fix: harden quantized checkpoint integration

tests/fixtures/quant_dequant_provenance.json

@@ -0,0 +1,21 @@
+{
+  "schema_version": 1,
+  "origin": "synthetic-runtime-generated",
+  "external_data": false,
+  "contract_sources": [
+    "https://docs.vllm.ai/projects/llm-compressor/en/stable/guides/entrypoints/convert/",
+    "https://github.com/huggingface/transformers/blob/main/docs/source/en/quantization/compressed_tensors.md"
+  ],
+  "generator": "tests/test_quant_dequant.py",
+  "torch_manual_seeds": [0, 1, 2, 3, 4, 5, 6, 7, 8],
+  "formats": {
+    "fp8": "IEEE-like E4M3 values generated by torch.float8_e4m3fn",
+    "nvfp4": "E2M1 values generated from OBLITERATUS's explicit 16-value test oracle"
+  },
+  "oracles": {
+    "fp8_relative_l2_max": 0.05,
+    "nvfp4_relative_l2_max": 0.15,
+    "nvfp4_cosine_min": 0.995
+  },
+  "reproducibility": "Fixtures are regenerated deterministically during each test run; no model weights, datasets, or network resources are consumed."
+}