[Nathan-D-R]

Write-side silent failures. Stacked on #57.

Findings fixed (tests in openqvd-arrow/tests/write_types.rs, run for Arrow 59 and 60; test_write_types.py)

• S1-3 non-microsecond timestamps, Time32/Time64 and Duration became text typed TIMESTAMP/TIME. The catch-all cast-to-string is gone. Timestamp (s/ms/us/ns), Time32/64, Duration, Date32 and Date64 (whole days) have exact mappings; every other type is UnsupportedArrowType (column and Arrow type named). Dictionary arrays are unpacked to their values; Utf8View is accepted. Tests: nanosecond_timestamps_are_timestamps_not_text, time_and_duration_are_times, second_timestamps_and_overflow, unmapped_types_are_errors_not_text, date64_must_be_whole_days, dictionaries_are_unpacked.
• S1-5 UInt32 wrapped; Int64 beyond 32 bits and UInt64 became text; Decimal became text. Integers beyond 32 bits are a dual of (exact float, exact decimal text), errors beyond 2^53; Decimal128/256 are a dual with the exact decimal text, errors when the unscaled value exceeds 2^53 or scale exceeds 22. The wrong "saturating" doc comment is replaced by the full mapping table. Tests: unsigned_integers_are_not_wrapped, large_int64_stays_numeric, integers_beyond_2_pow_53_are_errors, decimals_keep_their_digits, decimals_too_wide_for_a_float_are_errors.
• S1-12 time zones silently dropped. TimezoneNotSupported by default; lenient check timezone_to_utc writes UTC wall-clock time and attaches a Warning (WriteTable::warnings(), Python QvdWarning). Test: time_zone_is_an_error_unless_allowed, test_timezone_is_an_error_unless_allowed.
• S1-13 unchecked u32 casts for row count, symbol-table size, distinct values: now LayoutOverflow. (A table above 4 Gi rows cannot be built in a test; the checks are u32::try_from at the three sites.)
• Single-rounding timestamp serials, so write then read is exact for whole seconds (whole_second_timestamps_survive_a_round_trip_exactly, 5,000 values).

Not reproduced / deferred

• S1-13 has no executable test (needs more than 2^32 rows or 4 GiB of symbols).
• Nanosecond values are rounded to the nearest microsecond: a Qlik serial is an f64 day count and resolves about 0.6 us, so finer values cannot be stored. Documented in the mapping table.

API breaks

record_batch(es)_to_write_table returns tables carrying warnings(); new error variants and a new LenientCheck::TimezoneToUtc/WarningKind::TimezoneToUtc. Python write now returns None but emits warnings (the private _write returns them).

Tests changed

api.rs::leniency_is_a_set_of_named_checks and test_errors.py::test_leniency_basics no longer assume there is exactly one check.

Corpus

round_trip (read, re-encode with the writer, read, compare every cell) over /mnt/nas/data/QVD: 1,145 files, 1,072 ok, 15 over the size limit, 49 not QVD, 9 fail. The 9 are the three damaged fixtures plus the six NonNumericTemporal files from #57; no writer-related failures.

Lenient mode

Added: timezone_to_utc (write side; no corpus relevance, but tz-aware timestamps are common in Polars/pandas data and strict-only would block them with no workaround except converting first, which the error message explains).

Docs-site edits needed

Arrow-to-QVD type mapping table (from the record_batches_to_write_table doc), timezone handling, Python write warnings.

Review fixes

• S7: nanosecond remainders are ValueNotRepresentable; new lenient check sub_microsecond_precision rounds and warns once per column (nanosecond_digits_are_an_error_unless_rounding_is_allowed).
• S8: checked Date32/Date64 arithmetic; dates and timestamps limited to years 0001 to 9999 and checked against the reader's own conversion (openqvd::serial) so accepted values re-read strictly (dates_and_timestamps_the_reader_would_reject_are_not_written).
• S9: DuplicateColumn in WriteTable::new and when opening a file (duplicate_column_names_are_rejected, duplicate_field_names_are_an_error).
• N10: round-trip type drift is documented in the mapping table; Duration columns use the INTERVAL format.