[devin-ai-integration]

Summary

Fixes the 19 bugs found by end-to-end testing of the published v0.13.0 WASM against real PostgreSQL 16, MySQL 8.3 and SQLite (Ruby 3.1 – 3.4).

MySQL (mysql2)

• A sqlc.arg used twice generated def f(term:, term:) (syntax error, the whole file failed to load): now one keyword, sent twice. Repeated ? of one column become id:, id_2:, id_3:.
• Unaliased expressions (SELECT count(*)) were always nil: rows are now read by position (as: :array).
• :copyfrom stored the string "NULL" as SQL NULL: values are quoted, only nil becomes \N. The CSV is built without the csv gem, so the generated Gemfile no longer needs it.

SQLite (sqlite3)

• Types with modifiers (VARCHAR(255), DECIMAL(10,2)) aborted generation: Base#column_type strips the modifier, and its prefix check was reversed.
• :exec / :execrows with RETURNING never ran, or only touched the first row: the result is now drained.

PostgreSQL (pg)

• bytea containing NUL or non-UTF-8 bytes failed, both as a parameter and in COPY: now encoded with PG::TextEncoder::Bytea.
• JSON parameters holding a Ruby Array were sent as a PG array literal: now PG::TextEncoder::JSON.
• enum[] / uuid[] columns came back as the raw "{a,b}" string: they are decoded into arrays by a PG::TypeMapByOid built from pg_type.
• :execlastid crashed when RETURNING produced no row: it now returns nil.
• After PG::Connection#reset, every prepared query failed permanently: prepared state is now tracked per backend PID, so statements are prepared again.

All drivers

• Leaving transaction with break / return / throw neither committed nor rolled back, so the pooled connection kept an open transaction. It now commits, and rolls back on an exception or a killed thread.
• :copyfrom didn't quote identifiers (a column named order broke it): now quoted per driver.
• A casted sqlc.narg(x)::text became a required keyword: it now defaults to nil, as the README says.
• An expression sqlc can't type (any, e.g. coalesce(max(n), 0)) aborted generation: it is now untyped. Type errors also name the query.
• Name collisions:
  • Parameters named like generated locals (client, sql, ...) are renamed (client_:).
  • Columns named like row methods (hash) are renamed (hash_), and columns that aren't valid identifiers are snake_cased.
  • A constant or method defined twice is now a generation error.
  • Lowercase query names now get a valid ...Row class.
• Ruby 3.1: rows are keyword_init Structs (Row.new(name: ...) used to put the whole hash into the first field), RBS uses ::Struct[untyped] with keyword initialize / self.new, and long row constructors wrap, so the output is RuboCop-clean.
• Nested moduleName (MyApp::Db) is emitted as nested module blocks, so the files load without MyApp being defined first.
• rubyVersion: "3.4" is accepted. Unknown option keys, and preparedStatements with a driver other than pg, are now errors instead of being silently ignored.

Tests

• Added a regressions table and queries to examples/authors/{postgresql,mysql,sqlite}; examples/ and the test/fixtures/*.bin captures are regenerated from them.
• New E2E cases in tests/end2end_{pg,mysql2,sqlite3}.rb cover each runtime bug.
• New unit tests in test/features_test.rb cover the generation bugs.
• test/options_test.rb is updated for 3.4.
• Goldens and steep_expectations.yml are regenerated; the new expected diagnostics are the PG:: constants used by the array decoder.

Docs (README options table and semantics, docs/Design.md, CLAUDE.md) now say Ruby 3.1–3.4 and describe the new validation and naming rules.

Local results:

• rake test: 62 runs, 0 failures
• RuboCop and Steep on examples/: clean
• E2E: PostgreSQL 41/41, MySQL 36/36, SQLite 28/28
• scripts/tests/run_codegen.rb passes on Ruby 3.1 and 3.4

Worth reviewing:

• The pg connection setup now runs one pg_type query, to build the array decoder.
• Generated code for MySQL rows changed from hash access to positional access.

Checklist

• examples/ is regenerated (rake sqlc_generate_process) and committed
• Tests added or updated (rake test, and rake end2end_* when the generated code changes)
• The generated code stays RuboCop-clean (rake rubocop) and Steep-clean (rake steep)
• Labelled (feature, bug, breaking, documentation) so the release notes are grouped correctly

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

• Disable automatic comment, CI, and merge conflict monitoring