fix(pg_introspect): emit FK DDL before function stubs so unparseable routines can't drop references edges (#1854)

[sekmur]

Problem

The dropped-FK half of #1854 (closed as fixed-by-#1746, but reproducible on a fresh 0.9.16 — graphify/pg_introspect.py and graphify/extractors/sql.py are byte-identical from v0.9.14 through v0.9.16).

introspect_postgres() reconstructs one synthetic DDL document in the order tables → views → functions → FKs and hands it to extract_sql() in a single tree-sitter parse. A C-language routine's routine_definition is just the C symbol name, so its stub:

CREATE FUNCTION "public"."levenshtein"() RETURNS void AS $gfx$ levenshtein $gfx$ LANGUAGE c;

is unparseable, and tree-sitter's error recovery consumes the statements after it. With the FK ALTER TABLEs emitted last, they're what gets eaten — so any DB with a common extension installed (uuid-ossp, pgcrypto, pg_trgm, fuzzystrmatch, unaccent, PostGIS, …) silently loses its references edges. This is distinct from #1746 (FK reading privileges): the failure is downstream of a correct catalog read, in the parse.

Fix

Emit the FK block before the routine DDL — a pure block move. FK statements only reference tables, which are emitted first, so the order is always safe; an unparseable stub can now only damage what follows it (the remaining stubs, which contribute no edges from their bodies anyway).

Measured on a real DB (35 tables / 41 FKs / 41 extension routines from fuzzystrmatch + pg_trgm): 1/41 references edges before, 41/41 after.

Test

test_pg_introspect_fk_edges_survive_unparseable_function_stubs — 7 tables, 6 FKs, one LANGUAGE c routine + one plpgsql routine, asserts all 6 references edges survive. Fails on current v8 (0 or partial edges, statement-mix-dependent), passes with the fix. The existing baseline test never caught this because its mock routines only use SQL/PLPGSQL languages, never C.

Longer term this path may not need a parse at all — introspect_postgres() already holds tables and FK pairs as structured rows from pg_catalog, so emitting nodes/edges directly (instead of round-tripping through synthetic DDL) would make this failure class impossible. This PR is the minimal ordering fix; happy to discuss the direct-emission approach separately.

🤖 Generated with Claude Code