Recognize trainable pre-norm attention chains as graph-fusion candidates #871

Pattern\n\nExtend the synchronized-composite detector proposed in #870 to recognize trainable pre-norm attention boundaries hidden across shallow call edges:\n\nLayerNorm -> Q/K/V MatMul siblings -> packed noncausal SDPA/MHA -> output MatMul -> residual Add.\n\nThe high-value shape is repeated inside a transformer-depth loop on a synchronous accelerator backend. Forward-only analysis is incomplete: the matching VJP also crosses projection, SDPA-backward, output-projection, and LayerNorm-backward boundaries.\n\n## Independently validated GoAI case\n\nThe batch-8 ViT path used seven synchronous Metal operations around attention. A bounded shape-keyed MPSGraph now executes the forward as one submission and the explicit seven-input backward as one submission. Runtime epsilon remains a feed rather than a cache-key dimension.\n\nApple M2 Pro, F32, batch 8, sequence 65, dimension 128, 4 heads:\n\n- boundary median speedups across three fresh-process order-alternated count-7 campaigns: 2.968x, 2.986x, 3.059x;\n- depth-4 ViT training-step median speedups: 2.110x, 2.062x, 2.062x;\n- weakest of 21 aligned full-step pairs: 1.793x;\n- output, all seven gradients, full-model logits, all parameter gradients, and input immutability passed parity gates.\n\nThe control disabled only the new attention fusion, leaving the already-merged FFN graph fusion active in both arms. This isolates the incremental attention leverage.\n\n## Detector direction\n\nBuild a shallow interprocedural dispatch DAG rather than only a linear call sequence. Rank a candidate higher when:\n\n- one normalized activation fans out into three same-shape projection MatMuls;\n- the projections reconverge at a packed MHA/SDPA op, followed by output projection and residual Add;\n- the chain repeats in a depth loop;\n- the backend is synchronous or exposes commit-and-wait native calls;\n- stable shape, dtype, batch, head count, and epsilon make graph caching bounded;\n- matching backward dispatches exist for every differentiable input.\n\nReport forward and backward eager-op counts, the depth multiplier, feature gates, and a required same-binary end-to-end benchmark. Suppress or downgrade when bias, LoRA, explicit masks, causal behavior, dynamic shapes, device crossings, or missing fused backward capability change the semantics.\n\nThis is a concrete attention-specific specialization of #870 rather than evidence that dispatch-count reduction alone wins; the earlier 51-to-5 ViT preparation flattening remained neutral.