Bug/numeric parity fixes

[jlarson4]

Description

• MPT's MptMLP adds the residual inside its forward, so blocks.N.hook_mlp_out exposed resid_post instead of the additive contribution. This is a reoccurrence of the [Bug Report] BLOOM TransformerBridge maps hook_attn_out / hook_mlp_out to residual-added states, breaking HookedTransformer residual identities #1639 BLOOM issue in another arch. The alias now targets the down-projection output (same pattern as the BLOOM Fix BLOOM residual branch hook semantics #1640 fix), and MPT joins the residual-decomposition-identity integration fixture.
• Fixing MPT exposed general MLPBridge hook-fire bugs, symmetric on both edges: forward unconditionally re-fired out.hook_out with the module's final output (stamping the residual-added state over the projection output on residual-inside MLPs, double-applying interventions on ordinary container MLPs), and pre-fired in.hook_in on a tensor the replaced in projection then hooked again inside the wrapped forward. Both are now single-fire: the out-side re-fire is a fallback gated on whether the projection actually ran (fired flag on LinearBridge), and the in-side pre-fire suppresses the inner duplicate for one call. GatedMLPBridge's processed-weights functional path had the mirror-image gap — it bypasses the wrapped projections and never fired out.hook_out at all; it now fires it with the down-projection output. The fired/suppress flags are a declared class-level protocol on GeneralizedComponent, implemented identically by LinearBridge and Conv1DBridge (GPT-2-family Conv1D projections included), with object.__setattr__ writes to stay off the nn.Module setattr path.
• MLAAttentionBridge hardcoded qk_head_dim^-0.5, dropping the yarn mscale² factor HF folds into self.scaling — DeepSeek-V3/R1 and GLM4-MoE-Lite configs score ~1.87× too small ([Bug Report] TransformerBridge activations on gpt-oss-20b disagree with the HF model it wraps #1618-class silent divergence). The bridge now scores with the wrapped module's scaling, falling back to softmax_scale for trust_remote_code DeepSeek-V2 modules that use that name for the same quantity.
• MPTALiBiAttentionBridge hardcoded the same base scale, ignoring the module's resolved softmax_scale, and gated clip_qkv on is not None where HF gates on truthiness — a checkpoint spelling "disabled" as 0.0 clamped Q/K/V to zeros. Both now match HF; the Q/K/V clamp itself is extracted to a shared clamp_qkv utility used by both bridges, and the reconstructed path refuses (loudly) the untested clip-plus-post-reshape-norm combination no arch currently has.
• Falcon-H1 multiplies K by a learned key_multiplier inside its attention forward; the reconstructed path dropped it. PositionEmbeddingsAttentionBridge now applies it post-projection/pre-RoPE, matching HF order.
• OLMo v1 / OLMoE clip_qkv was never applied under hooks: the reconstruction had no clamp, and the old OLMo prepare_model patch additionally disabled the native module's clamp — so nothing clamped anywhere. The reconstruction now clamps Q/K/V out-of-place, after the pre-reshape qk-norm (OLMoE norms first, then clamps) and before RoPE; the patch is deleted from both adapters.
• HookedTransformer mirror: new HookedTransformerConfig.clip_qkv field, plumbed from the HF config in the OLMo-1B/7B and OLMoE branches, applied in abstract_attention between qk-norm and rotary/cache-append. OLMo checkpoints are attention-bias-free, so weight processing does not shift the clamp threshold.
• Tests, every one verified to fail on the unfixed tree: bridge-vs-HF parity against live OlmoAttention, OlmoeAttention (pins the norm-before-clamp order), FalconH1Attention, and DeepseekV3Attention modules — all eager with an explicit 4D causal mask, so parity is tight enough (~1e-5) that each restored term breaks it when reverted; a backward test that registers real full backward hooks on the projections, which an in-place clamp_ fails (the property the deleted patch protected); an HT test proving the clamp sits between projection and scores; and MLP-bridge hook-semantics tests asserting in.hook_in/out.hook_out fire exactly once with the projection's own tensor, that additive interventions apply exactly once, that residual-inside MLPs keep the projection output on out.hook_out, and that the gated processed path fires it. Each parity test also carries a negative control proving the guarded term is active. All test setup runs under torch.random.fork_rng, leaving global RNG state untouched.

Type of change

• Bug fix (non-breaking change which fixes an issue)

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