Add PostgreSQL semantic graph support

[dhritcu]

What changed

• add PostgreSQL parsing and semantic extraction using the vendored gmr/tree-sitter-postgres SQL grammar and its companion PL/pgSQL grammar
• preserve PostgreSQL quoted-name boundaries, case folding, object-class intent, search_path, transaction, savepoint, and temporary-relation semantics
• preprocess psql commands, variables, and COPY FROM STDIN payloads without changing source offsets, and report recovered parse errors
• extract static SQL and PL/pgSQL routine-body calls and table, type, and sequence references while honoring routine-local SET search_path
• model schemas, tables, altered columns, views, routines, types and enum mutations, triggers, policies, indexes, sequences, extensions, constraints, partitions, inheritance, and LIKE relationships
• synthesize migration-ordered table/view drop, re-create, rename, constraint-rename, enum, foreign-key, and schema-containment relationships
• keep strict ambiguity handling while selecting the correct preceding declaration at lifecycle boundaries
• surface related tables, foreign keys, triggers, routines, and renamed constraints in node trails
• bump the extraction version so existing indexes are prompted to rebuild

Why

The initial SQL support could locate declarations, but a useful database graph must answer relational questions without inventing dependencies. Looking up a table should expose its foreign-key neighbors, triggers, routines, and migration history. PostgreSQL naming, local search_path, psql input, repeated historical definitions, drops, re-creations, and renames all need ordered handling for those trails to stay accurate.

Validation

• npm run build
• PostgreSQL suite: 12 files, 108 tests passed
• GIT_CONFIG_GLOBAL=/dev/null npm test
  • 157 test files passed; 15 skipped
  • 2,645 tests passed; 175 skipped
• fresh Eversa development index using /private/tmp/codegraph-source/dist/bin/codegraph.js from this branch:
  • 7,553 files, 80,000 nodes, 257,404 edges
  • 110 PostgreSQL files, 3,096 PostgreSQL nodes, zero PostgreSQL parse errors
  • 148 foreign-key facts projected into 133 live table-to-table edges
  • all 7 altered enum values attached to their effective enum definitions
  • public.users exposes its FK neighbors, 6 triggers, routine/view/index references, and constraint names
  • renamed public.tags keeps its public.apps FK under tags_app_slug_fkey and its incoming public.tag_assignments FK
  • full index: 8.8s internal index phase, 26.18s process wall time, 1.74 GB peak RSS
  • no-change sync: 0.78s wall time, 133 MB peak RSS
  • 320.4 MB graph database; extraction version 27; zero pending changes or references

Intentional limits

• dynamic SQL assembled through PL/pgSQL EXECUTE is skipped; only statically identifiable routine-body SQL is linked
• migration chronology is inferred lexically within a directory; separate migration directories remain independent streams
• server catalog contents are not indexed, so implicit pg_catalog lookup and external Supabase/auth/storage objects can remain unresolved
• historical declarations and enum labels remain queryable; explicit rename transitions and synthesized live relationships carry current-state meaning