plan: tier2 incremental kuzu rebuild

[HumanBean17]

Summary

• Add implementation plan for incremental Kuzu graph rebuild (plans/active/PLAN-TIER2-INCREMENTAL-REBUILD.md)
• Four-PR sequence: foundation + deps index (T1), symmetric delete helpers (T2), incremental orchestrator (T3), CLI + decision engine (T4)
• Derived from propose/active/TIER2-INCREMENTAL-REBUILD-PROPOSE.md; reviewed for codebase accuracy (file paths, function signatures, Kuzu cascade semantics, _write_meta correctness in incremental mode)
• 52 named tests across all PRs; companion proposal INDEX-AUTO-MODE-PROPOSE.md for decision engine spec

Plan highlights

• PR-T1: FileDeps dataclass, .deps.json sidecar read/write, determinism test, perf baseline
• PR-T2: 12 symmetric delete_*_for_file helpers (Kuzu doesn't cascade — explicit DECLARES + member cleanup)
• PR-T3: build_ast_graph_incremental orchestrator, closure expansion (8 rules), per-pass subset functions, 21 tests including fixture equivalence and transaction rollback
• PR-T4: refresh_decision.py decision engine, CLI _cmd_increment integration, new refresh_code_index MCP tool

Key design decisions

• _write_meta in incremental mode queries live Kuzu DB for global stats (partial accumulator would produce wrong counts)
• Pass6 always reruns globally (fast, spans services)
• pass5 subset has no asts parameter (parent does del asts)
• refresh_code_index MCP tool does not exist yet — created in PR-T4
• Brownfield changes force full rebuild (PR-T5 optional refinement deferred)

Test plan

• Review plan for completeness against proposal §2-§7
• Verify file paths and function signatures match codebase
• Confirm test fixture strategy (Tier 3 per-test builds) is correct
• Validate cross-PR risk mitigations are adequate

🤖 Generated with Claude Code