perf: 47% faster JSON parse + graph build via AutoResearch

[jcfischer]

Summary

• 47.6% performance improvement on JSON parse + graph build (5,338ms → 2,797ms on 360MB / 1.68M node real Tana export)
• Ran 7 AutoResearch rounds (Karpathy-style autonomous optimization): 5 kept, 2 discarded
• All 29 parser/graph/indexer tests pass, identical output (same node/supertag/field/tag_application counts)

Key Changes

• Replace Zod per-node validation with fast structural check — Zod was 75% of parse time, validating 1.68M nodes against schema. Top-level structural check + type assertion instead.
• Single-pass graph build — merged 4 separate full iterations over docs into one, reducing from 5 passes to 2 (index build + classification)
• Single iteration with deferred resolution — further merged to 1 pass over docs + tiny candidate resolution pass (~1% of nodes)
• Remove unused defaults loop — inbound_refs, outbound_refs, editMode defaults were set on every node but never consumed
• Use Bun.file().json() — avoids intermediate string allocation

AutoResearch Results (results.tsv)

Commit	Total	Status	Description
af02f5e	5338.8ms	✓ keep	baseline
93fec81	3148.1ms	✓ keep	single-pass graph build
d205f3a	2982.8ms	✓ keep	remove defaults loop
d324b26	3121.9ms	✗ discard	early exit SYS scan
da26963	3195.1ms	✗ discard	trashIds Set
77d485f	2999.4ms	✓ keep	Bun.file().json()
9e5706d	2797.3ms	✓ keep	single iteration + deferred

Test plan

• 29 parser/graph/field-values tests pass on every round
• Output identical: 582 supertags, 1243 fields, 12121 tag applications
• Benchmarked on real 360MB export (1,679,904 nodes)
• Verify supertag sync works end-to-end with real workspace

🤖 Generated with Claude Code