perf(graph): bulk COPY FROM for _write_edges (PR-P1)

[HumanBean17]

Scope

Implements PR-P1 from plans/active/PLAN-INIT-INCREMENT-PERF.md: bulk COPY FROM for _write_edges (shared helper → accelerates both full and incremental paths).

What Changed

• Added _bulk_copy(conn, table_name, columns, rows) helper using in-memory pyarrow COPY FROM (step-1 spike result in docstring).
• Added column-order constants (_REL_*_COLUMNS, _UNRESOLVED_CALL_SITE_COLUMNS) matching _SCHEMA_* declarations; REL tables list FROM/TO first.
• Converted _write_edges to per-type row staging with bulk load:
  • CALLS dedup by (src, dst, argc, line) applied at staging.
  • callee_declaring_role materialized at staging via _callee_declaring_role_at_write.
  • UnresolvedCallSite nodes loaded before UNRESOLVED_AT edges (kuzu validates endpoint existence).
  • One _bulk_copy call per edge type (EXTENDS, IMPLEMENTS, INJECTS, DECLARES, OVERRIDES, CALLS, UNRESOLVED_AT).
• Deleted dead _CREATE_EXT/_CREATE_IMPL/_CREATE_INJ/_CREATE_DECL/_CREATE_OVERRIDES/_CREATE_CALL constants and local _CREATE_UNRESOLVED/_CREATE_UNRESOLVED_AT (removed with rewrite).
• Added tests/fixtures/graph_baseline_bank_chat.json fixture (node/edge counts, GraphMeta, sampled edge properties).
• Added four tests verifying equivalence, determinism, CALLS dedup preservation, and empty-REL-table no-op.

Out of Scope (PR-P2)

Nodes (_write_nodes_impl), routes/clients/producers (_write_routes_and_exposes, _write_clients_producers_and_calls), and GraphMeta (_write_meta) remain unchanged.

Tests

New tests (all pass):

• test_bulk_write_edges_match_per_row_baseline — bulk write matches baseline fixture.
• test_bulk_write_is_deterministic_double_build — two builds produce identical graphs.
• test_bulk_write_preserves_calls_dedup_and_callee_declaring_role — CALLS dedup and role preserved.
• test_bulk_write_empty_rel_table_is_noop — empty edge list is a no-op.

Regression (56 passed):

• Full test_ast_graph_build.py suite passes unchanged.
• Full test_incremental_graph.py suite passes (2 pre-existing failures in PR-P2 scope).
• test_bank_chat_brownfield_integration.py + test_call_edges_e2e.py pass unchanged.

Ruff: .venv/bin/ruff check . clean.

Sentinel Greps

All sentinel greps pass (from plan):

Must return zero (dead constants deleted):

grep -nE "_CREATE_(EXT|IMPL|INJ|DECL|OVERRIDES|CALL|UNRESOLVED|UNRESOLVED_AT)\b" build_ast_graph.py
# (returns empty)

Must be non-zero (retained for PR-P2):

grep -n "_MERGE_SYMBOL\b" build_ast_graph.py           # 2 hits (node upsert, PR-P2)
grep -n "_CREATE_CLIENT\b" build_ast_graph.py          # 3 hits (routes/clients, PR-P2)
grep -n "MERGE (r:Route" build_ast_graph.py            # 1 hit (pass5/6 dedup, kept)
grep -n 'COPY.*FROM.*rows' build_ast_graph.py          # 1 hit (bulk path present)

Manual Evidence

Build of tests/bank-chat-system (graph-write phase timing):

[graph] writing · nodes written in 5.11s
[graph] writing · edges written in 0.20s
[graph] writing · routes/exposes written in 0.44s

Counts (from java-codebase-rag meta):

• Node count: 959
• Edge counts: EXTENDS 18, IMPLEMENTS 21, INJECTS 94, DECLARES 606, OVERRIDES 38, CALLS 678, UNRESOLVED_AT 227

The bulk edges written in 0.20s (vs. ~250s baseline from per-row writes in the plan) demonstrates the speedup. Counts match the baseline fixture exactly.

Co-Authored-By: Claude noreply@anthropic.com