feat!: render through a QueryRenderer, and stop calling it a rewrite

[iamralch]

Started as "let a caller say where placeholder numbering starts" and grew, in review, into the API change that was actually needed. Two problems, one fix.

Nothing was being rewritten. A rewrite takes a statement and returns a changed statement. This takes typed values — a compiled filter, a parsed ordering, a decoded token — and produces SQL text plus binds. That is rendering. The word came from pgxaip, which took it from pgx's QueryRewriter hook, where something genuinely is rewritten; the inheritance stops here. render is now the verb in the public API and in the cursor and order modules behind it.

The arguments mixed the request with the schema. Six values reached a call site — filter, order_by, page_token, columns, dialect, param_offset — split evenly between what changes per request and what is fixed for a table in a deployment.

const VOLUMES: QueryRenderer<'static, dialect::Postgres> =
    QueryRenderer::new(dialect::Postgres).columns(VOLUME_COLUMNS);

let fragment = VOLUMES.render(&query)?;       // the whole request
let fragment = VOLUMES.at(3).render(&query)?; // ... spliced after $1, $2

new is const and the renderer is Copy, so it costs an entry in the binary and nothing at runtime — the dialects are zero-sized and Columns borrows.

The asymmetry is deliberate

The dialect is required by new; the column map is a builder method. That looks inconsistent until you compare failure modes. A renderer with no column map rejects every path — fail-closed, loud, on the first request. A renderer with a guessed dialect is not an error at all: it's $1 placeholders sent to MySQL, valid SQL for the wrong database. The one that cannot fail safely is required; the one that can is optional.

Knock-on

Query loses columns and its lifetime, leaving exactly the three fields protoc-gen-rust-aip generates onto a request — so copying one into the other is field-for-field with nothing left over, which was always the intent.

rewrite, rewrite_with and the Options re-export are gone; at replaces the last of them. Options is still right for sqlx-cel, which has one knob and no required argument to pair it with.

Breaking, deliberately, while nothing is published.

Testing

Against a real Postgres this time — not the CI-only path — plus the full feature sweep CI runs:

• cargo test --features sqlite,mysql with DATABASE_URL set: 59 passed, including the 6 end-to-end Postgres tests and the 6 SQLite ones
• all 7 feature combinations from the CI matrix: clean
• the negative case (--no-default-features --features postgres) still correctly refuses to build
• cargo fmt --check, cargo clippy --all-targets, cargo doc with -D warnings: clean