core: carry enums, views and missing placeholders through the analysis core

[kyleconroy]

Three fixes from a triage of every internal/endtoend replay case that passes as committed but fails under SQLCEXPERIMENT=coreanalyzer. Together they take the core context from 347 failing cases to 281, with no case regressing and the base context still fully green. Each commit stands on its own.

Enums absent from models (50 cases, 45 fixed)

The core catalog created an enum type on CREATE TYPE ... AS ENUM but kept none of its labels, and the dump that hands the catalog to codegen carried relations only. Every enum was missing from models.go, PostgreSQL enum columns were typed any and MySQL ENUM columns string.

• sql_enum_label stores an enum's labels, modeled on pg_enum.
• The schema package applies ALTER TYPE ADD VALUE (with BEFORE/AFTER and IF NOT EXISTS), RENAME VALUE, RENAME TO, SET SCHEMA and DROP TYPE.
• A MySQL ENUM or SET column declares an enum of its own, named <table>_<column> as the legacy catalog names it. It follows the column and the table through renames, is replaced on MODIFY, and goes with a dropped column or table.
• Type names keep the schema an engine reports on the TypeName itself and drop a spelled-out public., so foo.mood and DROP TYPE public.status resolve the way their column references do.

The five leftovers are blocked by other buckets (varchar typed any, MySQL BOOLEAN, sqlc.embed).

Views absent from models (10 cases, 8 fixed)

The catalog dump listed kind r relations only, so views, materialized views and CREATE TABLE AS had no model, and a query selecting all of one got a row struct of its own. The listing now covers the relations a query selects rows from the same way. The two leftovers are the anonymous-column naming bucket (SELECT 1 as a view) and CREATE SCHEMA.

Placeholders the analyzer walked past (17 cases, 14 fixed)

• INSERT ... SELECT: the engines report the source query with an empty VALUES list rather than none, so the analyzer took the VALUES branch and walked nothing. A placeholder selected directly now takes the type of the column it lands in.
• LIMIT on UPDATE, DELETE, and on UNION/INTERSECT/EXCEPT was never looked at.
• ON DUPLICATE KEY UPDATE assignments bind the way SET does.
• x IN (SELECT ...) on MySQL arrives wrapped in a sublink, which the IN node now looks through.
• x COLLATE c: SQLite puts the expression in the node's other field, so the placeholder underneath was neither found nor typed.
• CALL: the compiler now sends it through the core, CREATE PROCEDURE is recorded with a void pseudo type as its result, and a call types each placeholder from the declared parameter, by position or by name =>. The AST gains IsProcedure, set by both parsers, since a function with only OUT parameters also declares no return type.
• A placeholder on the left of an IN list takes its type from the members.

The three leftovers are MySQL BOOLEAN and LIMIT parameter naming. I tried naming LIMIT/OFFSET placeholders in the core and backed it out: it changed analyze output for SQLite and ClickHouse, where the database reports no name, so that convention belongs in the compiler bridge rather than the analysis.

Notes for review

• One judgement call: a MySQL ENUM column is now reported by sqlc analyze with the synthetic <table>_<column> type name rather than enum. That is what codegen needs and what the legacy compiler always did, but it is not what MySQL itself reports. goldeneye has no MySQL analyze check today.
• No test cases were added; the existing corpus covers the changes. Verified by running TestReplay/base (all green) and SQLC_TEST_CORE=1 TestReplay/core (347 → 281) before and after each commit.
• internal/core/catalogdb is regenerated from the catalog schema and query changes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01XZcKb4GFiZQbeo9oVmyF3m