tl.composite: fused epilogue ops with per-op scope

Extend tl.composite() with an ordered epilogue list. Each op carries
a scope flag - output_tile (default, runs once per (m,n) before
STORE), k_tile (every K-tile right after GEMM), or kernel. Plan
generator slots MATH stages by scope; pe_math reuses pe_dma's
local-loop pattern so chained epilogues (bias->relu) skip the port
hop. op_log captures per-stage params for telemetry. Topology
gains a gemm->math edge (snapshot test updated).

API stays backward-compatible - `epilogue=` is opt-in.

Example:
    h = tl.composite(
        op="gemm", a=a, b=b, out_ptr=int(out),
        epilogue=[
            {"op": "dequant", "scale": s_per_k, "scope": "k_tile"},
            {"op": "bias",    "bias":  bias_vec},
            {"op": "relu"},
            {"op": "scale",   "factor": 0.5},
        ],
    )
    tl.wait(h)

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

src/kernbench/components/builtin/tiling.py

@@ -23,6 +23,7 @@ def generate_gemm_plan(
     pe_prefix: str,
     a_pinned: bool = False,
     b_pinned: bool = False,
+    epilogue_specs: tuple = (),
 ) -> PipelinePlan:
     """Generate GEMM tile plan: M→N→K order.
 
@@ -46,7 +47,15 @@ def generate_gemm_plan(
     dma_id = f"{pe_prefix}.pe_dma"
     fetch_id = f"{pe_prefix}.pe_fetch_store"
     gemm_id = f"{pe_prefix}.pe_gemm"
-    # math_id = f"{pe_prefix}.pe_math"  # for K-accumulation if needed
+    math_id = f"{pe_prefix}.pe_math"
+
+    # Split epilogue_specs by scope. Lazy import to avoid circular dep
+    # between pe_commands ← pe_types ← tiling.
+    from kernbench.common.pe_commands import Scope as _Scope
+    k_tile_ops = [o for o in epilogue_specs
+                  if getattr(o, "scope", None) == _Scope.K_TILE]
+    out_tile_ops = [o for o in epilogue_specs
+                    if getattr(o, "scope", None) == _Scope.OUTPUT_TILE]
 
     tiles: list[TilePlan] = []
     tile_id = 0
@@ -106,9 +115,35 @@ def generate_gemm_plan(
                     },
                 ))
 
+                # K-tile-scope epilogue MATH stages (e.g. dequant) — run on
+                # every K-tile right after GEMM, before the accumulator
+                # advances to the next K slice.
+                for op in k_tile_ops:
+                    stages.append(Stage(
+                        stage_type=StageType.MATH,
+                        component=math_id,
+                        params={
+                            "op_kind": op.kind,
+                            "num_elements": tile_m * tile_n,
+                            "scope": "k_tile",
+                        },
+                    ))
+
                 # STORE + DMA_WRITE only on last K-tile per (m,n). The C
                 # accumulator stays in RegFile across the K loop.
                 if last_k:
+                    # Output-tile-scope epilogue MATH (bias, relu, ...) runs
+                    # ONCE per (m,n) after the final K-tile, before writeback.
+                    for op in out_tile_ops:
+                        stages.append(Stage(
+                            stage_type=StageType.MATH,
+                            component=math_id,
+                            params={
+                                "op_kind": op.kind,
+                                "num_elements": tile_m * tile_n,
+                                "scope": "output_tile",
+                            },
+                        ))
                     stages.append(Stage(
                         stage_type=StageType.STORE,
                         component=fetch_id,