Generalize dense-layer MLP hooks across interleaved MoE architectures

[jlarson4]

Description

• Fix DeepSeek V2 dense MLP hook semantics #1651 fixed the dense-prefix hook semantics for DeepSeek V2 with a per-arch MoEBridge subclass carrying fallback-list aliases. The same mapping exists in sixteen other adapters, so this generalizes the fix into MoEBridge itself and retires that subclass.
• MoEBridge.set_original_component now dispatches per layer at bind time: when a layer's module actually exposes the adapter-declared dense projections, that instance adopts the gated-MLP alias set (hook_pre, hook_pre_linear, hook_post) plus W_gate/W_in/W_out and b_*. Sparse layers are untouched. Detection is positive-only — declared and present — so no sparse block can be mis-detected by a naming quirk and stripped of its hooks.
• The keys are deliberately not gate/in/out: on a sparse layer of the same model gate is the router, so reusing it would make blocks.N.mlp.gate.hook_out mean the router on some layers and a d_mlp projection on others. dense_* is what llada2_moe, llama4, and laguna already declared, so those three are fixed by the shared component with no edit.
• Ungated dense feed-forwards are expressible too: dense_in/dense_out alone bind the MLP-style alias set (no hook_pre_linear, no W_gate), which is what Switch Transformers' wi/wo needs.
• Rollout: deepseek_v2, deepseek_v3, glm4_moe, glm4_moe_lite, glm_moe_dsa, ernie4_5_moe, afmoe, jamba, switch_transformers, laguna, llama4, llada2_moe, and the Qwen MoE family (qwen2_moe, qwen3_moe, qwen3_next, qwen3_vl_moe) — where the MoE-only submodules were additionally non-optional, so an mlp_only_layers / decoder_sparse_step checkpoint crashed at boot instead of degrading. qwen3_next and llama4 also gain the router hooks they never mapped. qwen3_5_moe is deliberately excluded: HF builds its sparse block unconditionally, so its dense declarations were dead and its MoE submodules are required.
• optional alone cannot say "required on sparse layers, absent on dense ones", which made a renamed HF router bind silently with no hooks — the failure this work removes. Adapters now declare sparse_required=(...); a sparse layer missing one raises from a validate_after_setup hook called once setup_submodules knows the skipped set. The tuple is validated at construction, since a typo would otherwise silently disable the guard. Symmetrically, a declared dense_gate that does not resolve on a dense layer is treated as a rename and raises, rather than binding as ungated and aliasing hook_pre to the up projection.
• get_params() keyed on submodules literally named in/out, so dense MoE layers fell into a bare except AttributeError and were silently zero-filled — real weights on the component, zeros in the parameter dict. It now resolves dense_* on dense binds, and the remaining zero-fill path logs a warning naming the layer instead of failing silent.
• Dense instances drop hook_router_scores rather than advertise a hook that can never fire (and forward no longer calls it), the dispatch is symmetric so a rebinding harness cannot leave a chimera, and hook_aliases/property_aliases are exempt from GeneralizedComponent.__setattr__'s passthrough — without that, a per-layer rebind is forwarded to the wrapped HF module whenever it exposes the attribute and silently vanishes.
• Alias collection: both collectors are now cycle-safe with a path-scoped guard (a globally-visited set cuts cycles but also silently drops a component legitimately shared under two names); _collect_component_aliases runs first on every hook_dict access and previously raised RecursionError on a 2-cycle. The block walk is unified into one loop, warns on unresolved aliases naming the stale-fallback consequence, and is memoized against (registry size, alias generation) — size alone cannot see a dense↔sparse rebind, which leaves the registry the same size while changing what the aliases point at. List-valued targets are skipped by the template collector, whose consumer reverse-matches with str.endswith and is lru_cached.
• Tests, each verified to fail with the fix reverted: unit dispatch coverage (dense/sparse binding, weight accessors pinned to the wrapped HF projections with a negative control, dropped router hook, gate-name non-overlap, template independence, symmetric rebind, __setattr__ passthrough survival, positive-detection-only, ungated dense, renamed-gate and renamed-router guards) bound through the real setup_submodules; a self-checking roster that greps the adapters directory and fails if a dense-declaring adapter is outside the guards; alias-traversal tests for cycles and diamonds; and integration on katuni4ka/tiny-random-deepseek-v3 covering hook shapes, alias identities, get_params weights, logits parity, and cache invalidation across a rebind.
• Bridge-vs-HF parity for the tiny roster now uses a measured, scale-relative tolerance instead of a platform-keyed constant. Measured across all 12 checkpoints on macOS-arm64: worst drift 2.2e-07, nine bit-identical, against a former 1e-2 gate — a 10% error injected into the reconstructed attention scale passed everywhere and now fails. Relative because the roster's logit spreads differ by ~700x, so no single absolute number is both stable on the widest checkpoint and discriminating on the narrowest.
• CI model caches: the tiny checkpoints these suites need are in the path lists of every job that runs them (including the macOS/MPS job, which ran them against a cache listing only TinyStories), the key is bumped so the additions actually save, and the benchmark job no longer shares that key with a one-path list.

Type of change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• This change requires a documentation update

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility