Summary

Follow-up to the init/increment-perf plan (PR-P1/P2/P3, #340–#342). A review fan-out over the landed code found that PR-P2 introduced a latent data-correctness regression: converting _write_nodes_impl to bulk COPY FROM replaced the per-row _MERGE_SYMBOL upsert with a skip-if-exists filter, which silently dropped the property refresh on preserved dependent type nodes.

A preserved dependent's role/capabilities depend on project-wide inputs (meta-annotation chain, brownfield overrides) and can shift without the dependent's own source changing. The _delete_file_scope contract (issue #305) deliberately preserves dependent nodes and relied on that re-MERGE to refresh them in place. With the skip filter, the role was recomputed and then discarded — so the incremental graph diverged from a full rebuild until the next full rebuild. Roles drive semantic-search ranking, so this silently degraded results. No crash, no node/edge loss, self-heals on full rebuild. CI green — the two tests meant to guard incremental≡full equivalence were broken.

This PR fixes the regression and the tests that let it ship.

How it was found

Fanned out four reviewer subagents over the landed plan. Two analyzed the same code path and disagreed (one said HIGH bug with a repro, one said benign/unreachable). Adjudicated empirically: the "benign" verdict came from a reviewer that couldn't run ladybug in its env (bare python — the system-shadows-venv trap) and over-reasoned; the "bug" verdict was confirmed with a deterministic repro:

initial full (V1)           : Svc.role = OTHER     ✓
increment (V2, in scope)    : Svc.role = OTHER     ← stale
full rebuild (V2, truth)    : Svc.role = SERVICE

Changes

build_ast_graph.py

• TypeIndexEntry / MemberEntry get a loaded_from_db flag set by _load_existing_{types,members}, distinguishing DB-loaded stubs (placeholder decls used only for cross-file resolution) from freshly-parsed scoped/dependent decls.
• _write_nodes_impl: after the bulk COPY of new nodes, re-SET every mutable Symbol field on preserved, non-stub, type-kind nodes (_SET_SYMBOL_BY_ID), restoring the upsert. Stubs are skipped (their stored values are authoritative); non-type kinds carry no mutable role/capabilities; the full path is unaffected (empty DB → no existing nodes → no SETs).
• _populate_declares_rows skips loaded members: their DECLARES edges already persist and re-emitting duplicated them (REL tables carry no PK). This second incremental≠full divergence surfaced once the equivalence test was fixed (incremental=7 vs full=5 DECLARES).
• Delete dead _existing_symbol_ids (zero callers); correct the misleading Route-MERGE loop comment (it iterates all routes, not phantoms).

Tests

• test_incremental_bulk_write_equivalent_to_full_rebuild now asserts real incremental≡full equivalence (node count, per-type edge counts, type roles) instead of the prior tautology (set(edge_type_keys) == set(edge_type_keys) + count > 0, which never failed because every rel type yields a count row even at 0).
• New test_incremental_refreshes_dependent_role_on_meta_chain_change guards the exact regression.
• Rename test_bulk_write_edges_match_per_row_baseline → test_bank_chat_bulk_build_matches_committed_baseline: the per-row path is gone, so the bulk-generated fixture is a drift anchor, not a per-row equivalence proof.

Validation

• .venv/bin/ruff check . — clean
• .venv/bin/python -m pytest tests -q — 848 passed, 14 skipped (HEAVY-gated)
• bank-chat fixture rebuild + meta: ontology_version 17 (unchanged), parse_errors 0, all counters present
• Repro (/tmp/repro_role_staleness2.py): increment now yields SERVICE = full rebuild (no divergence)

No ontology bump, no schema change, no re-index required, no public-surface change.

Known follow-ups (out of scope)

• Converting the remaining per-row Route MERGE to bulk (needs existing-id filtering to reproduce the dedup).
• OVERRIDES may have an analogous cross-file duplication; no test catches it yet.
• Editing an annotation definition does not pull annotation-users into incremental scope (a pre-existing scope-tracking gap, distinct from the upsert regression fixed here).

🤖 Generated with Claude Code