Fix Power BI DirectQuery slicers on text, interval, time and timestamptz columns

[adwk67]

What this changes

Issue: a Power BI DirectQuery slicer on a text column failed with Cannot cast varchar(…) to interval day to second. Fixing it exposed three more column types whose slicers silently selected no rows. Driver and connector
must be upgraded together.

Column type	Before	After
varchar	Folded CAST(… AS INTERVAL DAY TO SECOND) → error; unquoted constants	VARCHAR leads SQL_WVARCHAR in SQLGetTypeInfo (core preferred row); connector quotes and escapes text constants
interval …	Driver re-rendered -1-00, 0 00:00:00.5; Power BI's cast(col as VARCHAR) = '<shown>' matched nothing	Delivered as Trino's own text (-1-0, 0 00:00:00.500); SQL_C_INTERVAL_* still works via core
time	Power BI folds against 1899-12-30, Trino casts onto today → 0 rows	Connector reports SQL_CONVERT_TIME without SQL_CVT_TIMESTAMP → visible "couldn't fold" error; README workarounds (view with CAST(col AS VARCHAR), Import mode)
timestamp with time zone	Delivered in UTC; Trino reads the folded literal in the session zone → matched only in UTC sessions	Delivered as wall time in the session zone (SET TIME ZONE > TimeZone= > coordinator default via current_timezone() in the connect probe); time with time zone likewise, at the zone's current offset

Also: rustls 0.23.45 (RUSTSEC-2026-0285); Windows suites now run with UTF-8 output (two suites crashed on cp1252).
Core is pinned to 2f9f725 by a TEMPORARY rev — switch both Cargo.toml entries to the core release tag before
merging.

Spec basis

• SQLGetTypeInfo ordering ("how closely the data type maps"): see the core PR.
• SQL to C: Character → all C interval types: see the core PR. Interval columns are described as SQL_WVARCHAR
  (unchanged), so the character table is the one that governs.
• SQL to C: Time, footnote [c]: "The date fields of the timestamp structure are set to the current date" —
  https://learn.microsoft.com/en-us/sql/odbc/reference/appendixes/sql-to-c-time — Trino's CAST(time AS timestamp)
  follows ODBC; Power BI's 1899-12-30 anchor does not. The SQL_CONVERT_TIME override is therefore a deliberate
  misreport, documented in the connector and pinned by test_folding_contract.py.

Trino behaviour (measured, Trino 483)

• CAST(interval AS VARCHAR): -1-0, 0 00:00:00.500, 1 00:00:00.000.
• CAST(CAST(TIME '14:30:00' AS TIMESTAMP) AS DATE) = current_date → true.
• Berlin session: from_iso8601_timestamp('2025-10-26T00:30:00Z') = CAST('2025-10-26 02:30:00' AS TIMESTAMP) →
  false, …01:30:00Z… → true (Trino picks the later instant of the DST overlap; pinned in the slicer suite, README
  Troubleshooting).
• SET TIME ZONE sets the session property time_zone_id; current_timezone() may be a fixed offset (+05:30).
• CAST(TIME '12:00:00' AS TIME WITH TIME ZONE) uses the session zone's current offset (+02:00 in Europe/Berlin on
  2026-10-07).

Checklist

• pre-commit run --all-files passes (2026-10-08; run outside the nono sandbox, whose loopback restriction
  fails two backend::tests that need a local socket).
• CHANGELOG.md has entries under ## [Unreleased] (Changed + Fixed, including backfilled entries for the
  type-info and connector-quoting commits).
• No client error added or moved outside map_trino_error.
• New tests were checked failing first or by mutation: acceptance suite red before the fixes; interval unit/FFI
  tests red (3-07, -1-00); SQL_CONVERT_TIME contract red without the override; session-zone FFI tests fail
  when the driver delivers UTC or ignores SET TIME ZONE; timetz FFI test fails with UTC delivery.

If it applies

• Integration suite against a live Trino: all suites green (pbi slicer semantics 64/0, folding contract 41/0,
  type matrix 690/0, ffi 78/0).
• Windows suites in the VM (full run 2026-10-08, fe0e43f): integration ×4 32/0, sql surface 71/0, escape
  sequences 123/0, session keys 10/0, folding contract 41/0, pbi slicer semantics 64/0, tls 19/0, raw C ABI
  154/0, describe param 26/0, type matrix 690/0, transactions 10/0; spooling fails (pre-existing, see notes);
  oauth and harness unit tests skipped by design.
• Core defects fixed in core (preferred row, char→interval conversion, overflow); needs core 2f9f725 or the
  release that contains it.
• Connector rebuilt and loaded in Power BI Desktop (2026-10-08): text, interval, timestamptz slicers select
  their rows (UTC and TimeZone=Europe/Berlin); time slicer shows the fold error; Trino history: 0 failed
  queries.

Notes for the reviewer

• Visible to every application: interval text format and timestamptz/timetz values outside a UTC session change
  (CHANGELOG "Changed").
• Windows spooling fallback check fails with ABANDONED_QUERY — pre-existing, not caused by this branch: a
  2,000-row fetch takes 69–187 s on the Windows VM with both main and this branch (Linux 1.6 s), so 20,000 rows
  leave >5 min between page fetches. Not CPU-, disk- or memory-bound; under separate investigation. Explained: ODBC
  Driver Manager tracing was switched on in the test VM. With it off, the same 2,000 rows take 1.3–1.7 s on Windows
  (42–172 s with it on); not a driver defect.
• Not verified: Power BI Service via the on-premises data gateway.
• timestamp_with_utc_tz_returns_utc_via_get_data has the same "only in a UTC session" naming as the renamed test;
  left unchanged.

Update 2026-10-09: CI and test follow-ups

• 6e83771 test: MinIO for the spooling CI leg now comes from docker.io/pgsty/minio and docker.io/pgsty/mc
  (as stackabletech/demos uses); pulls of quay.io/minio were refused with "unauthorized". Spooling suite locally:
  10 passed, 1 skipped by design.
• 7c2722d test: a_timeout_names_the_cause_beneath_reqwest listens on the fixed loopback port 29871 instead of an
  OS-assigned one, so a per-port sandbox (Landlock) can allow it; pre-commit run --all-files now passes inside one.
• chore: core pin 2f9f725 → 53e5a89, which adds exact numeric → SQL_C_INTERVAL_* conversion and
  01S07 for a non-zero interval-text digit past the ninth (see the core PR). No driver code change; with the new
  pin cargo test --locked -- --include-ignored 514 passed, every integration suite passes (pbi slicer semantics
  64/0, spooling 10/0), pre-commit run --all-files 16/16.
• e1b3c91 chore: core pin → release tag v0.1.1 (4f5e802, the squash-merged core work plus the version bump; same code as
  53e5a89). Re-verified: 514 passed, all integration suites pass, pre-commit 16/16.
• a466d0b test: the Windows runner stops when ODBC Driver Manager tracing is on in the VM (it made reads 25–130×
  slower and caused the spooling ABANDONED_QUERY noted above); --allow-dm-trace overrides. Documented in
  WINDOWS.md; the README's troubleshooting section covers ABANDONED_QUERY / Query not found on large reads.
  Full Windows run green, spooling included.
• e26349e test: re-running setup.sh on a running stack no longer leaves Trino with an empty /etc/trino; a changed
  config fragment now recreates Trino (config hash as a compose label). Full Linux run green.
• c7c0b81 docs: WINDOWS.md, the README and the runner's doc comment describe the tracing slowdown qualitatively
  instead of quoting test-VM timings.
• 197cf66 docs: the README's Power BI section notes that incremental-refresh bounds on a timestamp with time zone
  column follow the session time zone (TimeZone). Background: in Power BI Desktop the RangeStart/RangeEnd filter
  folds into one Trino query for timestamp, timestamp with time zone and date (after Change Type → Date/Time)
  columns, with no folding warning in the policy dialog. The bounds arrive as plain TIMESTAMP literals, so for a
  timestamptz column Trino compares them in the session zone (checked in the Trino CLI: the same boundary comparison is
  false in UTC and true in Europe/Berlin). No code change.

🤖 Generated with Claude Code

[maltesander]

LGTM!