feat(linter): flat rule ids and static type checking

[psteinroe]

Restructures the linter around flat rule IDs and replaces the EXPLAIN-based typecheck with static type checking backed by an in-memory catalog. The linter infers the type of every expression and reports what Postgres would reject, without running the statement. EXPLAIN stays as a fallback. Existing configs and suppression comments keep working and print deprecation warnings. Completions and hover now see the tables and columns created earlier in the file.

This combines the former stack of #815, #819 and #823 into one PR.

Flat rule IDs and groups

Rules are identified by name: lint/banDropColumn, linter.rules.banDropColumn, pgls-ignore banDropColumn. Groups are metadata, configured in bulk under linter.groups, like oxlint's categories. Moving a rule to another group no longer breaks configs or suppressions. The rules were regrouped into correctness, safety, destructive, style, typecheck, and nursery.

// before
{ "linter": { "rules": { "safety": { "banDropColumn": "off" } } } }
// after
{ "linter": { "rules": { "banDropColumn": "off" }, "groups": { "style": "off" } } }

Precedence is rule > deprecated linter.rules.safety.<rule> > group > recommended/all preset. preferBigintOverInt and preferBigintOverSmallint are merged into preferBigInt with the options checkInt and checkSmallint. concurrentRefreshMatviewLock is removed; requireConcurrentRefreshMatview covers it, and the old name maps to it.

Backward compatibility and warnings

Everything users write keeps working:

• Config: linter.rules.safety.<rule>, safety.recommended, and safety.all still apply, including the removed rule names, which map to preferBigInt. The CLI prints one Warning: line per deprecated setting before it runs. The LSP shows a single warning message per session that lists them. The JSON schema marks linter.rules.safety as deprecated, so editors flag it in the config file.
• Suppressions: lint/safety/<rule>, lint/<group>, and the removed rule names still suppress; a removed rule stands for the rule that replaced it. Each such comment gets a warning diagnostic that names the flat form, like Use `banDropColumn` instead. lint/safety keeps its former meaning of every lint rule. safety is today's group.
• Rule selectors: the only and skip lists of the workspace API accept the old IDs and the removed rule names. safety and lint/safety still select every lint rule.
• API: The generated TypeScript types keep their names, including Safety and SchemaCache. So do the JSON formats, the schema export command, and the invalidateSchemaCache command.

What changes with an unchanged setup:

• Diagnostic categories: lint/safety/<rule> becomes lint/<rule>. Scripts that match on the LSP diagnostic code or the JSON, GitHub, GitLab, or JUnit reporter output need updating.
• linter.enabled: false now disables the linter. It used to be ignored.
• migrations.migrationsDir: when set, migration-only rules (most of safety and destructive) are skipped for files outside it.

Catalog and name resolution

The new crate pgls_catalog merges pgls_schema_cache and adds:

• a catalog that applies the file's DDL to the database snapshot: temp tables, CTAS and view columns, inheritance, renames, drops, and ROLLBACK / savepoints;
• the session state: search path, role, transactions, locks, check_function_bodies;
• a conservative name resolver.

No DDL runs against the database. Anything the catalog can't be sure about stays silent, including statements that don't parse and may have changed the catalog. EXPLAIN remains as a fallback for statements that only reference unchanged database objects. typecheck.enabled switches both off. The unused pgls_type_resolver crate is deleted.

Typecheck rules

All are recommended and report errors.

Rule	Reports	Postgres error
unknownRelation, unknownColumn, unknownSchema, unknownFunction, unknownType	references to objects that don't exist	42P01, 42703, 3F000, 42883, 42704
ambiguousColumn	a column of several FROM items	42702
insertColumnMismatch	more values than target columns, or the other way round	42601
missingFromClauseEntry	t.col without t in FROM	42P01
operatorTypeMismatch	1 + now(), '1' + '1'	42883, 42725
functionArgumentMismatch	length(1), ambiguous overloads	42883, 42725
invalidCast	now()::int	42846
assignmentTypeMismatch	insert into t (qty) values (now()), also UPDATE and ON CONFLICT	42804
functionReturnTypeMismatch	LANGUAGE sql functions whose final statement returns the wrong number or types of columns	42P13

How typing works

The snapshot now loads casts, operators, and typing metadata for types and functions. pgls_catalog::typing ports the algorithms of the latest supported Postgres (18, pinned to REL_18_6) for coercion, common types, polymorphic types, and function and operator overload selection. Each port links the Postgres function it follows at that tag. Postgres 15 to 17 resolve types the same way; where 18 accepts more (old/new in RETURNING), it is gated on the server version. The resolver types every clause Postgres types. Anything it doesn't model is unknown, and unknown never reports. AGENTS.md describes how to port and test the type rules.

No false positives

Two tests compare the analysis with Postgres:

• resolve/differential_tests.rs checks a corpus of statements against a live Postgres on every run. Statements Postgres accepts must give no findings, inferred column types must match what Postgres describes, and expected errors must be found.
• pgls_analyser/tests/postgres_regress.rs runs Postgres' own regression SQL for 15, 16, 17 and 18 (the fixtures from test: add Postgres regression fixtures for every supported version #824, about 170,000 statements) through the linter, against the catalog of a fresh database on each version. It is a normal cargo test without network or database and takes about 11 seconds. Any finding on a statement Postgres accepted fails it. The findings on statements Postgres rejected are snapshotted per version, so changes in detection show up in review. just record-regress <major> --catalog-only re-records a catalog.

It found 644 false positives on Postgres 15 and 426 more on 16 to 18; all are fixed. One of them affected the editor beyond typing: statements that don't parse were dropped before linting, so a CREATE TABLE with newer syntax made every later use of the table report a missing relation. Such statements now make missing objects unknown for the rest of the file, unless they can't change the catalog (queries, COPY, maintenance, ...).

Schema cache

The SchemaCache JSON gains casts, operators, and typing fields on types and functions. All of them are optional, so older exports still load.

Review focus: false positives in pgls_catalog/src/resolve, catalog/ddl and typing. The known gaps are listed in the module docs of catalog/ddl, resolve and typing. Statements with psql identifier parameters (:schema.table) are not typechecked.

Fixes #624
Fixes #692
Fixes #369
Fixes #431