Support mixed-width QMoE export

docs/design/per-component-quantization-loading.md

@@ -437,6 +437,27 @@ preprocessor can still pack QMoE expert banks and tie floating-point tables.
 A per-component `WeightBundle` adapter is a future migration target, not the
 currently implemented interface.
 
+Native QMoE additionally has a projection-specific adapter for Olive fused
+K-last routed experts. The initial mixed layout is deliberately narrow:
+model-wide INT4, FC1 `gate_up_proj` overridden to INT2, FC2 `down_proj`
+remaining INT4, and one common power-of-two group size of at least 16. Exact
+and `re:` Olive overrides are resolved against the original Hugging Face paths;
+plain Olive exclusions retain their producer-defined substring matching.
+GPTQ and AWQ instead use the generic plain-path subtree and `re:` full-match
+rules. FC1 and FC2 buffers are then validated independently. The adapter
+requires every routed layer and validates rank, dtype, expert count, packed byte
+width, scale geometry, and optional zero-point geometry before binding.
+Qwen3.5-VL keeps the decoder's full module plan
+(``preserve_module_plan=True``) while constructing the split package: expert
+overrides are resolved against authoritative ``model.language_model`` source
+paths, while validation and binding use the decoder-prefixed checkpoint roots
+created by the VL weight router.
+It emits QMoE's FC-specific bit attributes (including FC3=FC1 for fused
+SwiGLU). Uniform INT4 omits those attributes and retains the legacy payload
+path. Mixed-width execution remains an ONNX Runtime `NOT_IMPLEMENTED`
+boundary; Mobius tests this serialized graph contract without claiming
+runtime numerical support.
+
 Appropriate model-specific operations include:
 
 - HuggingFace-to-ONNX name alignment;

src/mobius/_configs/_base.py

@@ -651,6 +651,8 @@ class ArchitectureConfig(BaseModelConfig):
     num_experts_per_tok: int | None = None
     moe_intermediate_size: int | None = None
     shared_expert_intermediate_size: int | None = None
+    # Internal HF path aliases used to resolve routed-expert module overrides.
+    qmoe_source_paths: tuple[str, ...] = ()
     norm_topk_prob: bool = True
     # When True, the decoder layer uses post-norm style (FlexOLMo): norms are applied
     # to sub-layer outputs instead of inputs, with an extra post_feedforward_layernorm.

src/mobius/_configs/_quantization.py

@@ -100,8 +100,9 @@ class QuantizationConfig:
     # Full HuggingFace module paths (including descendants) or ``re:``-prefixed
     # full-match regexes that remain floating point inside this component.
     modules_to_not_convert: tuple[str, ...] = ()
-    # The same path/regex matching rules apply to per-module overrides.
-    # Insertion order is significant: the first matching override wins.
+    # Generic component collapse treats a plain path as a subtree. Producer-
+    # specific adapters may apply stricter matching (Olive uses literal
+    # equality); insertion order remains significant in either case.
     overrides: dict[str, QuantizationOverride] = dataclasses.field(default_factory=dict)
     # Keep this field last: QuantizationConfig has historically supported
     # positional construction, so inserting a field earlier would silently