Lab — ship the adapter — step 2 of 3

Write audit_qlora_payload(p) — the judgment gate on the train config. Checks run in this order; the first failure wins:

1. p["job"] is not "ci-triage-labeler" → "hold: job must be ci-triage-labeler"
2. p["base_model"] not in the allowed 0.5–3B Qwen2.5 Instruct set → "hold: base must be Qwen2.5 0.5-3B Instruct"
3. p["method"] is not "qlora" → "hold: method must be qlora"
4. p["bits"] is not 4 → "hold: QLoRA wants 4-bit"
5. p["lora_r"] not in (8, 16) → "hold: lora_r should be 8 or 16"
6. p["reviewer"] is not "human" → "hold: rows need a human reviewer"
7. p["split_unit"] not in ("company", "doc") → "hold: split by company or doc, not row"
8. set(p["train_ids"]) overlaps set(p["eval_ids"]) → "hold: train/eval ids overlap"
9. otherwise → "pass"

Expected output:

good      pass
70b       hold: base must be Qwen2.5 0.5-3B Instruct
no-review hold: rows need a human reviewer
leak      hold: train/eval ids overlap