Track main correctness fixes for CAST initialization, day-time intervals, and FORCE join validation (PR #30)

[liuyongvs]

Problem and related PR

Tracking issue for the main-branch correctness fixes proposed in #30 (still open and unmerged when this issue was created). These problems belong on main independently of the Flink 2.1.3 backport. This issue provides a place to track the defects, regression evidence, and completion of that PR's scope.

Independently reproduced: non-nullable CAST initialization

Confirmed on main ebe550c69122909cd08fca8ee89380478a92cc5c, released Apache Flink 2.2.1, JDK 17, UTC, default compatibility settings.

Use a bounded runtime source with schema s CHAR(8) NOT NULL and these valid, non-null values:

1234.567
9876.543

Run the same query in two fresh streaming TableEnvironments, first on stock Flink and then with NativePlanner.install(tableEnvironment):

SELECT CAST(s AS DOUBLE) AS cast_value FROM input_table;

The source must supply runtime rows rather than a directly folded constant expression. The audited query routes through NativeCalc, with one actual native substitution.

Expected: both engines return 1234.567 and 9876.543.

Observed on stock Flink: both rows are returned successfully.

Observed on native execution: the job fails during function initialization, before the valid input values can be evaluated:

IllegalStateException: failed to open UDF tech.streamfusion.planner.HostCastFunction
TableRuntimeException: Error when casting CHAR(8) NOT NULL to DOUBLE NOT NULL.
NullPointerException: Cannot invoke "org.apache.flink.table.data.binary.BinaryStringData.trim()"
because "<local2>" is null

HostCastFunction.initializeExecutor() warms the generated cast by calling executor.cast(null) repeatedly, even when the input type is non-nullable. A valid non-null warmup value is required for those generated executors. This is an initialization failure for valid input, not malformed-input behavior.

Remaining scope described by PR #30

The following two areas are included in the existing PR's report and proposed changes; the independent audit evidence above specifically reconfirms the CAST defect:

1. Day-time interval output and predicates: preserve the millisecond representation when interval expressions pass through native operators; reject unsupported interval filter/join predicates at planning time so they fall back rather than fail during execution. Duration and year-month interval gaps should remain explicitly documented.
2. FORCE join validation: preserve stock Flink's validation failure under table.optimizer.delta-join.strategy=FORCE; native substitution must not hide an error that Flink is supposed to raise. This does not require adding native delta-join support.

Acceptance criteria

• Valid non-nullable CAST queries initialize successfully and match stock Flink; add nullable/non-nullable regression coverage without swallowing real cast-initialization errors.
• Supported day-time interval projections, keys, filters and join residuals preserve Flink semantics, with explicit planning-time fallback for unsupported combinations.
• FORCE delta-join validation errors remain observable as they are in stock Flink.
• Relevant regression tests and coverage documentation are included in the main-branch fix.
• Link the shipping fix to this issue and verify the reproduced CAST case on the merged main revision.

Proposed implementation: #30.