feat(types): refuse wrong-dialect functions and missing sources at the type level - #106

Db.run and compile find a function of the wrong dialect only when the query compiles, as a defect. A column of a source that is not in the query's FROM or JOIN is found only by the server. Both are now type errors, with a message that names the cause.

Closes #66.
Part of #52.

Depends on #97; review only the commits after it.

Changes

• New DialectEntry<Set>. It rides in the params channel that a query already carries. Each built-in function adds ClickHouseOnly or PostgresOnly to its result. A portable function (nullIf, /clickhouse's lower) and the shared operators add nothing. The tag follows the same set that assertSet checks at run time.
• Dialect<Functions>: clickhouseDialect is Dialect<"clickhouse">, postgresDialect is Dialect<"postgres">. A plain Dialect is not checked.
• compile, compileUnsafe, compileUnion and compileUnionUnsafe check DialectSatisfied<Params, Functions>. The target is options.dialect, else the entry's dialect. New types Compile<Default>, CompileUnsafe<Default>, CompileUnion<Default>, CompileUnionUnsafe<Default> and the options types QueryCompileOptions and UnionCompileOptions, so /postgres gets a real "postgres" default (it used typeof compileCH before). The query overload is now the last one, so a refused query shows its own reason (wrongDialect, paramsRequired) and not the write overload's.
• fromSqlClient returns DatabaseApi<Functions>. With postgresDialect, run refuses ClickHouse functions in the type. DatabaseApi<"postgres"> still provides the Database tag.
• A query with functions of two dialects is refused everywhere (mixedDialects), the Database in context and custom dialects included.
• New SourceEntry<Name>. A column ref carries its source: the FROM alias, else the table name, or the join alias. CHQuery gets a sixth type parameter, Source (default never). Each query method removes its own sources (BindSources). An entry that is left names a missing source, and compile and run refuse it (sourceNotInQuery). A later JOIN of that name also provides it.
• Correlation: an inner query can use the outer query's accessor (e.Name.eq($.Name), which writes s.Name). The inner query carries s on, and the outer query provides it. Docs show this as a checked alternative to a string outerRef.
• Fix: orderBy(($) => [...]) carried only the first sort key's params. A param in a later key was not required. It now carries all of them.
• The runtime QueryBuilderDefect for a wrong-dialect function stays as the safety net. soundness.test.ts keeps testing it under @ts-expect-error.
• Tests: src/ch/dialect.test-d.ts and src/ch/sources.test-d.ts (@ts-expect-error for each refused case, expectTypeOf for each message), an orderBy case in params-propagation.test-d.ts, and a runtime docs test for the correlated accessor.
• Docs: new section "Checked in the type" in docs/params-and-compilation.md, plus docs/reference.md, docs/postgres.md, docs/database.md, docs/joins-and-subqueries.md and design/dialects.md. One minor changeset.

Left out

• A CompiledQuery does not carry its dialect in the type. pg.run(CH.compileUnsafe(q)) is still refused only at run time (DialectMismatch). Adding it changes CompiledQuery for every caller.
• The Database tag stays untyped (DatabaseApi<string>), so Db.run checks mixing only. A typed tag per dialect is an API decision for the maintainers.
• CurrentDialect is read at run time, so the type cannot see it. Under a CurrentDialect of another dialect, pass options.dialect too.
• No runtime check for a missing source. The server still reports it, and a captured ref to a table of the same name in a subquery still binds to the inner table, as SQL does.
• Writes (update, deleteFrom, insertInto) do not tag their own accessors. A ref from a select query used in a write is still refused.
• A source whose name the type does not know (Table<string, …>, a non-literal alias) is not checked.
• Effect on existing code: a ClickHouse function's result type is now Expr<T, ClickHouseOnly>. Expr<T> annotations still accept it, but toEqualTypeOf<Expr<T>> assertions change. Thirteen such assertions in three test files of this repo are updated.
• No literal, non-empty check for derived-source aliases (effect-qb's third point). FreshAlias already covers clashes.
• MySQL, SQLite and D1 are not part of this PR. A new dialect only needs Dialect<"its-set"> and builtins("its-set", …).

Validation

• bun run build: pass.
• bun run typecheck: pass.
• bunx vitest run --testTimeout=60000 --hookTimeout=60000 with ClickHouse 26.8.2.7 at 127.0.0.1:8123 (EFFECT_ORM_CLICKHOUSE_* set): 60 files pass, 0 skipped; 1040 tests pass.
• node scripts/check-doc-citations.mjs: ok (25 files). node scripts/check-exports-documented.mjs: ok (303 /clickhouse, 221 /postgres). node scripts/check-doc-examples.mjs: ok (21 files).
• bun scripts/check-package.ts: pass (tarball imports, types, CommonJS consumer).
• Upstream CI for a fork PR waits for a maintainer to approve the run.