diff --git a/README.md b/README.md
index 3667a60..03cf5a3 100644
--- a/README.md
+++ b/README.md
@@ -486,7 +486,7 @@ Quantization roughly halves the GPU count at each step down. A 70B model that ne
 | FP8 | DeepSeek-style block-wise (`weight_scale_inv` + `weight_block_size`), compressed-tensors per-channel, ModelOpt FP8 |
 | NVFP4 | ModelOpt (`weight` + `weight_scale` + `weight_scale_2`), compressed-tensors NVFP4 |
 
-Two things to know: peak VRAM is the **BF16 size** of the model (not the quantized size), and the output is saved as BF16 — re-quantize afterward with llm-compressor or modelopt if you want a quantized serving artifact. Other quantization schemes (fbgemm, quanto, W4A4) fail loudly with a message naming the scheme.
+Two things to know: peak VRAM is the **BF16 size** of the model (not the quantized size), and the output is saved as BF16 — re-quantize afterward with llm-compressor or modelopt if you want a quantized serving artifact. Unsupported or ambiguous quantization layouts fail loudly with a message naming the scheme.
 
 ### GPU calculator
 
diff --git a/ci/conditional-test-policy.json b/ci/conditional-test-policy.json
index 2312d65..043bde4 100644
--- a/ci/conditional-test-policy.json
+++ b/ci/conditional-test-policy.json
@@ -128,7 +128,11 @@
       "runner": "self-hosted, linux, x64, cuda",
       "prerequisites": "ENABLE_CUDA_GATE=true and a dedicated CUDA runner",
       "expected_cost": "under 15 self-hosted runner-minutes",
-      "coverage_paths": ["obliteratus/device.py", "obliteratus/models/loader.py"]
+      "coverage_paths": [
+        "obliteratus/device.py",
+        "obliteratus/models/loader.py",
+        "obliteratus/models/quant_dequant.py"
+      ]
     },
     {
       "id": "bitsandbytes-runtime",
diff --git a/ci/test-risk-map.json b/ci/test-risk-map.json
index 31d0bb6..9de29aa 100644
--- a/ci/test-risk-map.json
+++ b/ci/test-risk-map.json
@@ -75,6 +75,7 @@
         "obliteratus/mlx_backend.py",
         "obliteratus/model_profile.py",
         "obliteratus/models/loader.py",
+        "obliteratus/models/quant_dequant.py",
         "obliteratus/presets.py",
         "obliteratus/runtime_contracts.py",
         "obliteratus/study_presets.py"
@@ -86,6 +87,7 @@
         "tests/test_config_properties.py",
         "tests/test_device_boundaries.py",
         "tests/test_loader_boundaries.py",
+        "tests/test_quant_dequant.py",
         "tests/test_mlx_backend_boundaries.py",
         "tests/test_model_profile.py",
         "tests/test_model_profile_contracts.py",
@@ -350,6 +352,13 @@
       "required_tests": ["tests/test_loader_boundaries.py", "tests/test_offline_integration.py"],
       "conditional_gates": ["cuda-runtime", "bitsandbytes-runtime"]
     },
+    {
+      "path": "obliteratus/models/quant_dequant.py",
+      "risk_class": "mixed-runtime",
+      "risk": "FP8 and NVFP4 format detection, numerical dequantization, Hub revision integrity, and checkpoint materialization",
+      "required_tests": ["tests/test_quant_dequant.py"],
+      "conditional_gates": ["cuda-runtime"]
+    },
     {
       "path": "obliteratus/mlx_backend.py",
       "risk_class": "mixed-runtime",
diff --git a/docs/theory_journal.md b/docs/theory_journal.md
index 6f2c180..0c47cc6 100644
--- a/docs/theory_journal.md
+++ b/docs/theory_journal.md
@@ -1821,10 +1821,13 @@ protocol; anything here that surprised us is marked **(finding)**.
    detection must accept both `_scale`/`_scale_2` suffixes and the dotted
    `.weight_scale` family **(finding: first state-dict test run failed on
    exactly this)**.
-4. **compressed-tensors stores reciprocal scales.** Dequant divides rather
-   than multiplies. Round-tripping reciprocals through FP8-E4M3 storage is
-   measurably coarser (cosine ~0.986 vs ~0.999 for direct storage on random
-   weights) — expected, not a bug.
+4. **compressed-tensors inverts only the global NVFP4 scale.** Its documented
+   converter maps ModelOpt `weight` to `weight_packed`, leaves `weight_scale`
+   unchanged, and maps `weight_scale_2` to the reciprocal
+   `weight_global_scale`. Dequantization therefore multiplies by the FP8 block
+   scale and divides by the global scale. Treating both scales as reciprocal
+   produced a self-consistent synthetic test but did not match the published
+   checkpoint contract **(maintainer audit finding)**.
 5. **torch 2.13.0+cpu**: `torch.float4_e2m1fn_x2` exists but the runtime
    upcast probe fails on this CPU build, so the manual nibble-LUT path is
    what actually runs there; the probe design (rather than a version check)
@@ -1833,9 +1836,11 @@ protocol; anything here that surprised us is marked **(finding)**.
    so the FP8/NVFP4 loud-failure guard in `_dequantize_weight` must run
    *before* the bitsandbytes branch or the error message is misleading
    **(finding: unit test caught the wrong branch firing)**.
-7. NVFP4 round-trip error on random Gaussian weights: relative Frobenius
-   error ~0.10 direct / ~0.18 reciprocal, cosine >0.98 — this is the
-   intrinsic precision of the format, not implementation error.
+7. NVFP4 round-trip error on the deterministic random-Gaussian fixtures is
+   below 0.15 relative Frobenius error with cosine above 0.995 for both
+   ModelOpt and compressed-tensors global-scale storage. Seeds, construction,
+   and thresholds are recorded in
+   `tests/fixtures/quant_dequant_provenance.json`.
 8. **Real checkpoints are mixed-precision.** `nvidia/Nemotron-3-Nano-Omni-
    30B-A3B-Reasoning-NVFP4` ships `quant_method: modelopt` with
    `quant_algo: MIXED_PRECISION`: FP8 (8-bit float, per-tensor
@@ -1857,7 +1862,12 @@ Smoke-test results on real checkpoints: see Appendix E continuation below
 
 ---
 
-## Appendix E2: Nemotron-3-Nano-Omni smoke runs (2026-07-24, 1x A100-80GB)
+## Appendix E2: Contributor-reported Nemotron-3-Nano-Omni smoke runs (2026-07-24)
+
+The following A100 results were supplied with the contribution. They do not
+include immutable checkpoint revisions or retained machine-verifiable logs and
+were not independently reproduced during maintainer integration. They are
+historical engineering notes, not release evidence or a support claim.
 
 Real-checkpoint validation of the FP8/NVFP4 dequantization path. Both
 targets loaded with **zero quantized tensors left, zero NaNs, and
@@ -1944,10 +1954,11 @@ Notes:
 - The compressed-tensors per-channel FP8 dense alternate
   (`RedHatAI/Llama-3.1-8B-Instruct-FP8-dynamic`) was not run: the exact
   code path (`dequantize_fp8_per_channel`, direct scales) was already
-  exercised by the Nemotron FP8 checkpoint, and Llama-3.1 is a gated
-  repo. CT-NVFP4 (reciprocal scales) is covered by synthetic unit tests
-  only — no public checkpoint was needed after the Nemotron layout
-  parsed cleanly.
+  exercised by the contributor's Nemotron FP8 run, and Llama-3.1 is a gated
+  repo. CT-NVFP4 is covered by deterministic synthetic contract tests only;
+  those tests follow the published `weight_packed`/`weight_scale`/
+  `weight_global_scale` mapping. A pinned public-checkpoint run remains
+  necessary before making a runtime support claim.
 
 ---
 
diff --git a/obliteratus/abliterate.py b/obliteratus/abliterate.py
index a547f95..a6a8b19 100644
--- a/obliteratus/abliterate.py
+++ b/obliteratus/abliterate.py
@@ -46,6 +46,20 @@ _QUANT_SCALE_ATTRS = (
     "weight_global_scale",
 )
 
+_FUSED_QUANT_SCALE_SUFFIXES = (
+    "_scale",
+    "_scale_2",
+    "_scale_inv",
+    "_global_scale",
+)
+
+
+def _has_fused_quant_scale(container: nn.Module, name: str) -> bool:
+    return any(
+        hasattr(container, f"{name}{suffix}")
+        for suffix in _FUSED_QUANT_SCALE_SUFFIXES
+    )
+
 # Reduce CUDA memory fragmentation for large models.  Must be set before any
 # CUDA allocations, so we do it at import time.  This is the PyTorch-recommended
 # fix for "reserved but unallocated" memory issues.
@@ -5021,7 +5035,9 @@ class AbliterationPipeline:
                 data = param.data
                 # FP8/packed fused params are quantized storage — the loader
                 # dequantizes them; raw upcasting here would corrupt scales.
-                if data.dtype in _FP8_DTYPES or data.dtype == torch.uint8:
+                if data.dtype in _FP8_DTYPES or (
+                    data.dtype == torch.uint8 and _has_fused_quant_scale(container, name)
+                ):
                     raise RuntimeError(
                         f"FP8/NVFP4 fused-expert weight '{name}' reached surgery "
                         f"without dequantization — the loader should have "
@@ -5611,8 +5627,13 @@ class AbliterationPipeline:
             if data.shape[-1] != hidden_dim and data.shape[-2] != hidden_dim:
                 continue
 
-            if (data.dtype in _FP8_DTYPES or data.dtype == torch.uint8) \
-                    and not hasattr(param, "quant_state"):
+            if (
+                data.dtype in _FP8_DTYPES
+                or (
+                    data.dtype == torch.uint8
+                    and _has_fused_quant_scale(container, pname)
+                )
+            ) and not hasattr(param, "quant_state"):
                 raise RuntimeError(
                     f"FP8/NVFP4 fused-expert weight '{pname}' reached surgery "
                     f"without dequantization — the loader should have "
@@ -5732,8 +5753,13 @@ class AbliterationPipeline:
             if data.shape[-1] != hidden_dim and data.shape[-2] != hidden_dim:
                 continue
 
-            if (data.dtype in _FP8_DTYPES or data.dtype == torch.uint8) \
-                    and not hasattr(param, "quant_state"):
+            if (
+                data.dtype in _FP8_DTYPES
+                or (
+                    data.dtype == torch.uint8
+                    and _has_fused_quant_scale(container, pname)
+                )
+            ) and not hasattr(param, "quant_state"):
                 raise RuntimeError(
                     f"FP8/NVFP4 fused-expert weight '{pname}' reached surgery "
                     f"without dequantization — the loader should have "
diff --git a/obliteratus/models/loader.py b/obliteratus/models/loader.py
index 9341f4e..b0eb208 100644
--- a/obliteratus/models/loader.py
+++ b/obliteratus/models/loader.py
@@ -259,45 +259,88 @@ except Exception:
 # legacy ``_tied_weights_keys`` (list of names or dict) across modules into
 # the {param_name: target_name} mapping accelerate expects.
 # Affected: Nemotron Omni (NemotronH_Nano_Omni_Reasoning_V3)
+def _install_all_tied_weights_keys_compat(model_cls) -> bool:
+    """Install the legacy tied-weight adapter only when the class lacks it."""
+    if hasattr(model_cls, "all_tied_weights_keys"):
+        return False
+
+    def _get_all_tied_weights_keys(self):
+        cached = self.__dict__.get("_all_tied_weights_keys_store")
+        if cached is not None:
+            return cached
+        expander = getattr(self, "get_expanded_tied_weights_keys", None)
+        if callable(expander):
+            try:
+                expanded = expander(all_submodels=True)
+            except Exception:
+                # Legacy remote implementations may still expose the modern
+                # method name while keeping an incompatible list contract.
+                pass
+            else:
+                self.__dict__["_all_tied_weights_keys_store"] = expanded
+                return expanded
+        tied: dict = {}
+        for module in self.modules():
+            keys = getattr(module, "_tied_weights_keys", None)
+            if not keys:
+                continue
+            if isinstance(keys, dict):
+                candidates = keys
+            else:  # legacy list of patterns/names — one tied group
+                names = list(keys)
+                candidates = {name: names[0] for name in names}
+            for name, target in candidates.items():
+                # Only expose keys that actually resolve on this model —
+                # multimodal wrappers inherit "lm_head.weight" from their
+                # LM submodule's config but have no top-level lm_head.
+                try:
+                    self.get_submodule(name.rsplit(".", 1)[0])
+                except AttributeError:
+                    continue
+                tied.setdefault(name, target)
+        return tied
+
+    def _set_all_tied_weights_keys(self, value):
+        self.__dict__["_all_tied_weights_keys_store"] = value
+
+    model_cls.all_tied_weights_keys = property(
+        _get_all_tied_weights_keys, _set_all_tied_weights_keys,
+    )
+    return True
+
+
 try:
     from transformers import PreTrainedModel as _PTM
-    if not hasattr(_PTM, "all_tied_weights_keys"):
-        def _get_all_tied_weights_keys(self):
-            cached = self.__dict__.get("_all_tied_weights_keys_store")
-            if cached is not None:
-                return cached
-            tied: dict = {}
-            for module in self.modules():
-                keys = getattr(module, "_tied_weights_keys", None)
-                if not keys:
-                    continue
-                if isinstance(keys, dict):
-                    candidates = keys
-                else:  # legacy list of patterns/names — one tied group
-                    names = list(keys)
-                    candidates = {name: names[0] for name in names}
-                for name, target in candidates.items():
-                    # Only expose keys that actually resolve on this model —
-                    # multimodal wrappers inherit "lm_head.weight" from their
-                    # LM submodule's config but have no top-level lm_head.
-                    try:
-                        self.get_submodule(name.rsplit(".", 1)[0])
-                    except AttributeError:
-                        continue
-                    tied.setdefault(name, target)
-            return tied
 
-        def _set_all_tied_weights_keys(self, value):
-            # transformers 5.x post_init() assigns this attribute.
-            self.__dict__["_all_tied_weights_keys_store"] = value
-
-        _PTM.all_tied_weights_keys = property(
-            _get_all_tied_weights_keys, _set_all_tied_weights_keys,
-        )
+    _install_all_tied_weights_keys_compat(_PTM)
 except Exception:
     pass
 
 
+def _force_eager_config(config) -> None:
+    """Recursively replace unavailable Flash Attention 2 selections."""
+    from transformers import PretrainedConfig
+
+    if getattr(config, "_attn_implementation", None) == "flash_attention_2":
+        config._attn_implementation = "eager"
+    for subconfig in vars(config).values():
+        if isinstance(subconfig, PretrainedConfig):
+            _force_eager_config(subconfig)
+
+
+def _config_uses_flash_attention_2(config) -> bool:
+    """Return whether this config or a nested Transformers config selects FA2."""
+    from transformers import PretrainedConfig
+
+    if getattr(config, "_attn_implementation", None) == "flash_attention_2":
+        return True
+    return any(
+        _config_uses_flash_attention_2(subconfig)
+        for subconfig in vars(config).values()
+        if isinstance(subconfig, PretrainedConfig)
+    )
+
+
 # ── Deferred shims ──────────────────────────────────────────────────
 # Some patches must wait until the _LazyModule has fully initialized
 # (it replaces its __dict__ during bootstrap).  We apply these once,
@@ -661,7 +704,12 @@ def load_model(
     # FP8 / NVFP4 checkpoints: dequantize to a plain float copy on disk,
     # then load through the normal float path.  Anything quantized that we
     # don't explicitly support fails loudly here, before any weight loads.
-    quant_detection = qd.detect_quant_scheme(model_name, token=token)
+    quant_detection = qd.detect_quant_scheme(
+        model_name,
+        token=token,
+        revision=revision,
+        local_files_only=local_files_only,
+    )
     _dequant_tmp = None
     _dequant_source = None
     if quant_detection.scheme is qd.QuantScheme.UNSUPPORTED:
@@ -677,7 +725,12 @@ def load_model(
         )
         _dequant_source = model_name
         _dequant_tmp, _ = qd.materialize_dequantized_checkpoint(
-            model_name, quant_detection, out_dtype=torch_dtype, token=token,
+            model_name,
+            quant_detection,
+            out_dtype=torch_dtype,
+            token=token,
+            revision=revision,
+            local_files_only=local_files_only,
         )
         model_name = _dequant_tmp
         config = AutoConfig.from_pretrained(
@@ -709,25 +762,7 @@ def load_model(
     # If flash-attn isn't installed, loading would fail — fall back to eager.
     # Multimodal wrappers pin it on nested sub-configs too (e.g. Nemotron
     # Omni's llm_config), so walk those as well.
-    def _force_eager_cfg(cfg) -> None:
-        from transformers import PretrainedConfig as _PTC
-        if getattr(cfg, "_attn_implementation", None) == "flash_attention_2":
-            cfg._attn_implementation = "eager"
-        for sub in vars(cfg).values():
-            if isinstance(sub, _PTC):
-                _force_eager_cfg(sub)
-
-    def _has_flash_attention(cfg) -> bool:
-        from transformers import PretrainedConfig as _PTC
-        if getattr(cfg, "_attn_implementation", None) == "flash_attention_2":
-            return True
-        return any(
-            _has_flash_attention(sub)
-            for sub in vars(cfg).values()
-            if isinstance(sub, _PTC)
-        )
-
-    if _has_flash_attention(config):
+    if _config_uses_flash_attention_2(config):
         try:
             from transformers.utils import is_flash_attn_2_available
             _fa2 = is_flash_attn_2_available()
@@ -738,7 +773,7 @@ def load_model(
                 "config.json pins flash_attention_2 but flash-attn is not "
                 "installed — falling back to eager attention."
             )
-            _force_eager_cfg(config)
+            _force_eager_config(config)
 
     model_cls = _select_model_class(task, config)
     load_kwargs: dict = {
@@ -861,12 +896,14 @@ def load_model(
                 f"pip install git+https://github.com/huggingface/transformers.git"
             ) from e
         raise
-    except Exception:
-        # Any other load failure — don't leak the temp dequantized checkpoint.
+    finally:
+        # from_pretrained fully materializes weights before returning. Clean up
+        # on success and across every typed/retry failure path above.
         if _dequant_tmp is not None:
             import shutil as _shutil
+
             _shutil.rmtree(_dequant_tmp, ignore_errors=True)
-        raise
+            logger.info("Removed temporary dequantized checkpoint %s", _dequant_tmp)
 
     if load_policy.move_to_resolved_device:
         # Explicit devices and auto-selected MPS/CPU load on CPU before moving.
@@ -937,12 +974,7 @@ def load_model(
     dev.empty_cache()
 
     if _dequant_tmp is not None:
-        # Weights are fully materialized by from_pretrained at this point —
-        # the temp dequantized copy is no longer needed.
-        import shutil as _shutil
         model._obliteratus_dequantized_scheme = quant_detection.scheme.value
-        _shutil.rmtree(_dequant_tmp, ignore_errors=True)
-        logger.info("Removed temporary dequantized checkpoint %s", _dequant_tmp)
 
     try:
         tokenizer = AutoTokenizer.from_pretrained(
diff --git a/obliteratus/models/quant_dequant.py b/obliteratus/models/quant_dequant.py
index a82bb0e..9e6a95f 100644
--- a/obliteratus/models/quant_dequant.py
+++ b/obliteratus/models/quant_dequant.py
@@ -29,8 +29,11 @@ NVFP4 (ModelOpt, ``quant_algo: "NVFP4"``):
     ``<name>.weight_scale_2`` float32 scalar — global scale (amax/2688)
     dequant: e2m1_values * weight_scale * weight_scale_2
 
-NVFP4 (compressed-tensors): same layout but scales are stored as
-    reciprocals and the global scale may be named ``weight_global_scale``.
+NVFP4 (compressed-tensors, ``format: "nvfp4-pack-quantized"``):
+    ``<name>.weight_packed``       uint8 packed E2M1 values
+    ``<name>.weight_scale``        FP8 block scales (multiplicative)
+    ``<name>.weight_global_scale`` FP32 reciprocal global scale
+    dequant: e2m1_values * weight_scale / weight_global_scale.
 """
 
 from __future__ import annotations
@@ -97,12 +100,17 @@ class QuantDetection:
     reason: str = ""
     block_size: Tuple[int, int] = FP8_DEFAULT_BLOCK
     group_size: int = NVFP4_GROUP_SIZE
-    scale_is_inverse: bool = False  # compressed-tensors stores reciprocals
+    scale_is_inverse: bool = False
+    global_scale_is_inverse: bool = False
     raw_quant_config: dict = field(default_factory=dict)
 
 
 def _load_json_from_checkpoint(
-    model_name_or_path: str, filename: str, token: Optional[str] = None,
+    model_name_or_path: str,
+    filename: str,
+    token: Optional[str] = None,
+    revision: Optional[str] = None,
+    local_files_only: bool = False,
 ) -> Optional[dict]:
     """Read a JSON metadata file from a local dir or the HF hub."""
     if os.path.isdir(model_name_or_path):
@@ -117,7 +125,13 @@ def _load_json_from_checkpoint(
     except Exception:  # pragma: no cover - huggingface_hub always ships w/ transformers
         return None
     try:
-        path = hf_hub_download(model_name_or_path, filename, token=token)
+        path = hf_hub_download(
+            model_name_or_path,
+            filename,
+            token=token,
+            revision=revision,
+            local_files_only=local_files_only,
+        )
     except EntryNotFoundError:
         return None
     except Exception as exc:
@@ -128,12 +142,20 @@ def _load_json_from_checkpoint(
 
 
 def _safetensors_key_names(
-    model_name_or_path: str, config_json: Optional[dict], token: Optional[str] = None,
+    model_name_or_path: str,
+    config_json: Optional[dict],
+    token: Optional[str] = None,
+    revision: Optional[str] = None,
+    local_files_only: bool = False,
 ) -> set:
     """Collect tensor key names from the safetensors index (no weight loads)."""
     keys: set = set()
     index = _load_json_from_checkpoint(
-        model_name_or_path, "model.safetensors.index.json", token=token,
+        model_name_or_path,
+        "model.safetensors.index.json",
+        token=token,
+        revision=revision,
+        local_files_only=local_files_only,
     )
     if index and "weight_map" in index:
         keys.update(index["weight_map"].keys())
@@ -144,7 +166,13 @@ def _safetensors_key_names(
     else:
         try:
             from huggingface_hub import hf_hub_download
-            st_path = hf_hub_download(model_name_or_path, "model.safetensors", token=token)
+            st_path = hf_hub_download(
+                model_name_or_path,
+                "model.safetensors",
+                token=token,
+                revision=revision,
+                local_files_only=local_files_only,
+            )
         except Exception:
             return keys
     if not os.path.exists(st_path):
@@ -158,7 +186,12 @@ def _safetensors_key_names(
     return keys
 
 
-def detect_quant_scheme(model_name_or_path: str, token: Optional[str] = None) -> QuantDetection:
+def detect_quant_scheme(
+    model_name_or_path: str,
+    token: Optional[str] = None,
+    revision: Optional[str] = None,
+    local_files_only: bool = False,
+) -> QuantDetection:
     """Classify a checkpoint's quantization without loading any weights.
 
     Peeks at ``config.json``'s ``quantization_config`` plus safetensors
@@ -166,7 +199,13 @@ def detect_quant_scheme(model_name_or_path: str, token: Optional[str] = None) ->
     reported as UNSUPPORTED with a human-readable reason — the loader
     turns that into a loud error rather than silent corruption.
     """
-    config_json = _load_json_from_checkpoint(model_name_or_path, "config.json", token=token) or {}
+    config_json = _load_json_from_checkpoint(
+        model_name_or_path,
+        "config.json",
+        token=token,
+        revision=revision,
+        local_files_only=local_files_only,
+    ) or {}
     qcfg = config_json.get("quantization_config")
     if not qcfg:
         return QuantDetection(QuantScheme.NONE)
@@ -182,7 +221,13 @@ def detect_quant_scheme(model_name_or_path: str, token: Optional[str] = None) ->
                 block_size=(int(block[0]), int(block[1])),
                 raw_quant_config=raw,
             )
-        keys = _safetensors_key_names(model_name_or_path, config_json, token=token)
+        keys = _safetensors_key_names(
+            model_name_or_path,
+            config_json,
+            token=token,
+            revision=revision,
+            local_files_only=local_files_only,
+        )
         if any(k.endswith("weight_scale_inv") for k in keys):
             return QuantDetection(QuantScheme.FP8_BLOCKWISE, raw_quant_config=raw)
         return QuantDetection(
@@ -235,25 +280,30 @@ def detect_quant_scheme(model_name_or_path: str, token: Optional[str] = None) ->
         wtype = str(wcfg.get("type", "")).lower()
         if num_bits == 8 and wtype == "float":
             return QuantDetection(QuantScheme.FP8_PER_CHANNEL_CT, raw_quant_config=raw)
-        if num_bits == 4 and wtype == "float":
+        if (
+            num_bits == 4
+            and wtype == "float"
+            and qcfg.get("format") == "nvfp4-pack-quantized"
+        ):
             gs = int(wcfg.get("group_size") or NVFP4_GROUP_SIZE)
             return QuantDetection(
                 QuantScheme.NVFP4_CT,
                 group_size=gs,
-                scale_is_inverse=True,
+                global_scale_is_inverse=True,
                 raw_quant_config=raw,
             )
         return QuantDetection(
             QuantScheme.UNSUPPORTED,
             reason=(
                 f"compressed-tensors weights num_bits={num_bits} type={wtype!r} "
-                f"not supported (need float 8-bit or NVFP4 4-bit)"
+                f"format={qcfg.get('format')!r} not supported (need float 8-bit "
+                f"or NVFP4 4-bit with format 'nvfp4-pack-quantized')"
             ),
             raw_quant_config=raw,
         )
 
     if quant_method in ("gptq", "awq", "bitsandbytes", "bitsandbytes_4bit",
-                        "bitsandbytes_8bit", "mxfp4", ""):
+                        "bitsandbytes_8bit", "mxfp4"):
         # Handled elsewhere in the loader / surgery layer.
         return QuantDetection(QuantScheme.NONE, raw_quant_config=raw)
 
@@ -416,6 +466,7 @@ def dequantize_nvfp4(
     global_scale: Optional[torch.Tensor],
     out_shape: Optional[Tuple[int, ...]] = None,
     scale_is_inverse: bool = False,
+    global_scale_is_inverse: bool = False,
     group_size: int = NVFP4_GROUP_SIZE,
     force_manual: bool = False,
 ) -> torch.Tensor:
@@ -424,9 +475,11 @@ def dequantize_nvfp4(
     ``packed_uint8``: (M, N/2) uint8, two E2M1 values per byte, low nibble
     first along the input dim.  ``block_scale``: one FP8-E4M3 scale per
     ``group_size`` elements.  ``global_scale``: FP32 scalar (ModelOpt
-    ``weight_scale_2`` = amax/2688); None means 1.0.  With
-    ``scale_is_inverse`` (compressed-tensors) the scales are reciprocals
-    and are divided out instead of multiplied in.
+    ``weight_scale_2`` = amax/2688); None means 1.0.  ModelOpt stores both
+    scales multiplicatively.  Compressed-tensors keeps the block scale but
+    stores the global scale as its reciprocal, selected with
+    ``global_scale_is_inverse``.  ``scale_is_inverse`` is retained for
+    explicitly reciprocal block-scale layouts.
     """
     vals = unpack_e2m1(packed_uint8, force_manual=force_manual)
     *lead, N = vals.shape
@@ -444,7 +497,7 @@ def dequantize_nvfp4(
     vals = vals.reshape(*lead, N)
     if global_scale is not None:
         gs = global_scale.to(torch.float32)
-        if scale_is_inverse:
+        if global_scale_is_inverse:
             vals.div_(gs)
         else:
             vals.mul_(gs)
@@ -482,35 +535,51 @@ def _strip_known_suffix(key: str) -> Optional[str]:
     return None
 
 
-def find_scale_siblings(state_dict: Dict[str, torch.Tensor], weight_key: str) -> dict:
-    """Locate the scale tensors belonging to ``weight_key`` (``foo.weight``)."""
-    base = weight_key[: -len(".weight")] if weight_key.endswith(".weight") else weight_key
-    dotted = weight_key.endswith(".weight")
-    out = {}
-    for name, attr in [
-        (".weight_scale_inv", "scale_inv"),
-        (".weight_scale_2", "global_scale"),
-        (".weight_global_scale", "global_scale"),
-        (".weight_scale", "block_scale"),
-    ]:
-        k = base + name
-        if k in state_dict:
-            out[attr] = state_dict[k]
-    if not dotted:
-        # Fused-MoE naming: experts.gate_up_proj{,_scale,_scale_2,_scale_inv}
-        for name, attr in [
+def _scale_sibling_names(weight_key: str) -> tuple[tuple[str, str], ...]:
+    """Return candidate ``(key, role)`` pairs for one quantized weight."""
+    if weight_key.endswith(".weight_packed"):
+        base = weight_key[: -len(".weight_packed")]
+        dotted = True
+    elif weight_key.endswith(".weight"):
+        base = weight_key[: -len(".weight")]
+        dotted = True
+    else:
+        base = weight_key
+        dotted = False
+    suffixes = (
+        (
+            (".weight_scale_inv", "scale_inv"),
+            (".weight_scale_2", "global_scale"),
+            (".weight_global_scale", "global_scale"),
+            (".weight_scale", "block_scale"),
+        )
+        if dotted
+        else (
+            # Fused-MoE naming:
+            # experts.gate_up_proj{,_scale,_scale_2,_scale_inv}
             ("_scale_inv", "scale_inv"),
             ("_scale_2", "global_scale"),
             ("_global_scale", "global_scale"),
             ("_scale", "block_scale"),
-        ]:
-            k = base + name
-            if k in state_dict and attr not in out:
-                out[attr] = state_dict[k]
+        )
+    )
+    return tuple((base + suffix, role) for suffix, role in suffixes)
+
+
+def find_scale_siblings(state_dict: Dict[str, torch.Tensor], weight_key: str) -> dict:
+    """Locate the scale tensors belonging to ``weight_key`` (``foo.weight``)."""
+    out = {}
+    for key, role in _scale_sibling_names(weight_key):
+        if key in state_dict and role not in out:
+            out[role] = state_dict[key]
     return out
 
 
-def is_scale_key(key: str, state_dict: Optional[Dict[str, torch.Tensor]] = None) -> bool:
+def is_scale_key(
+    key: str,
+    state_dict: Optional[Dict[str, torch.Tensor]] = None,
+    known_keys: Optional[set[str]] = None,
+) -> bool:
     """True for keys holding quantization scales, not real parameters.
 
     With ``state_dict`` given, the key only counts as a scale if its base
@@ -521,9 +590,15 @@ def is_scale_key(key: str, state_dict: Optional[Dict[str, torch.Tensor]] = None)
     base = _strip_known_suffix(key)
     if base is None:
         return False
-    if state_dict is not None:
-        # Dotted scales pair with ``base.weight``; fused-MoE scales with ``base``.
-        if base not in state_dict and (base + ".weight") not in state_dict:
+    keys = known_keys if known_keys is not None else state_dict
+    if keys is not None:
+        # Dotted scales pair with dense or packed weights; fused-MoE scales
+        # pair directly with ``base``.
+        if (
+            base not in keys
+            and (base + ".weight") not in keys
+            and (base + ".weight_packed") not in keys
+        ):
             return False
     return True
 
@@ -532,6 +607,8 @@ def dequantize_state_dict(
     state_dict: Dict[str, torch.Tensor],
     detection: QuantDetection,
     out_dtype: torch.dtype = torch.bfloat16,
+    sibling_tensors: Optional[Dict[str, torch.Tensor]] = None,
+    known_keys: Optional[set[str]] = None,
 ) -> Dict[str, torch.Tensor]:
     """Dequantize every quantized weight group in one checkpoint shard.
 
@@ -540,13 +617,16 @@ def dequantize_state_dict(
     over naming conventions (dense linears and fused MoE experts alike)
     because grouping is by key suffix, not by module type.
     """
+    scale_lookup = state_dict
+    if sibling_tensors:
+        scale_lookup = {**state_dict, **sibling_tensors}
     out: Dict[str, torch.Tensor] = {}
     for key, tensor in state_dict.items():
-        if is_scale_key(key, state_dict):
+        if is_scale_key(key, state_dict, known_keys):
             # Scales are consumed together with their weight (or dropped if
             # the weight is absent — e.g. quantization status tensors).
             continue
-        sib = find_scale_siblings(state_dict, key)
+        sib = find_scale_siblings(scale_lookup, key)
         if is_fp8_dtype(tensor.dtype):
             if "scale_inv" in sib:
                 w = dequantize_fp8_blockwise(tensor, sib["scale_inv"], detection.block_size)
@@ -567,6 +647,7 @@ def dequantize_state_dict(
                 w = dequantize_nvfp4(
                     tensor, scale, gs,
                     scale_is_inverse=detection.scale_is_inverse,
+                    global_scale_is_inverse=detection.global_scale_is_inverse,
                     group_size=detection.group_size,
                 )
             else:
@@ -574,8 +655,9 @@ def dequantize_state_dict(
                     f"packed uint8 tensor {key!r} in a {detection.scheme.value} "
                     f"checkpoint — unsupported layout; report a bug"
                 )
-            out[key] = w.to(out_dtype)
-        elif tensor.dtype == torch.uint8 and key.endswith(".weight"):
+            output_key = key[: -len("_packed")] if key.endswith(".weight_packed") else key
+            out[output_key] = w.to(out_dtype)
+        elif tensor.dtype == torch.uint8 and key.endswith((".weight", ".weight_packed")):
             raise RuntimeError(
                 f"uint8 weight {key!r} has no recognizable scale siblings "
                 f"(keys present: {[k for k in state_dict if k.startswith(key[:-7])]}) "
@@ -596,6 +678,8 @@ def materialize_dequantized_checkpoint(
     detection: QuantDetection,
     out_dtype: torch.dtype = torch.bfloat16,
     token: Optional[str] = None,
+    revision: Optional[str] = None,
+    local_files_only: bool = False,
 ) -> Tuple[str, str]:
     """Write a dequantized float copy of a checkpoint to a temp dir.
 
@@ -608,6 +692,7 @@ def materialize_dequantized_checkpoint(
     import shutil
     import tempfile
 
+    from safetensors import safe_open
     from safetensors.torch import load_file, save_file
 
     if os.path.isdir(model_name_or_path):
@@ -618,6 +703,8 @@ def materialize_dequantized_checkpoint(
         src = snapshot_download(
             model_name_or_path,
             token=token,
+            revision=revision,
+            local_files_only=local_files_only,
             allow_patterns=[
                 "*.json", "*.safetensors", "*.py", "tokenizer*", "*.model",
                 "*.txt", "chat_template*", "special_tokens_map.json",
@@ -631,52 +718,86 @@ def materialize_dequantized_checkpoint(
         detection.scheme.value, model_name_or_path, out_dtype, tmp,
     )
 
-    # Copy non-weight files; strip quantization_config from config.json.
-    for name in os.listdir(src):
-        s = os.path.join(src, name)
-        if not os.path.isfile(s) or name.endswith(".safetensors"):
-            continue
-        d = os.path.join(tmp, name)
-        if name == "config.json":
-            with open(s, "r", encoding="utf-8") as fh:
-                cfg = json.load(fh)
-            cfg.pop("quantization_config", None)
-            with open(d, "w", encoding="utf-8") as fh:
-                json.dump(cfg, fh, indent=2)
+    try:
+        # Copy non-weight files; strip quantization_config from config.json.
+        for name in os.listdir(src):
+            s = os.path.join(src, name)
+            if not os.path.isfile(s) or name.endswith(".safetensors"):
+                continue
+            d = os.path.join(tmp, name)
+            if name == "config.json":
+                with open(s, "r", encoding="utf-8") as fh:
+                    cfg = json.load(fh)
+                cfg.pop("quantization_config", None)
+                with open(d, "w", encoding="utf-8") as fh:
+                    json.dump(cfg, fh, indent=2)
+            else:
+                shutil.copy2(s, d)
+
+        # Locate shards.
+        index_path = os.path.join(src, "model.safetensors.index.json")
+        old_index = None
+        if os.path.exists(index_path):
+            with open(index_path, "r", encoding="utf-8") as fh:
+                old_index = json.load(fh)
+            shards = sorted(set(old_index["weight_map"].values()))
+        elif os.path.exists(os.path.join(src, "model.safetensors")):
+            shards = ["model.safetensors"]
         else:
-            shutil.copy2(s, d)
+            raise RuntimeError(
+                f"Quantized checkpoint '{model_name_or_path}' has no safetensors "
+                f"weights (pytorch_model.bin quantized checkpoints are not "
+                f"supported) — use a BF16 checkpoint"
+            )
 
-    # Locate shards.
-    index_path = os.path.join(src, "model.safetensors.index.json")
-    old_index = None
-    if os.path.exists(index_path):
-        with open(index_path, "r", encoding="utf-8") as fh:
-            old_index = json.load(fh)
-        shards = sorted(set(old_index["weight_map"].values()))
-    elif os.path.exists(os.path.join(src, "model.safetensors")):
-        shards = ["model.safetensors"]
-    else:
-        raise RuntimeError(
-            f"Quantized checkpoint '{model_name_or_path}' has no safetensors "
-            f"weights (pytorch_model.bin quantized checkpoints are not "
-            f"supported) — use a BF16 checkpoint"
-        )
+        weight_map = dict(old_index.get("weight_map") or {}) if old_index else {}
+        known_keys = set(weight_map) if weight_map else None
+        new_weight_map: Dict[str, str] = {}
+        new_total_size = 0
+        for i, shard in enumerate(shards, 1):
+            sd = load_file(os.path.join(src, shard), device="cpu")
+            external_scales: Dict[str, torch.Tensor] = {}
+            external_by_shard: Dict[str, list[str]] = {}
+            for weight_key in sd:
+                for scale_key, _role in _scale_sibling_names(weight_key):
+                    scale_shard = weight_map.get(scale_key)
+                    if scale_shard and scale_shard != shard:
+                        external_by_shard.setdefault(scale_shard, []).append(scale_key)
+            for scale_shard, scale_keys in external_by_shard.items():
+                with safe_open(
+                    os.path.join(src, scale_shard), framework="pt", device="cpu",
+                ) as handle:
+                    for scale_key in scale_keys:
+                        external_scales[scale_key] = handle.get_tensor(scale_key)
 
-    new_weight_map: Dict[str, str] = {}
-    for i, shard in enumerate(shards, 1):
-        sd = load_file(os.path.join(src, shard), device="cpu")
-        out_sd = dequantize_state_dict(sd, detection, out_dtype=out_dtype)
-        save_file(out_sd, os.path.join(tmp, shard), metadata={"format": "pt"})
-        for k in out_sd:
-            new_weight_map[k] = shard
-        logger.info(
-            "dequantized shard %d/%d (%s): %d tensors -> %d float tensors",
-            i, len(shards), shard, len(sd), len(out_sd),
-        )
+            out_sd = dequantize_state_dict(
+                sd,
+                detection,
+                out_dtype=out_dtype,
+                sibling_tensors=external_scales,
+                known_keys=known_keys,
+            )
+            if out_sd:
+                save_file(out_sd, os.path.join(tmp, shard), metadata={"format": "pt"})
+            for k, tensor in out_sd.items():
+                new_weight_map[k] = shard
+                new_total_size += tensor.numel() * tensor.element_size()
+            logger.info(
+                "dequantized shard %d/%d (%s): %d tensors -> %d float tensors",
+                i, len(shards), shard, len(sd), len(out_sd),
+            )
 
-    if old_index is not None:
-        metadata = dict(old_index.get("metadata") or {})
-        with open(os.path.join(tmp, "model.safetensors.index.json"), "w", encoding="utf-8") as fh:
-            json.dump({"metadata": metadata, "weight_map": new_weight_map}, fh, indent=2)
+        if old_index is not None:
+            metadata = dict(old_index.get("metadata") or {})
+            metadata["total_size"] = new_total_size
+            with open(
+                os.path.join(tmp, "model.safetensors.index.json"),
+                "w",
+                encoding="utf-8",
+            ) as fh:
+                json.dump({"metadata": metadata, "weight_map": new_weight_map}, fh, indent=2)
+    except BaseException:
+        shutil.rmtree(tmp, ignore_errors=True)
+        raise
 
     return tmp, src
diff --git a/scripts/smoke_load_quant.py b/scripts/smoke_load_quant.py
index 69a0c2b..f17c80e 100644
--- a/scripts/smoke_load_quant.py
+++ b/scripts/smoke_load_quant.py
@@ -1,55 +1,100 @@
-"""Smoke test: load a quantized checkpoint through the OBLITERATUS loader.
+"""Smoke-load a quantized checkpoint through the OBLITERATUS loader."""
 
-Verifies: detection, dequantization to BF16, no quantized tensors left in
-the model, config stripped of quantization_config, and a short generation.
-"""
+from __future__ import annotations
 
-import sys
+import argparse
 
 import torch
 
 from obliteratus.models import quant_dequant as qd
 from obliteratus.models.loader import load_model
 
-repo = sys.argv[1]
 
-det = qd.detect_quant_scheme(repo)
-print(f"[detection] {det.scheme.value} (inverse={det.scale_is_inverse}, group={det.group_size})")
-assert det.scheme not in (qd.QuantScheme.NONE, qd.QuantScheme.UNSUPPORTED), det
+def build_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("repo", help="Local checkpoint path or Hugging Face repository")
+    parser.add_argument(
+        "--revision",
+        help="Immutable Hub commit, tag, or branch to load",
+    )
+    parser.add_argument(
+        "--local-files-only",
+        action="store_true",
+        help="Refuse network access and use only locally cached files",
+    )
+    parser.add_argument(
+        "--trust-remote-code",
+        action="store_true",
+        help=(
+            "Explicitly allow checkpoint-provided Python code to execute. "
+            "Off by default; review and pin the repository revision first."
+        ),
+    )
+    return parser
 
-handle = load_model(
-    repo,
-    task="causal_lm",
-    device="auto",
-    dtype="bfloat16",
-    trust_remote_code=True,
-)
-model = handle.model
-print(f"[load] ok — scheme tag: {getattr(model, '_obliteratus_dequantized_scheme', None)}")
 
-bad = []
-n_params = 0
-for name, p in model.named_parameters():
-    n_params += 1
-    if p.dtype in qd.FP8_DTYPES or p.dtype == torch.uint8:
-        bad.append((name, str(p.dtype)))
-    if torch.isnan(p.data).any().item() if p.data.is_floating_point() else False:
-        bad.append((name, "NaN"))
-print(f"[check] {n_params} params, quantized/NaN leftovers: {bad[:10] or 'NONE'}")
-assert not bad, bad
+def main(argv: list[str] | None = None) -> int:
+    args = build_parser().parse_args(argv)
+    det = qd.detect_quant_scheme(
+        args.repo,
+        revision=args.revision,
+        local_files_only=args.local_files_only,
+    )
+    print(
+        f"[detection] {det.scheme.value} "
+        f"(block_inverse={det.scale_is_inverse}, "
+        f"global_inverse={det.global_scale_is_inverse}, group={det.group_size})"
+    )
+    if det.scheme in (qd.QuantScheme.NONE, qd.QuantScheme.UNSUPPORTED):
+        raise RuntimeError(det.reason or f"checkpoint is not FP8/NVFP4: {args.repo}")
 
-assert getattr(model.config, "quantization_config", None) is None, "quantization_config survived"
+    handle = load_model(
+        args.repo,
+        task="causal_lm",
+        device="auto",
+        dtype="bfloat16",
+        trust_remote_code=args.trust_remote_code,
+        revision=args.revision,
+        local_files_only=args.local_files_only,
+    )
+    model = handle.model
+    print(
+        "[load] ok — scheme tag: "
+        f"{getattr(model, '_obliteratus_dequantized_scheme', None)}"
+    )
 
-tok = handle.tokenizer
-prompt = "The capital of France is"
-try:
-    inputs = tok(prompt, return_tensors="pt").to(model.device)
-except Exception:
-    # Omni processors: fall back to bare tokenizer encode
-    ids = tok.encode(prompt, return_tensors="pt").to(model.device)
-    inputs = {"input_ids": ids}
-with torch.no_grad():
-    out = model.generate(**inputs, max_new_tokens=16, do_sample=False)
-text = tok.decode(out[0][-16:] if out.dim() > 1 else out[-16:], skip_special_tokens=True)
-print(f"[generate] {text!r}")
-print("SMOKE_LOAD_OK")
+    bad = []
+    n_params = 0
+    for name, param in model.named_parameters():
+        n_params += 1
+        if param.dtype in qd.FP8_DTYPES or param.dtype == torch.uint8:
+            bad.append((name, str(param.dtype)))
+        if param.data.is_floating_point() and torch.isnan(param.data).any().item():
+            bad.append((name, "NaN"))
+    print(f"[check] {n_params} params, quantized/NaN leftovers: {bad[:10] or 'NONE'}")
+    if bad:
+        raise RuntimeError(f"quantized or non-finite tensors remain: {bad[:10]}")
+
+    if getattr(model.config, "quantization_config", None) is not None:
+        raise RuntimeError("quantization_config survived dequantization")
+
+    tokenizer = handle.tokenizer
+    prompt = "The capital of France is"
+    try:
+        inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
+    except Exception:
+        ids = tokenizer.encode(prompt, return_tensors="pt").to(model.device)
+        inputs = {"input_ids": ids}
+    with torch.no_grad():
+        output = model.generate(**inputs, max_new_tokens=16, do_sample=False)
+    text = tokenizer.decode(
+        output[0][-16:] if output.dim() > 1 else output[-16:],
+        skip_special_tokens=True,
+    )
+    print(f"[generate] {text!r}")
+    print("SMOKE_LOAD_OK")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/tests/fixtures/quant_dequant_provenance.json b/tests/fixtures/quant_dequant_provenance.json
new file mode 100644
index 0000000..7e41a45
--- /dev/null
+++ b/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."
+}
diff --git a/tests/test_cli_boundaries.py b/tests/test_cli_boundaries.py
index 310f513..cc6285a 100644
--- a/tests/test_cli_boundaries.py
+++ b/tests/test_cli_boundaries.py
@@ -80,6 +80,18 @@ def test_refusal_max_tokens_cli_default_and_positive_override(monkeypatch):
     assert command.call_args.args[0].refusal_max_tokens == 512
 
 
+def test_trust_remote_code_requires_explicit_cli_opt_in(monkeypatch):
+    command = Mock()
+    monkeypatch.setattr(cli, "_cmd_abliterate", command)
+
+    cli.main(["abliterate", "local/model"])
+    assert command.call_args.args[0].trust_remote_code is False
+
+    command.reset_mock()
+    cli.main(["abliterate", "local/model", "--trust-remote-code"])
+    assert command.call_args.args[0].trust_remote_code is True
+
+
 @pytest.mark.parametrize("invalid", ["0", "-1", "not-an-integer"])
 def test_refusal_max_tokens_cli_rejects_invalid_values(invalid):
     with pytest.raises(SystemExit) as exc:
@@ -510,6 +522,7 @@ def test_abliterate_pipeline_callbacks_residue_and_contribution(monkeypatch, tmp
     cli._cmd_abliterate(args)
     assert factory.call_args.kwargs["refusal_max_tokens"] == 512
     assert factory.call_args.kwargs["gpu_memory_utilization"] == 0.95
+    assert factory.call_args.kwargs["trust_remote_code"] is False
     assert (result_path / "hard_negative_residue.json").is_file()
     telemetry.assert_called_once_with(pipeline)
 
diff --git a/tests/test_loader_boundaries.py b/tests/test_loader_boundaries.py
index d60acdf..c1f2871 100644
--- a/tests/test_loader_boundaries.py
+++ b/tests/test_loader_boundaries.py
@@ -23,6 +23,7 @@ def test_compatibility_shims_tolerate_independent_missing_or_broken_imports(monk
     original_import = builtins.__import__
     failure_keys = {
         ("transformers", ("AutoModelForImageTextToText",)): ImportError,
+        ("transformers", ("PreTrainedModel",)): RuntimeError,
         ("transformers.utils", ("output_capturing",)): ImportError,
         ("transformers.utils.import_utils", ()): RuntimeError,
         ("transformers.utils", ()): RuntimeError,
@@ -109,6 +110,72 @@ def test_compatibility_shims_tolerate_rejected_generic_attribute_patch(monkeypat
     assert _execute_loader_source()["TASK_MODEL_MAP"]
 
 
+def test_all_tied_weights_compat_aggregates_filters_and_caches():
+    class LegacyModel(torch.nn.Module):
+        pass
+
+    assert loader._install_all_tied_weights_keys_compat(LegacyModel) is True
+    assert loader._install_all_tied_weights_keys_compat(LegacyModel) is False
+
+    model = LegacyModel()
+    model.lm_head = torch.nn.Linear(2, 2)
+    model._tied_weights_keys = ["lm_head.weight", "missing.weight"]
+    child = torch.nn.Module()
+    child._tied_weights_keys = {"lm_head.bias": "lm_head.weight"}
+    model.child = child
+
+    assert model.all_tied_weights_keys == {
+        "lm_head.weight": "lm_head.weight",
+        "lm_head.bias": "lm_head.weight",
+    }
+    model.all_tied_weights_keys = {"cached.weight": "source.weight"}
+    assert model.all_tied_weights_keys == {"cached.weight": "source.weight"}
+
+
+def test_all_tied_weights_compat_prefers_transformers_expansion_contract():
+    class ExpandableModel(torch.nn.Module):
+        def get_expanded_tied_weights_keys(self, *, all_submodels):
+            assert all_submodels is True
+            return {"lm_head.weight": "model.embed_tokens.weight"}
+
+    assert loader._install_all_tied_weights_keys_compat(ExpandableModel) is True
+    model = ExpandableModel()
+
+    assert model.all_tied_weights_keys == {
+        "lm_head.weight": "model.embed_tokens.weight",
+    }
+    assert model.__dict__["_all_tied_weights_keys_store"] == {
+        "lm_head.weight": "model.embed_tokens.weight",
+    }
+
+
+def test_all_tied_weights_compat_falls_back_when_expansion_is_legacy():
+    class LegacyExpandableModel(torch.nn.Module):
+        def get_expanded_tied_weights_keys(self, *, all_submodels):
+            raise TypeError("legacy list contract")
+
+    assert loader._install_all_tied_weights_keys_compat(LegacyExpandableModel) is True
+    model = LegacyExpandableModel()
+    model.lm_head = torch.nn.Linear(2, 2)
+    model._tied_weights_keys = ["lm_head.weight"]
+
+    assert model.all_tied_weights_keys == {"lm_head.weight": "lm_head.weight"}
+
+
+def test_flash_attention_config_helpers_recurse_into_transformers_subconfigs():
+    from transformers import PretrainedConfig
+
+    outer = PretrainedConfig()
+    inner = PretrainedConfig()
+    outer.inner = inner
+    inner._attn_implementation = "flash_attention_2"
+
+    assert loader._config_uses_flash_attention_2(outer) is True
+    loader._force_eager_config(outer)
+    assert inner._attn_implementation == "eager"
+    assert loader._config_uses_flash_attention_2(outer) is False
+
+
 def _config(**overrides):
     values = {
         "model_type": "gpt2",
@@ -200,6 +267,158 @@ def test_revision_trust_and_offline_flags_reach_every_provider(loader_boundary):
     loader.dev.empty_cache.assert_called_once_with()
 
 
+def test_unavailable_flash_attention_falls_back_to_eager(
+    loader_boundary, monkeypatch,
+):
+    import transformers.utils
+
+    loader_boundary.config._attn_implementation = "flash_attention_2"
+    monkeypatch.setattr(
+        transformers.utils,
+        "is_flash_attn_2_available",
+        lambda: False,
+    )
+    monkeypatch.setattr(
+        loader.qd,
+        "detect_quant_scheme",
+        lambda *args, **kwargs: loader.qd.QuantDetection(loader.qd.QuantScheme.NONE),
+    )
+
+    loader.load_model("local/model", skip_snapshot=True)
+
+    assert loader_boundary.config._attn_implementation == "eager"
+
+
+def test_broken_flash_attention_probe_falls_back_to_eager(
+    loader_boundary, monkeypatch,
+):
+    loader_boundary.config._attn_implementation = "flash_attention_2"
+    real_import = builtins.__import__
+
+    def rejecting_import(name, globals=None, locals=None, fromlist=(), level=0):
+        if (
+            name == "transformers.utils"
+            and "is_flash_attn_2_available" in tuple(fromlist or ())
+        ):
+            raise RuntimeError("flash-attention availability probe failed")
+        return real_import(name, globals, locals, fromlist, level)
+
+    monkeypatch.setattr(builtins, "__import__", rejecting_import)
+    loader.load_model("local/model", skip_snapshot=True)
+
+    assert loader_boundary.config._attn_implementation == "eager"
+
+
+def test_available_flash_attention_preserves_requested_implementation(
+    loader_boundary, monkeypatch,
+):
+    import transformers.utils
+
+    loader_boundary.config._attn_implementation = "flash_attention_2"
+    monkeypatch.setattr(
+        transformers.utils,
+        "is_flash_attn_2_available",
+        lambda: True,
+    )
+
+    loader.load_model("local/model", skip_snapshot=True)
+
+    assert loader_boundary.config._attn_implementation == "flash_attention_2"
+
+
+def test_multimodal_unwrap_generation_cache_and_legacy_ties(
+    loader_boundary, monkeypatch,
+):
+    class FakeLanguageModel(torch.nn.Module):
+        def __init__(self):
+            super().__init__()
+            self.config = _config(tie_word_embeddings=False)
+            self._tied_weights_keys = ["lm_head.weight"]
+            self.last_cache_position = None
+
+        def prepare_inputs_for_generation(
+            self, input_ids, *args, cache_position=None, **kwargs,
+        ):
+            self.last_cache_position = cache_position
+            return {"input_ids": input_ids, "cache_position": cache_position}
+
+    class MultimodalWrapper(torch.nn.Module):
+        def __init__(self):
+            super().__init__()
+            self.language_model = FakeLanguageModel()
+
+        def forward(self, input_ids, pixel_values):
+            return input_ids, pixel_values
+
+    wrapper = MultimodalWrapper()
+    loader_boundary.model_class.from_pretrained.return_value = wrapper
+    monkeypatch.setattr(loader, "PreTrainedModel", FakeLanguageModel)
+    monkeypatch.setattr(
+        loader.qd,
+        "detect_quant_scheme",
+        lambda *args, **kwargs: loader.qd.QuantDetection(loader.qd.QuantScheme.NONE),
+    )
+
+    handle = loader.load_model("local/model", skip_snapshot=True)
+
+    assert handle.model is wrapper.language_model
+    assert handle.model._tied_weights_keys == {}
+    input_ids = torch.ones(1, 2, dtype=torch.long)
+
+    class Cache:
+        @staticmethod
+        def get_seq_length():
+            return 4
+
+    prepared = handle.model.prepare_inputs_for_generation(
+        input_ids, past_key_values=Cache(),
+    )
+    assert prepared["cache_position"].tolist() == [4, 5]
+
+    legacy_cache = ((torch.zeros(1, 1, 3, 1),),)
+    prepared = handle.model.prepare_inputs_for_generation(
+        input_ids, past_key_values=legacy_cache,
+    )
+    assert prepared["cache_position"].tolist() == [3, 4]
+
+    prepared = handle.model.prepare_inputs_for_generation(
+        input_ids, past_key_values=object(),
+    )
+    assert prepared["cache_position"].tolist() == [0, 1]
+
+    prepared = handle.model.prepare_inputs_for_generation(input_ids)
+    assert prepared["cache_position"].tolist() == [0, 1]
+
+    explicit = torch.tensor([7, 8])
+    prepared = handle.model.prepare_inputs_for_generation(
+        input_ids, cache_position=explicit,
+    )
+    assert prepared["cache_position"] is explicit
+
+
+def test_multimodal_signature_probe_failure_preserves_wrapper(
+    loader_boundary, monkeypatch,
+):
+    class FakeLanguageModel(torch.nn.Module):
+        def __init__(self):
+            super().__init__()
+            self.config = _config()
+
+    class WrapperWithUninspectableForward(torch.nn.Module):
+        def __init__(self):
+            super().__init__()
+            self.language_model = FakeLanguageModel()
+            self.forward = object()
+
+    wrapper = WrapperWithUninspectableForward()
+    loader_boundary.model_class.from_pretrained.return_value = wrapper
+    monkeypatch.setattr(loader, "PreTrainedModel", FakeLanguageModel)
+
+    handle = loader.load_model("local/model", skip_snapshot=True)
+
+    assert handle.model is wrapper
+
+
 def test_hf_token_is_forwarded_without_logging_value(loader_boundary, monkeypatch):
     monkeypatch.setenv("HF_TOKEN", "secret-token")
     loader.load_model("x", skip_snapshot=True)
diff --git a/tests/test_quant_dequant.py b/tests/test_quant_dequant.py
index 401559a..a7792b1 100644
--- a/tests/test_quant_dequant.py
+++ b/tests/test_quant_dequant.py
@@ -2,7 +2,7 @@
 
 Covers:
 - FP8 block-wise (DeepSeek-style) and per-channel round-trips
-- NVFP4 round-trips (direct + reciprocal scale conventions), native and
+- NVFP4 round-trips (ModelOpt + compressed-tensors global-scale conventions), native and
   manual nibble-unpack paths agreeing with each other
 - detect_quant_scheme classification from config.json metadata
 - surgery guards rejecting un-dequantized FP8/NVFP4 weights
@@ -13,6 +13,8 @@ from __future__ import annotations
 
 import json
 import os
+from pathlib import Path
+from types import SimpleNamespace
 
 import pytest
 import torch
@@ -61,12 +63,12 @@ def _nearest_e2m1(x: torch.Tensor) -> torch.Tensor:
     return sign * lut[idx]
 
 
-def _pack_nvfp4(w: torch.Tensor, group: int = 16, reciprocal: bool = False):
+def _pack_nvfp4(w: torch.Tensor, group: int = 16, global_reciprocal: bool = False):
     """Float (M, N) weight → (packed uint8, e4m3 block scales, fp32 global).
 
     Follows the ModelOpt convention: w ≈ e2m1 * block_scale * global_scale.
-    With reciprocal=True the returned scales are 1/scale (compressed-tensors
-    convention).
+    Compressed-tensors keeps the block scale but stores the reciprocal of
+    the global scale.
     """
     M, N = w.shape
     wg = w.reshape(M, N // group, group)
@@ -82,8 +84,8 @@ def _pack_nvfp4(w: torch.Tensor, group: int = 16, reciprocal: bool = False):
     high = codes[:, 1::2]
     packed = (high << 4) | low
     bs_fp8 = bs.to(torch.float8_e4m3fn)
-    if reciprocal:
-        return packed, (1.0 / bs_fp8.float()).to(torch.float8_e4m3fn), 1.0 / global_scale
+    if global_reciprocal:
+        return packed, bs_fp8, 1.0 / global_scale
     return packed, bs_fp8, global_scale.float()
 
 
@@ -139,13 +141,17 @@ def test_fp8_per_tensor_scalar_scale():
 # ---------------------------------------------------------------------------
 
 @requires_fp8
-@pytest.mark.parametrize("reciprocal", [False, True])
-def test_nvfp4_roundtrip(reciprocal):
+@pytest.mark.parametrize("global_reciprocal", [False, True])
+def test_nvfp4_roundtrip(global_reciprocal):
     torch.manual_seed(4)
     w = torch.randn(64, 128)
-    packed, bs, gs = _pack_nvfp4(w, reciprocal=reciprocal)
+    packed, bs, gs = _pack_nvfp4(w, global_reciprocal=global_reciprocal)
     out = qd.dequantize_nvfp4(
-        packed, bs, gs, scale_is_inverse=reciprocal, force_manual=True,
+        packed,
+        bs,
+        gs,
+        global_scale_is_inverse=global_reciprocal,
+        force_manual=True,
     )
     assert out.shape == w.shape
     # NVFP4 is coarse — check correlation and relative error, not equality.
@@ -153,9 +159,8 @@ def test_nvfp4_roundtrip(reciprocal):
         w.flatten(), out.flatten(), dim=0,
     )
     rel = (out - w).norm() / w.norm()
-    # Reciprocal storage double-rounds the inverse scales — slightly coarser.
-    assert cos > (0.98 if reciprocal else 0.995), f"NVFP4 cosine {cos:.4f}"
-    assert rel < (0.20 if reciprocal else 0.15), f"NVFP4 relative error {rel:.4f}"
+    assert cos > 0.995, f"NVFP4 cosine {cos:.4f}"
+    assert rel < 0.15, f"NVFP4 relative error {rel:.4f}"
 
 
 @requires_fp8
@@ -206,6 +211,67 @@ def _write_config(tmp_path, cfg: dict):
         json.dump(cfg, fh)
 
 
+def test_load_json_local_missing_returns_none(tmp_path):
+    assert qd._load_json_from_checkpoint(str(tmp_path), "missing.json") is None
+
+
+def test_safetensors_key_names_reads_single_file_and_index(tmp_path):
+    from safetensors.torch import save_file
+
+    save_file({"a.weight": torch.ones(1)}, tmp_path / "model.safetensors")
+    assert qd._safetensors_key_names(str(tmp_path), None) == {"a.weight"}
+    (tmp_path / "model.safetensors.index.json").write_text(json.dumps({
+        "weight_map": {"b.weight": "model.safetensors"},
+    }))
+    assert qd._safetensors_key_names(str(tmp_path), None) == {"b.weight"}
+
+
+def test_safetensors_key_names_handles_missing_and_invalid_files(tmp_path):
+    assert qd._safetensors_key_names(str(tmp_path), None) == set()
+    (tmp_path / "model.safetensors").write_text("not safetensors")
+    assert qd._safetensors_key_names(str(tmp_path), None) == set()
+
+
+def test_safetensors_key_names_remote_propagates_hub_policy(tmp_path, monkeypatch):
+    from safetensors.torch import save_file
+
+    checkpoint = tmp_path / "remote.safetensors"
+    save_file({"remote.weight": torch.ones(1)}, checkpoint)
+    monkeypatch.setattr(qd, "_load_json_from_checkpoint", lambda *args, **kwargs: None)
+    calls = []
+
+    def fake_download(repo, filename, **kwargs):
+        calls.append((repo, filename, kwargs))
+        return str(checkpoint)
+
+    monkeypatch.setattr("huggingface_hub.hf_hub_download", fake_download)
+    assert qd._safetensors_key_names(
+        "org/model",
+        None,
+        token="token",
+        revision="immutable-sha",
+        local_files_only=True,
+    ) == {"remote.weight"}
+    assert calls == [(
+        "org/model",
+        "model.safetensors",
+        {
+            "token": "token",
+            "revision": "immutable-sha",
+            "local_files_only": True,
+        },
+    )]
+
+
+def test_safetensors_key_names_remote_download_failure_is_empty(monkeypatch):
+    monkeypatch.setattr(qd, "_load_json_from_checkpoint", lambda *args, **kwargs: None)
+    monkeypatch.setattr(
+        "huggingface_hub.hf_hub_download",
+        lambda *args, **kwargs: (_ for _ in ()).throw(OSError("offline")),
+    )
+    assert qd._safetensors_key_names("org/model", None) == set()
+
+
 def test_detect_none(tmp_path):
     _write_config(tmp_path, {"model_type": "gpt2"})
     det = qd.detect_quant_scheme(str(tmp_path))
@@ -232,6 +298,18 @@ def test_detect_fp8_blockwise_via_scale_keys(tmp_path):
     assert det.scheme is qd.QuantScheme.FP8_BLOCKWISE
 
 
+def test_detect_fp8_without_supported_layout_fails_closed(tmp_path):
+    _write_config(tmp_path, {
+        "quantization_config": {
+            "quant_method": "fp8",
+            "activation_scheme": "static",
+        },
+    })
+    det = qd.detect_quant_scheme(str(tmp_path))
+    assert det.scheme is qd.QuantScheme.UNSUPPORTED
+    assert "without weight_block_size" in det.reason
+
+
 def test_detect_fp8_ct(tmp_path):
     _write_config(tmp_path, {
         "quantization_config": {
@@ -247,15 +325,33 @@ def test_detect_nvfp4_ct(tmp_path):
     _write_config(tmp_path, {
         "quantization_config": {
             "quant_method": "compressed-tensors",
+            "format": "nvfp4-pack-quantized",
             "config_groups": {"group_0": {"weights": {"num_bits": 4, "type": "float", "group_size": 16}}},
         },
     })
     det = qd.detect_quant_scheme(str(tmp_path))
     assert det.scheme is qd.QuantScheme.NVFP4_CT
-    assert det.scale_is_inverse is True
+    assert det.scale_is_inverse is False
+    assert det.global_scale_is_inverse is True
     assert det.group_size == 16
 
 
+def test_detect_nvfp4_ct_without_packed_format_fails_closed(tmp_path):
+    _write_config(tmp_path, {
+        "quantization_config": {
+            "quant_method": "compressed-tensors",
+            "config_groups": {
+                "group_0": {
+                    "weights": {"num_bits": 4, "type": "float", "group_size": 16},
+                },
+            },
+        },
+    })
+    det = qd.detect_quant_scheme(str(tmp_path))
+    assert det.scheme is qd.QuantScheme.UNSUPPORTED
+    assert "nvfp4-pack-quantized" in det.reason
+
+
 def test_detect_nvfp4_modelopt(tmp_path):
     _write_config(tmp_path, {
         "quantization_config": {"quant_method": "modelopt", "quant_algo": "NVFP4"},
@@ -294,6 +390,23 @@ def test_detect_modelopt_mixed_fp8_only(tmp_path):
     assert det.scheme is qd.QuantScheme.FP8_PER_CHANNEL_CT
 
 
+def test_detect_modelopt_fp8_and_unknown(tmp_path):
+    _write_config(tmp_path, {
+        "quantization_config": {"quant_method": "modelopt", "quant_algo": "FP8"},
+    })
+    assert qd.detect_quant_scheme(str(tmp_path)).scheme is qd.QuantScheme.FP8_PER_CHANNEL_CT
+
+    _write_config(tmp_path, {
+        "quantization_config": {
+            "quant_method": "modelopt",
+            "kv_cache_quant_algo": "INT8",
+        },
+    })
+    det = qd.detect_quant_scheme(str(tmp_path))
+    assert det.scheme is qd.QuantScheme.UNSUPPORTED
+    assert "INT8" in det.reason
+
+
 def test_detect_unsupported_fbgemm(tmp_path):
     _write_config(tmp_path, {
         "quantization_config": {"quant_method": "fbgemm_fp8"},
@@ -303,6 +416,15 @@ def test_detect_unsupported_fbgemm(tmp_path):
     assert "fbgemm" in det.reason
 
 
+def test_detect_nonempty_quant_config_without_method_fails_closed(tmp_path):
+    _write_config(tmp_path, {
+        "quantization_config": {"bits": 4, "format": "unknown-packed-layout"},
+    })
+    det = qd.detect_quant_scheme(str(tmp_path))
+    assert det.scheme is qd.QuantScheme.UNSUPPORTED
+    assert "quant_method ''" in det.reason
+
+
 @pytest.mark.parametrize("method", ["gptq", "awq", "bitsandbytes"])
 def test_detect_passthrough_schemes(tmp_path, method):
     _write_config(tmp_path, {"quantization_config": {"quant_method": method}})
@@ -310,10 +432,335 @@ def test_detect_passthrough_schemes(tmp_path, method):
     assert det.scheme is qd.QuantScheme.NONE
 
 
+def test_detect_remote_checkpoint_pins_revision_and_offline_mode(
+    tmp_path, monkeypatch,
+):
+    config_path = tmp_path / "config.json"
+    config_path.write_text(json.dumps({
+        "quantization_config": {
+            "quant_method": "fp8",
+            "weight_block_size": [128, 128],
+        },
+    }))
+    calls = []
+
+    def fake_download(repo, filename, **kwargs):
+        calls.append((repo, filename, kwargs))
+        return str(config_path)
+
+    monkeypatch.setattr("huggingface_hub.hf_hub_download", fake_download)
+    det = qd.detect_quant_scheme(
+        "org/model",
+        token="token",
+        revision="immutable-sha",
+        local_files_only=True,
+    )
+
+    assert det.scheme is qd.QuantScheme.FP8_BLOCKWISE
+    assert calls == [(
+        "org/model",
+        "config.json",
+        {
+            "token": "token",
+            "revision": "immutable-sha",
+            "local_files_only": True,
+        },
+    )]
+
+
+def test_loader_propagates_revision_and_offline_mode_to_detection(monkeypatch):
+    from obliteratus.models import loader as loader_mod
+
+    captured = {}
+    monkeypatch.setattr(
+        loader_mod.AutoConfig,
+        "from_pretrained",
+        staticmethod(lambda *args, **kwargs: SimpleNamespace(quantization_config=None)),
+    )
+
+    def fake_detect(model_name, **kwargs):
+        captured["model_name"] = model_name
+        captured.update(kwargs)
+        return qd.QuantDetection(qd.QuantScheme.UNSUPPORTED, reason="test sentinel")
+
+    monkeypatch.setattr(qd, "detect_quant_scheme", fake_detect)
+    with pytest.raises(RuntimeError, match="test sentinel"):
+        loader_mod.load_model(
+            "org/model",
+            task="causal_lm",
+            device="cpu",
+            dtype="float32",
+            revision="immutable-sha",
+            local_files_only=True,
+        )
+
+    assert captured == {
+        "model_name": "org/model",
+        "token": None,
+        "revision": "immutable-sha",
+        "local_files_only": True,
+    }
+
+
+def test_materialize_remote_checkpoint_pins_revision_and_offline_mode(
+    tmp_path, monkeypatch,
+):
+    from safetensors.torch import save_file
+
+    source = tmp_path / "source"
+    source.mkdir()
+    _write_config(source, {
+        "model_type": "gpt2",
+        "quantization_config": {"quant_method": "fp8"},
+    })
+    save_file({"plain.weight": torch.ones(2, 2)}, source / "model.safetensors")
+    captured = {}
+
+    def fake_snapshot(repo, **kwargs):
+        captured["repo"] = repo
+        captured.update(kwargs)
+        return str(source)
+
+    monkeypatch.setattr("huggingface_hub.snapshot_download", fake_snapshot)
+    output, returned_source = qd.materialize_dequantized_checkpoint(
+        "org/model",
+        qd.QuantDetection(qd.QuantScheme.FP8_BLOCKWISE),
+        token="token",
+        revision="immutable-sha",
+        local_files_only=True,
+    )
+    try:
+        assert returned_source == str(source)
+        assert captured["repo"] == "org/model"
+        assert captured["token"] == "token"
+        assert captured["revision"] == "immutable-sha"
+        assert captured["local_files_only"] is True
+        assert "*.safetensors" in captured["allow_patterns"]
+    finally:
+        import shutil
+
+        shutil.rmtree(output)
+
+
+def test_materialize_failure_removes_partial_checkpoint(tmp_path, monkeypatch):
+    import tempfile
+
+    source = tmp_path / "source"
+    source.mkdir()
+    _write_config(source, {
+        "model_type": "gpt2",
+        "quantization_config": {"quant_method": "fp8"},
+    })
+    partial = tmp_path / "partial-output"
+
+    def fake_mkdtemp(*, prefix):
+        assert prefix == "obliteratus_dequant_"
+        partial.mkdir()
+        return str(partial)
+
+    monkeypatch.setattr(tempfile, "mkdtemp", fake_mkdtemp)
+    with pytest.raises(RuntimeError, match="no safetensors weights"):
+        qd.materialize_dequantized_checkpoint(
+            str(source),
+            qd.QuantDetection(qd.QuantScheme.FP8_BLOCKWISE),
+        )
+
+    assert not partial.exists()
+
+
+def test_materialize_sharded_checkpoint_rewrites_index_and_copies_metadata(tmp_path):
+    import shutil
+
+    from safetensors.torch import load_file, save_file
+
+    save_file({"a.weight": torch.ones(1)}, tmp_path / "part-1.safetensors")
+    save_file({"b.weight": torch.ones(1)}, tmp_path / "part-2.safetensors")
+    (tmp_path / "model.safetensors.index.json").write_text(json.dumps({
+        "metadata": {"total_size": 8},
+        "weight_map": {
+            "a.weight": "part-1.safetensors",
+            "b.weight": "part-2.safetensors",
+        },
+    }))
+    _write_config(tmp_path, {
+        "model_type": "gpt2",
+        "quantization_config": {"quant_method": "fp8"},
+    })
+    (tmp_path / "tokenizer_config.json").write_text('{"test": true}')
+    output, _ = qd.materialize_dequantized_checkpoint(
+        str(tmp_path),
+        qd.QuantDetection(qd.QuantScheme.FP8_BLOCKWISE),
+        out_dtype=torch.float32,
+    )
+    try:
+        index = json.loads(Path(output, "model.safetensors.index.json").read_text())
+        assert index["metadata"] == {"total_size": 8}
+        assert index["weight_map"] == {
+            "a.weight": "part-1.safetensors",
+            "b.weight": "part-2.safetensors",
+        }
+        assert load_file(Path(output, "part-1.safetensors"))["a.weight"].item() == 1
+        assert json.loads(Path(output, "tokenizer_config.json").read_text()) == {
+            "test": True,
+        }
+    finally:
+        shutil.rmtree(output)
+
+
+@requires_fp8
+def test_materialize_resolves_scale_tensor_across_shard_boundary(tmp_path):
+    import shutil
+
+    from safetensors.torch import load_file, save_file
+
+    weight = torch.tensor([[0.5, 1.0]], dtype=torch.float8_e4m3fn)
+    save_file({"a.weight": weight}, tmp_path / "part-1.safetensors")
+    save_file(
+        {
+            "a.weight_scale": torch.tensor([2.0]),
+            "norm.weight": torch.ones(1),
+        },
+        tmp_path / "part-2.safetensors",
+    )
+    (tmp_path / "model.safetensors.index.json").write_text(json.dumps({
+        "metadata": {"total_size": 6},
+        "weight_map": {
+            "a.weight": "part-1.safetensors",
+            "a.weight_scale": "part-2.safetensors",
+            "norm.weight": "part-2.safetensors",
+        },
+    }))
+    _write_config(tmp_path, {
+        "model_type": "gpt2",
+        "quantization_config": {"quant_method": "modelopt", "quant_algo": "FP8"},
+    })
+
+    output, _ = qd.materialize_dequantized_checkpoint(
+        str(tmp_path),
+        qd.QuantDetection(qd.QuantScheme.FP8_PER_CHANNEL_CT),
+        out_dtype=torch.float32,
+    )
+    try:
+        index = json.loads(Path(output, "model.safetensors.index.json").read_text())
+        assert index["metadata"]["total_size"] == 12
+        assert index["weight_map"] == {
+            "a.weight": "part-1.safetensors",
+            "norm.weight": "part-2.safetensors",
+        }
+        assert load_file(Path(output, "part-1.safetensors"))["a.weight"].tolist() == [
+            [1.0, 2.0],
+        ]
+        assert set(load_file(Path(output, "part-2.safetensors"))) == {"norm.weight"}
+    finally:
+        shutil.rmtree(output)
+
+
+def test_loader_failure_removes_materialized_checkpoint(tmp_path, monkeypatch):
+    from obliteratus.models import loader as loader_mod
+
+    materialized = tmp_path / "materialized"
+    materialized.mkdir()
+    config = SimpleNamespace(quantization_config=None)
+    monkeypatch.setattr(
+        loader_mod.AutoConfig,
+        "from_pretrained",
+        staticmethod(lambda *args, **kwargs: config),
+    )
+    monkeypatch.setattr(
+        qd,
+        "detect_quant_scheme",
+        lambda *args, **kwargs: qd.QuantDetection(qd.QuantScheme.FP8_BLOCKWISE),
+    )
+    monkeypatch.setattr(
+        qd,
+        "materialize_dequantized_checkpoint",
+        lambda *args, **kwargs: (str(materialized), "source"),
+    )
+    monkeypatch.setattr(loader_mod, "_estimate_model_memory_gb", lambda *args: 0.0)
+
+    class FailingModelClass:
+        @staticmethod
+        def from_pretrained(**kwargs):
+            raise OSError("load failed")
+
+    monkeypatch.setattr(loader_mod, "_select_model_class", lambda *args: FailingModelClass)
+    with pytest.raises(OSError, match="load failed"):
+        loader_mod.load_model(
+            "org/model",
+            task="causal_lm",
+            device="cpu",
+            dtype="float32",
+        )
+
+    assert not materialized.exists()
+
+
+def test_smoke_loader_requires_explicit_remote_code_trust():
+    from scripts import smoke_load_quant
+
+    parser = smoke_load_quant.build_parser()
+    assert parser.parse_args(["org/model"]).trust_remote_code is False
+    assert parser.parse_args([
+        "org/model", "--trust-remote-code",
+    ]).trust_remote_code is True
+
+
+def test_numerical_fixture_provenance_is_machine_readable():
+    provenance = json.loads(
+        (Path(__file__).parent / "fixtures" / "quant_dequant_provenance.json").read_text()
+    )
+    assert provenance["origin"] == "synthetic-runtime-generated"
+    assert provenance["external_data"] is False
+    assert provenance["torch_manual_seeds"] == list(range(9))
+    assert provenance["oracles"]["fp8_relative_l2_max"] == 0.05
+    assert provenance["oracles"]["nvfp4_relative_l2_max"] == 0.15
+
+
 # ---------------------------------------------------------------------------
 # State-dict dequantization
 # ---------------------------------------------------------------------------
 
+
+@requires_fp8
+def test_upcast_raw_fp8_bytes_and_reject_plain_float():
+    fp8 = torch.tensor([0.5, -1.0], dtype=torch.float8_e4m3fn)
+    raw = fp8.view(torch.uint8)
+    assert torch.equal(qd._upcast_fp8(raw), fp8.float())
+    with pytest.raises(RuntimeError, match="cannot upcast"):
+        qd._upcast_fp8(torch.ones(2, dtype=torch.float32))
+
+
+@requires_fp8
+def test_fp8_blockwise_rejects_incompatible_scale_shape():
+    weight = torch.ones(4, 4, dtype=torch.float8_e4m3fn)
+    with pytest.raises(RuntimeError, match="incompatible"):
+        qd.dequantize_fp8_blockwise(weight, torch.ones(1, 1), block_size=(2, 2))
+
+
+def test_unpack_native_failure_falls_back_to_manual(monkeypatch):
+    packed = torch.tensor([[0x1B]], dtype=torch.uint8)
+    monkeypatch.setattr(qd, "_native_fp4_upcast_works", lambda: True)
+    monkeypatch.setattr(
+        qd,
+        "_unpack_e2m1_native",
+        lambda value: (_ for _ in ()).throw(RuntimeError("unsupported")),
+    )
+    assert qd.unpack_e2m1(packed).tolist() == [[-1.5, 0.5]]
+
+
+def test_nvfp4_inverse_block_scale_and_output_shape():
+    packed = torch.tensor([[0x11] * 8], dtype=torch.uint8)
+    output = qd.dequantize_nvfp4(
+        packed,
+        torch.full((1, 1), 2.0),
+        None,
+        out_shape=(2, 8),
+        scale_is_inverse=True,
+        force_manual=True,
+    )
+    assert output.shape == (2, 8)
+    assert torch.all(output == 0.25)
+
 @requires_fp8
 def test_dequantize_state_dict_fp8_blockwise():
     torch.manual_seed(7)
@@ -350,6 +797,75 @@ def test_dequantize_state_dict_nvfp4_modelopt():
     assert cos > 0.995
 
 
+@requires_fp8
+def test_dequantize_state_dict_nvfp4_compressed_tensors_contract():
+    torch.manual_seed(8)
+    w = torch.randn(32, 64)
+    packed, bs, gs = _pack_nvfp4(w, global_reciprocal=True)
+    sd = {
+        "layers.0.mlp.weight_packed": packed,
+        "layers.0.mlp.weight_scale": bs,
+        "layers.0.mlp.weight_global_scale": gs,
+        "layers.0.mlp.input_global_scale": torch.tensor(1.0),
+    }
+    det = qd.QuantDetection(
+        qd.QuantScheme.NVFP4_CT,
+        global_scale_is_inverse=True,
+    )
+    out = qd.dequantize_state_dict(sd, det, out_dtype=torch.float32)
+    assert set(out) == {"layers.0.mlp.weight"}
+    cos = torch.nn.functional.cosine_similarity(
+        w.flatten(), out["layers.0.mlp.weight"].flatten(), dim=0,
+    )
+    rel = (out["layers.0.mlp.weight"] - w).norm() / w.norm()
+    assert cos > 0.995
+    assert rel < 0.15
+
+
+@requires_fp8
+def test_materialize_nvfp4_compressed_tensors_checkpoint_contract(tmp_path):
+    import shutil
+
+    from safetensors.torch import load_file, save_file
+
+    torch.manual_seed(8)
+    w = torch.randn(32, 64)
+    packed, bs, gs = _pack_nvfp4(w, global_reciprocal=True)
+    save_file({
+        "layers.0.mlp.weight_packed": packed,
+        "layers.0.mlp.weight_scale": bs,
+        "layers.0.mlp.weight_global_scale": gs,
+        "layers.0.mlp.input_global_scale": torch.tensor(1.0),
+        "layers.0.norm.weight": torch.ones(32),
+    }, tmp_path / "model.safetensors")
+    _write_config(tmp_path, {
+        "model_type": "gpt2",
+        "quantization_config": {
+            "quant_method": "compressed-tensors",
+            "format": "nvfp4-pack-quantized",
+        },
+    })
+    det = qd.QuantDetection(
+        qd.QuantScheme.NVFP4_CT,
+        global_scale_is_inverse=True,
+    )
+    output, source = qd.materialize_dequantized_checkpoint(
+        str(tmp_path), det, out_dtype=torch.float32,
+    )
+    try:
+        assert source == str(tmp_path)
+        state = load_file(os.path.join(output, "model.safetensors"))
+        assert set(state) == {"layers.0.mlp.weight", "layers.0.norm.weight"}
+        rel = (state["layers.0.mlp.weight"] - w).norm() / w.norm()
+        assert rel < 0.15
+        materialized_config = json.loads(
+            Path(output, "config.json").read_text(encoding="utf-8")
+        )
+        assert "quantization_config" not in materialized_config
+    finally:
+        shutil.rmtree(output)
+
+
 @requires_fp8
 def test_dequantize_state_dict_fp8_missing_scale_raises():
     q = torch.randn(16, 16).to(torch.float8_e4m3fn)
@@ -358,6 +874,47 @@ def test_dequantize_state_dict_fp8_missing_scale_raises():
         qd.dequantize_state_dict({"a.weight": q}, det)
 
 
+@requires_fp8
+def test_dequantize_state_dict_fp8_per_channel():
+    state = {
+        "a.weight": torch.tensor([[0.5, 1.0]], dtype=torch.float8_e4m3fn),
+        "a.weight_scale": torch.tensor([2.0]),
+    }
+    out = qd.dequantize_state_dict(
+        state,
+        qd.QuantDetection(qd.QuantScheme.FP8_PER_CHANNEL_CT),
+        out_dtype=torch.float32,
+    )
+    assert set(out) == {"a.weight"}
+    assert out["a.weight"].tolist() == [[1.0, 2.0]]
+
+
+def test_dequantize_state_dict_rejects_packed_weight_for_wrong_scheme():
+    state = {
+        "a.weight": torch.zeros(1, 8, dtype=torch.uint8),
+        "a.weight_scale": torch.ones(1, 1),
+    }
+    with pytest.raises(RuntimeError, match="unsupported layout"):
+        qd.dequantize_state_dict(
+            state,
+            qd.QuantDetection(qd.QuantScheme.FP8_BLOCKWISE),
+        )
+
+
+def test_dequantize_state_dict_rejects_unscaled_packed_weight():
+    with pytest.raises(RuntimeError, match="no recognizable scale siblings"):
+        qd.dequantize_state_dict(
+            {"a.weight_packed": torch.zeros(1, 8, dtype=torch.uint8)},
+            qd.QuantDetection(qd.QuantScheme.NVFP4_CT),
+        )
+
+
+def test_scale_key_requires_matching_weight():
+    state = {"logit_scale": torch.tensor(1.0)}
+    assert qd.is_scale_key("logit_scale", state) is False
+    assert qd.is_scale_key("ordinary.weight", state) is False
+
+
 # ---------------------------------------------------------------------------
 # Surgery guards
 # ---------------------------------------------------------------------------
@@ -403,6 +960,81 @@ def test_dequantize_weight_plain_uint8_still_converts():
     assert w.is_floating_point()
 
 
+def test_fused_uint8_guard_requires_quantization_scale_sibling():
+    from obliteratus.abliterate import AbliterationPipeline
+
+    container = torch.nn.Module()
+    container.w1 = torch.nn.Parameter(
+        torch.zeros(2, 4, 2, dtype=torch.uint8), requires_grad=False,
+    )
+    container.w1_scale = torch.ones(2, 4, 1)
+    with pytest.raises(RuntimeError, match="without dequantization"):
+        AbliterationPipeline._project_fused_3d(
+            container,
+            torch.ones(2, 1),
+            ["w1"],
+            norm_preserve=False,
+            scale=1.0,
+        )
+
+
+@requires_fp8
+def test_fused_fp8_guard_rejects_raw_storage():
+    from obliteratus.abliterate import AbliterationPipeline
+
+    container = torch.nn.Module()
+    container.w1 = torch.nn.Parameter(
+        torch.zeros(2, 4, 2).to(torch.float8_e4m3fn), requires_grad=False,
+    )
+    with pytest.raises(RuntimeError, match="without dequantization"):
+        AbliterationPipeline._project_fused_3d(
+            container,
+            torch.ones(2, 1),
+            ["w1"],
+            norm_preserve=False,
+            scale=1.0,
+        )
+
+
+def test_granular_fused_uint8_guard_rejects_packed_storage():
+    from obliteratus.abliterate import AbliterationPipeline
+
+    container = torch.nn.Module()
+    container.w1 = torch.nn.Parameter(
+        torch.zeros(2, 4, 2, dtype=torch.uint8), requires_grad=False,
+    )
+    container.w1_scale = torch.ones(2, 4, 1)
+    with pytest.raises(RuntimeError, match="without dequantization"):
+        AbliterationPipeline._project_fused_3d_granular(
+            container,
+            torch.ones(2),
+            {},
+            ["w1"],
+            norm_preserve=False,
+            scale=1.0,
+        )
+
+
+def test_selective_fused_uint8_guard_rejects_packed_storage():
+    from obliteratus.abliterate import AbliterationPipeline
+
+    container = torch.nn.Module()
+    container.w1 = torch.nn.Parameter(
+        torch.zeros(2, 4, 2, dtype=torch.uint8), requires_grad=False,
+    )
+    container.w1_global_scale = torch.ones(1)
+    with pytest.raises(RuntimeError, match="without dequantization"):
+        AbliterationPipeline._project_fused_3d_selective_inversion(
+            container,
+            torch.ones(2),
+            ["w1"],
+            safety_indices={0},
+            reflect_scale=2.0,
+            remove_scale=1.0,
+            norm_preserve=False,
+        )
+
+
 # ---------------------------------------------------------------------------
 # End-to-end loader on tiny synthetic checkpoints
 # ---------------------------------------------------------------------------
diff --git a/tests/test_strategy_navigation_contracts.py b/tests/test_strategy_navigation_contracts.py
index 40eb2ca..ee1b37e 100644
--- a/tests/test_strategy_navigation_contracts.py
+++ b/tests/test_strategy_navigation_contracts.py
@@ -63,6 +63,15 @@ class _NoEmbeddingModel(nn.Module):
         self.model.layers = nn.ModuleList([nn.Module()])
 
 
+class _NemotronHModel(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.backbone = nn.Module()
+        layer = nn.Module()
+        layer.mixer = nn.Module()
+        self.backbone.layers = nn.ModuleList([layer])
+
+
 def _handle(model: nn.Module, *, architecture: str, hidden_size: int = 8, num_layers: int = 1, num_heads: int = 2):
     return ModelHandle(
         model=model,
@@ -93,6 +102,15 @@ def test_strategy_navigation_resolves_fallback_layers_and_missing_attention():
         get_attention_module(broken, "qwen3_5_moe")
 
 
+def test_nemotron_h_navigation_uses_backbone_layers_and_mixer():
+    handle = _handle(_NemotronHModel(), architecture="nemotron_h")
+
+    layers = get_layer_modules(handle)
+    assert layers is handle.model.backbone.layers
+    assert get_attention_module(layers[0], handle.architecture) is layers[0].mixer
+    assert get_ffn_module(layers[0], handle.architecture) is layers[0].mixer
+
+
 def test_head_pruning_zeros_qkv_and_output_slices_for_standard_attention():
     model = nn.Module()
     model.model = nn.Module()