⚡ Bolt: [performance improvement] Optimize HullKVCache hot paths

.jules/bolt.md

@@ -0,0 +1,3 @@
+## 2025-04-06 - Dict setdefault in hot loops
+**Learning:** `dict.setdefault(key, complex_default)` eagerly evaluates `complex_default` on every iteration, leading to massive memory allocation and performance overhead when creating default lists/dicts in hot loops (like in `HullKVCache._rebuild_if_needed`).
+**Action:** Replace `setdefault` with an explicit `if key not in dict:` membership check to lazily evaluate the default structure only when necessary.

scripts/export_p3_paper_freeze.py

@@ -6,7 +6,6 @@
 import json
 from pathlib import Path
 import re
-from typing import Any
 
 from utils import detect_runtime_environment
 

scripts/export_p5_public_surface_sync.py

@@ -4,7 +4,6 @@
 
 import json
 from pathlib import Path
-from typing import Any
 
 from utils import detect_runtime_environment
 

scripts/export_r10_d0_same_endpoint_cost_attribution.py

@@ -21,7 +21,7 @@
     subroutine_braid_long_program,
     verify_program,
 )
-from exec_trace import ExecutionState, Program, TraceEvent, replay_trace, TraceInterpreter
+from exec_trace import Program, TraceEvent, replay_trace, TraceInterpreter
 from geometry import brute_force_hardmax_2d
 from model.exact_hardmax import encode_latest_write_query
 from model.free_running_executor import (

scripts/export_r2_systems_baseline_gate.py

@@ -2,7 +2,7 @@
 
 from __future__ import annotations
 
-from collections import Counter, defaultdict
+from collections import defaultdict
 import csv
 import json
 from pathlib import Path

scripts/export_r3_d0_exact_execution_stress_gate.py

@@ -4,7 +4,6 @@
 
 import json
 from pathlib import Path
-from typing import Any
 
 from bytecode import (
     lower_program,

scripts/export_release_worktree_hygiene_snapshot.py

@@ -5,7 +5,6 @@
 import json
 from pathlib import Path
 import subprocess
-from typing import Any
 
 from utils import detect_runtime_environment
 

src/geometry/hull_kv.py

@@ -17,7 +17,6 @@
     HardmaxResult,
     NumberLike,
     ValueLike,
-    _as_fraction,
     _coerce_key,
     _coerce_value,
     _normalize_number,
@@ -205,15 +204,18 @@ def _rebuild_if_needed(self) -> None:
         total_value_sum = [Fraction(0) for _ in range(self._value_width or 0)]
 
         for index, (key, value) in enumerate(self._entries):
-            bucket = aggregates.setdefault(
-                key,
-                {"value_sum": [Fraction(0) for _ in value], "count": 0, "entry_indices": []},
-            )
-            for coord_index, coord in enumerate(value):
-                bucket["value_sum"][coord_index] += coord
-                total_value_sum[coord_index] += coord
-            bucket["count"] += 1
-            bucket["entry_indices"].append(index)
+            if key not in aggregates:
+                bucket = {"value_sum": list(value), "count": 1, "entry_indices": [index]}
+                aggregates[key] = bucket
+                for coord_index, coord in enumerate(value):
+                    total_value_sum[coord_index] += coord
+            else:
+                bucket = aggregates[key]
+                for coord_index, coord in enumerate(value):
+                    bucket["value_sum"][coord_index] += coord
+                    total_value_sum[coord_index] += coord
+                bucket["count"] += 1
+                bucket["entry_indices"].append(index)
 
         points: list[_PointAggregate] = []
         for key in sorted(aggregates):
@@ -243,9 +245,11 @@ def _scan_maximizers(
 
         best_score = target_score
         maximizers: list[_PointAggregate] = []
+        qx, qy = query
 
         for point in self._points:
-            score = dot_2d(point.key, query)
+            px, py = point.key
+            score = (px * qx) + (py * qy)
             if best_score is None or score > best_score:
                 best_score = score
                 maximizers = [point]

tests/test_bytecode_harness.py

@@ -1,7 +1,6 @@
 from __future__ import annotations
 
 from bytecode import (
-    accumulator_loop_program,
     alternating_memory_loop_bytecode_program,
     arithmetic_smoke_program,
     branch_then_call_false_program,

tests/test_bytecode_memory_surfaces.py

@@ -7,7 +7,6 @@
 from bytecode import (
     analyze_memory_surfaces,
     call_frame_roundtrip_program,
-    countdown_helper_call_program,
     memory_surface_cases,
     memory_surface_negative_programs,
     run_memory_surface_case,

tests/test_model_softmax_baseline.py

@@ -13,7 +13,6 @@
     evaluate_free_running_rollout,
     evaluate_teacher_forced_model,
     require_torch,
-    serialize_event_tokens,
     SoftmaxBaselineConfig,
     SoftmaxTrainingConfig,
     summarize_trace_sequences,