0.52.100

What's Changed

• Decline the sequence-concatenation stage for an operand produced by a call on a tensor-library value (wala#911) by @khatchad in #845
• Note the null-URL getScript face of the stale-fixture-jar trap by @khatchad in #819
• Bump the maven-plugins group with 3 updates by @dependabot[bot] in #823
• Keep the rank of a top_k whose k is not a resolvable constant (wala#609) by @khatchad in #840
• Ignore .mvn/maven.config by @khatchad in #844
• Read top_k's input shape in the caller's frame when its points-to set is empty (wala#609) by @khatchad in #841
• Read a from_generator dataset's element shapes from the generator's own yields (wala#903) by @khatchad in #846

Fixed Since 0.52.99

The known issue named in 0.52.99's notes is narrowed by this release. That release typed a scalar_literal + tensor broadcast add as a rank-1 sequence, because the concatenation typing assumed + was Python sequence concatenation. The stage now declines when the other operand is produced by a call whose attribute chain roots at a modeled tensor-library namespace, so those cases fall back to an unknown rank rather than a wrong one.

Known Issues

• The narrowing above is a negative filter and does not reach zero (wala#911, which stays open). A scalar_literal + tensor broadcast add is still typed as rank 1 when the operand's chain cannot be rooted at a modeled library: an unmodeled third-party library, a chain broken at an unmodeled hop, or a binding deeper than the walk's bound. A wrong rank narrows what a consumer accepts and can reject a valid caller, where an unknown rank suppresses the constraint and lets the program run.

Full Changelog: 0.52.99...0.52.100