[ump45nose]

Which issue does this PR close?

• Closes Reader does not promote date partition values and column bounds to timestamp / timestamp_ns #3503.

What changes are included in this PR?

Follow-up to #3483. With v3 type promotion a date column can become timestamp or timestamp_ns, but older manifests keep day-count lower/upper bounds. Compared as raw numbers those day counts sit far below any timestamp in micros/nanos, so bound pruning wrongly drops files.

• Datum::promote_to converts a date (days from epoch) to timestamp (micros) or timestamp_ns (nanos) via checked multiplication; a date outside the target range returns DataInvalid. Matching is on (self.data_type(), target) so a bare Int is not treated as a date.
• Datum::try_from_bytes accepts 4-byte payloads under a timestamp / timestamp_ns schema (manifests that carried an older date bound forward) and reuses the same conversion.
• InclusiveMetricsEvaluator promotes date lower/upper bounds before comparing them to a timestamp predicate (field name attached on error) and propagates overflow as DataInvalid.

Partition date → timestamp / timestamp_ns is not promoted: the spec rules that out for identity transforms.

timestamptz targets are left alone (date → timestamptz is not an allowed promotion).

Are these changes tested?

Yes, new / folded unit tests:

• spec::values::tests::test_datum_promote_date: table-driven date → timestamp / timestamp_ns (including negative), timestamp_ns edges day 106751 fits / 106752 doesn't, overflow → DataInvalid, bare Int does not promote.
• spec::values::tests::avro_bytes_timestamp_from_date and test_parse_manifest_promotes_date_bounds_under_timestamp_schema: 4-byte date bounds under a timestamp schema.
• inclusive_metrics_evaluator::eval_promotes_date_bounds: table-driven < / > / = / IN with inside (kept) and outside (pruned) literals for both targets; overflow asserts ErrorKind::DataInvalid.

AI Disclosure

I used an AI coding assistant to help write the code and tests. I reviewed the changes and ran the tests locally.

[ewoodbury]

Contributing my review here, changes look solid. Agree with the approach of promoting the bound where it is compared with a checked multiply, and it follows #3483. I confirmed the new evaluator tests fail with the date arms reverted, so they pin the fix.

Main asks from me before merge:

• Handle the spec's 4-byte bound case under a timestamp/timestamp_ns schema. Those manifests still fail to load.
• Settle whether the partition-value half belongs here. My reading of the spec is that it rules out this promotion for identity partitions.
• Pin the pruned side of <, > and IN.