Refactor temporal reuse: fold fill/drain into regular actions with post-hoc correction

Replace the complex in-line temporal reuse detection (parent-named
BuffetStats attributes, halo factor, bypass-zone walking) with a
simpler two-phase approach: analyze_storage folds fill/drain into
regular read/write actions, then _apply_temporal_reuse_corrections
walks per-tensor mappings to find contiguous irrelevant temporal
loops and divides out inflated stats as a post-processing step.

- Remove ~200 lines from symbolic.py (parent attrs, halo logic,
  _has_temporal_reuse, _compute_overlap_fallback, partial_overlap_info)
- Add _apply_temporal_reuse_corrections, _get_parent_buffet, and
  _compute_buffet_tile_shapes to sparse_adjustment.py
- Simplify memory.py and run_model.py action counting
- Improve energy.py KeyError message for missing actions
- Update temporal reuse test to use structural reuse mapping
- Add spatial fanout temporal reuse test and regression comparison tool
- Update sparseloop reproduction notebook outputs and regression reference

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: fisherxue

accelforge/model/_looptree/energy.py

@@ -202,8 +202,11 @@ def compute_energy_from_actions(
         component_obj = components[key.level]
         try:
             energy_per_ac = component_obj.actions[key.action].energy
-        except (KeyError, TypeError):
-            energy_per_ac = 0
+        except KeyError as e:
+            raise KeyError(
+                f"Action {key.action} not found in component {key.level}. Action occurred "
+                f"{counts.total} times."
+            ) from None
         energy_result[key] = counts.total * energy_per_ac
 
     for component_obj in spec.arch.get_nodes_of_type(arch.Component):

accelforge/model/_looptree/latency/memory.py

@@ -66,37 +66,25 @@ def component_latency(
             actions[f"{action.name}_actions"] += 0
 
         if isinstance(name2component[component], TensorHolder):
-            read_actions_val = (
-                buffet_stats.max_per_unit_read_actions
-                + buffet_stats.max_per_parent_drain_read_actions
-            )
+            # On main, max_per_unit_read_actions already includes drain reads
+            # (folded in by analyze_storage).
+            read_actions_val = buffet_stats.max_per_unit_read_actions
             actions["read_actions"] += read_actions_val
             per_tensor_reads[component][buffet.tensor] += read_actions_val
-            # Per-unit computation-path reads only (no fill/drain).
-            # Use for PE buffer BW where fills go through the parent's port.
+            # Per-unit computation-path reads (on main, same as read_actions
+            # since fill/drain are folded in).
             actions["pu_read_actions"] += buffet_stats.max_per_unit_read_actions
             # Total actions across all spatial instances (for BW throttling
             # of shared levels above spatial, e.g. shared_glb)
-            total_read_actions_val = (
-                buffet_stats.total_read_actions
-                + buffet_stats.total_parent_drain_read_actions
-            )
-            actions["total_read_actions"] += total_read_actions_val
+            actions["total_read_actions"] += buffet_stats.total_read_actions
             if not isinstance(name2component[component], arch.Toll):
-                write_actions_val = (
-                    buffet_stats.max_per_unit_write_actions
-                    + buffet_stats.max_per_parent_fill_write_actions
-                )
+                write_actions_val = buffet_stats.max_per_unit_write_actions
                 actions["write_actions"] += write_actions_val
                 per_tensor_writes[component][buffet.tensor] += write_actions_val
                 actions["pu_write_actions"] += (
                     buffet_stats.max_per_unit_write_actions
                 )
-                total_write_actions_val = (
-                    buffet_stats.total_write_actions
-                    + buffet_stats.total_parent_fill_write_actions
-                )
-                actions["total_write_actions"] += total_write_actions_val
+                actions["total_write_actions"] += buffet_stats.total_write_actions
         elif isinstance(name2component[component], arch.Compute):
             pass
         else: