unchanged for standard models.
• Full offline suite: no new failures. The 5 gemma2-sdpa failures in tests/test_attention.py also fail on main since transformers 5.21 (Gemma2 no longer accepts SDPA); they will be fixed in a separate PR.