[psteinroe]

Stacked on #976.

Completions inserted names exactly as they are stored. Postgres downcases unquoted identifiers and rejects most keywords as names, so the inserted text often meant a different name:

create table "MyTable" ("Id" int, "select" int);
create function "My Schema".helper() returns int ...;

select | from "MyTable"      -- before: Id, select   now: "Id", "select"
select * from |              -- before: MyTable      now: "MyTable"
select |                     -- before: My Schema.helper()   now: "My Schema".helper()
select | from users          -- before: pg_catalog.RI_FKey_cascade_del()   now: pg_catalog."RI_FKey_cascade_del"()

What changed

• pgls_catalog::quote_identifier(name, position) and needs_quotes, next to identifier_name, of which they are the inverse. A name stays unquoted only if it's an identifier by scan.l's rules that downcasing leaves unchanged, and if it isn't a keyword the grammar rejects at that position (name classification in gram.y):
  • ColId (the first part of a table, column or schema name) rejects reserved keywords and those that name types or functions.
  • type_function_name (an unqualified function) rejects reserved keywords and those that name columns.
  • After a dot, any keyword is fine.
  • Otherwise the name is quoted, with " escaped as "". This is more precise than Postgres' own quote_identifier, which quotes every keyword that isn't unreserved.
• The keyword categories come from pgls_lexer, which already generates them from kwlist.h. It now has a public keyword_category(word), and pgls_catalog depends on pgls_lexer.
• with_schema_or_alias takes the position of the item:
  • Without quotes, the prefix is quoted as a ColId and the item after the dot as a label.
  • After an opening quote ("| or "|"), only " is escaped, and no second quote is added.
  • Aliases are passed in as names, so from users "pr" now inserts pr.email instead of "pr".email. Both refer to the same alias.
• Schemas now set a completion text, so they're quoted too. After an opening quote, they also get the closing one: select * from "| inserts private".

Snapshot changes

• pg_catalog.RI_FKey_cascade_del() is now pg_catalog."RI_FKey_cascade_del"().
• After an opening quote, schemas get their closing quote.
• "pr".email is now pr.email.

Tests: unit tests for needs_quotes/quote_identifier, covering the four keyword categories, uppercase, spaces, ", a leading digit, $, non-ASCII, and the round trip through identifier_name. A completions test covers tables, schemas, schema-qualified tables, the opening-quote case, columns (keywords quoted alone but not after a dot), and functions with and without a schema. pgls_catalog, pgls_lexer, pgls_completions, pgls_workspace --lib and clippy pass.

Not done: the pretty printer keeps its own needs_quoting in pgls_pretty_print/src/nodes/string.rs. It uses its codegen's keyword table, and sharing the predicate would mean a new dependency for pgls_pretty_print and changes in string.rs while its PR stack is open. It can switch to pgls_catalog::needs_quotes, or the predicate can move into pgls_lexer, once that stack has landed. Completing while typing inside an unterminated quote ("My|) still returns nothing, as on main. Policies and roles are still inserted unquoted.

Update: Merged main, now that #976 has landed (the merge only takes main's version of the #976 changes; this PR's own diff is unchanged). cargo clippy --workspace --all-targets -- -D warnings and the tests of pgls_catalog, pgls_lexer and pgls_completions pass on the merge. #998 is stacked on this PR and switches expand star to quote_identifier.