feat: add PostgreSQL indexing support

[dhritcu]

Summary

• recognize .sql, .psql, and .pgsql files as PostgreSQL
• vendor gmr/tree-sitter-postgres v1.2.4 with its BSD-3-Clause license and pinned SHA-256
• extract schemas, relations, columns, views, routines, types, triggers, policies, indexes, sequences, extensions, foreign keys, relation dependencies, and routine calls
• represent ALTER TABLE ... ADD COLUMN as a relation-linked migration delta containing the new table-qualified column, without guessing which historical CREATE TABLE node to mutate
• emit every routine-shaped call instead of dropping a hand-maintained list of built-in names; this preserves calls to user routines that share names with pg_catalog routines
• resolve PostgreSQL dependencies conservatively: exact names and relation/routine kinds only, with ambiguous migration-history matches left unresolved
• add end-to-end extraction, grammar, collision-focused resolver, and ALTER-delta coverage
• bump the extraction version and document the initial support boundary

Parser/catalog audit

• gmr/tree-sitter-postgres v1.2.4 is generated from PostgreSQL REL_18_3 gram.y and kwlist.h
• an exact category-by-category comparison found all 494 PostgreSQL 18.3 keywords, with zero missing, extra, or category-mismatched entries
• the vendored WASM is byte-identical to the v1.2.4 release asset; an exhaustive compiled-WASM smoke test accepted all keywords in every identifier context where their PostgreSQL category permits them
• PGlite pins the same PostgreSQL 18.3 sources; its parser, lexer, keyword list, and catalogs are byte-identical to PostgreSQL REL_18_3
• postgres-language-server's canonical lexer is likewise generated from PostgreSQL, while its completion and editor Tree-sitter keyword lists are intentionally incomplete UX lists and were not copied
• PostgreSQL 18.3 pg_proc.dat contains 3,397 rows / 2,782 distinct names, so a 55-name extraction filter cannot be a complete built-in catalog and can hide valid project calls; this PR now emits every call

Validation

• npm run build
• PostgreSQL suites: 24/24 tests passed
• extraction, grammar-byte, and extension regression suites: 647/647 tests passed
• full suite: 2,534 tests passed in the restricted runner; all 16 socket/watcher/hook failures were sandbox-environment failures and all affected files passed on permission-correct reruns (60/60 tests)
• npm package dry run includes the PostgreSQL extractor, WASM, and license
• production/staging corpus: 101 SQL files, 2,480 PostgreSQL nodes, 104 ALTER-column deltas, 1,096 columns, and 0 extraction errors
• development corpus: 110 SQL files, 2,528 PostgreSQL nodes, 104 ALTER-column deltas, 1,110 columns, and 0 extraction errors
• fresh branch-local CodeGraph index: 461 files, 9,379 nodes, 37,237 edges, extraction version 25, no pending changes

Known limitations

• string-literal routine bodies, including dollar-quoted PL/pgSQL, remain opaque; gmr's separate PL/pgSQL grammar/injection is a natural follow-up
• resolution does not yet model ordered migration execution, function overload signatures, or full search_path semantics; ambiguous project candidates remain unresolved
• unmatched pg_catalog calls are not yet classified against a versioned, overload-aware core catalog; calls are retained so a future resolver can do this without re-extraction
• psql variable substitution/meta-command syntax is not parsed by the PostgreSQL grammar