Autoresearch v2: 42% faster via graph surgery + executor optimizations

[sunapi386]

Context

Why

The v1 autoresearch round optimized small arithmetic DAGs but didn't exercise the engine's heaviest code paths: ForEachNode sub-workflow expansion (graph surgery), large DAGs (100+ nodes), retry/backoff loops, and yield/resume. These are the patterns real production workflows hit.

What

Second autoresearch round targeting the 4 new heavy benchmarks. 16 experiments, 11 kept, 5 reverted. 42% faster, 70% less memory vs v1 baseline.

How

Focused on eliminating validation overhead in hot paths (_construct_trusted, model_construct), optimizing ForEachNode.run() loop (pre-computed templates, skip per-iteration with_namespace), and switching the topological executor to incremental successor-based ready-node tracking.

Summary

Component	Change	Rationale
core/workflow.py	_construct_trusted bypasses DAG validation; defaultdict for edges_by_target; Edge.model_construct in expand_node	Validation is redundant for programmatic construction from trusted code
core/node.py	Preserve cached properties in with_namespace; fast-path ShouldRetry; model_construct in _cast_input	Avoids re-creating create_model dynamic types for every namespaced copy
core/edge.py	model_construct in with_namespace	Skip Pydantic validation for simple ID prefix
nodes/iteration.py	Rewritten ForEachNode.run: skip workflow.with_namespace per iteration, pre-compute adapter templates	Biggest single target — was ~100ms, now ~39ms
execution/topological.py	Successor-based incremental ready-node tracking	Avoids scanning all edges on every node completion; large_100 3x faster
execution/parallel.py	Dispatch all initially-ready nodes	Was only dispatching input_node first

Results

Metric	v1 End	v2 End	Improvement
total_time_s	0.4151	0.2432	41% faster
peak_memory_mb	1.06	0.32	70% less memory
correctness	16/16	16/16	Maintained

Per-benchmark highlights (topological):

Benchmark	v1	v2	Speedup
foreach_expand	~100ms	39ms	2.6x
large_100	~91ms	31ms	2.9x
retry_chain	~27ms	19ms	1.4x

Test plan

• All 413 existing tests pass
• Benchmark correctness: 16/16 across all workflows and both executors
• No new dependencies added
• No test files modified

🤖 Generated with Claude Code