fix(wire): honor binary result-format codes (Power BI/Tableau preview ERROR)

[lfrancke]

Stacked on #8. This PR is based on chore/update-dependencies because it needs the same cargo-deny advisory fixes (drop rustls-pemfile, bump tokio-postgres) to pass CI. Review/merge #8 first; GitHub will retarget this PR to main automatically once #8 lands. The diff shown here is only the binary-wire-format change.

What

Honor the client-requested binary result-column format in the extended-query protocol. Previously the gateway always emitted text-format DataRows, so clients that bind result columns for binary (Power BI's Npgsql typed path, Tableau/pgjdbc, tokio-postgres) misdecoded everything except varchar.

Why

A customer reported the Power BI Desktop Navigator preview showing ERROR on Float / Integer / Timestamp columns while string columns rendered, and the full data load worked. Root cause: binary-decoding our text bytes fails for every type except the string family (whose binary and text wire forms are byte-identical). Real PostgreSQL honors the Bind result-format request and sends binary — the one concrete behavioural difference. This was originally mis-filed as an upstream pgwire gap; pgwire already supports binary output, so the fix is ours.

How

• types::encode_cell dispatches per column on the requested FieldFormat, converting Trino JSON values to typed Rust values so pgwire's DataRowEncoder emits PostgreSQL binary. Supported: bool, int2/4/8, float4/8, numeric, date, time, timestamp (no tz); the string family passes through as raw bytes. SQL NULL handled once (format-independent).
• Fails closed (SQLSTATE 0A000) for types without a faithful binary encoder rather than emitting bytes the client would misread — we cannot silently downgrade to text because the client decodes strictly per the format it bound.
• trino_stream::build_pg_schema takes the portal's result_column_format, threaded through process_query → execute_trino_query. The schema drives both the RowDescription and the per-cell encoding, so they cannot diverge. Simple-query path stays text.

Scope / follow-ups (not in this PR)

• The static catalog/intercept path (pg_type etc.) remains text-only. Type-loading drivers request text there, so it's not user-visible; pinned by a regression test.
• timestamptz, interval, uuid, bytea, inet, json/jsonb, arrays fail closed under binary — add encoders as needed.

Testing

• Byte-level unit tests for encode_cell (no Trino).
• Re-enabled the three #[ignore]'d extended-protocol tests + a new test for the customer's exact types (bigint, real, double, timestamp).
• Validated end-to-end against Trino 479 (SDP): the binary tests pass with the fix and fail with error deserializing column when the schema-format line is reverted — confirming they exercise the fix.
• cargo test (257 across 5 suites, with Trino), cargo clippy --all-targets -- -D warnings, cargo fmt --check all clean.

Review notes

• New runtime deps: chrono and rust_decimal (versions unify with what pgwire's default features already pull into postgres-types; rust_decimal uses db-postgres). Both Apache-2.0 / MIT.
• Worth a close look at the per-type conversions in types.rs::encode_binary_cell (range checks, NaN/Infinity-as-string, decimal precision fail-closed, timestamp parse formats).

🤖 Generated with Claude Code