feat: make every adapter reproducible from a PEFT or Unsloth recipe

[Yash-Chindam]

What

Closes the adapter half of the model lifecycle in the spec (§6, §7.5): "PEFT/Unsloth creates LoRA or QLoRA adapters" from "immutable base-model revisions".

How

• config/adapters/<id>.yaml: one reviewed recipe per catalog adapter, holding the base commit, dataset version and LoRA hyperparameters.
• llm_router.adapters:
  • check fails when a catalog adapter has no recipe, or a recipe names a different base, dataset, domain or quantization than its catalog entry. CD runs it.
  • plan prints the exact arguments PEFT (LoraConfig, BitsAndBytesConfig) or Unsloth (from_pretrained, get_peft_model) would be given.
  • train runs the job with whichever framework the recipe selects.
  • register runs the artifact scan on a finished adapter and prints its catalog entry: stage development, no measured gain, revision equal to the artifact's digest.
• A recipe is rejected if its base revision is a branch or tag rather than a commit, or if its rank is above what the serving configuration loads.
• Training never writes checkpoints into the artifact, since they carry pickled optimizer state the scanner would refuse.

Tests

tests/unit/test_adapters.py (22 tests): committed recipes agree with the catalog; each kind of disagreement is reported; unreproducible or unservable recipes are rejected; PEFT and Unsloth arguments; scan-gated registration; CLI. 100% coverage of everything except train.

Limits

• train has never been executed. It needs a GPU and the training extra; it is written against the documented PEFT, Transformers and Unsloth interfaces and excluded from coverage.
• hf_repo and hf_revision in the committed recipes are placeholders, because the catalog's base models are mocks.
• The training datasets the recipes point at are not in this repository.

🤖 Generated with Claude Code