Vector search metric guard is ineffective and ANN segments are scored with the wrong distance metric after the metric changes

[LuciferYang]

Search before asking

• I searched in the issues and found no similar issues.

Paimon version

master (1.5-SNAPSHOT)

Compute Engine

Any engine using a primary-key table with a pk-vector ANN index (the distance metric is configured via fields.<col>.pk-vector.distance.metric).

Minimal reproduce step

1. Create a primary-key table with a pk-vector column and distance.metric = inner_product.
2. Write rows so the ANN index segments are built.
3. Change the column's distance.metric to a different value.
4. Run a vector search.

What doesn't meet your expectations?

The search silently scores the already-built ANN segments with the new metric, so the nearest-neighbor results are wrong. PkVectorAnnSegmentSearcher has a guard that is meant to reject searching a segment whose metric differs from the configured one, but it compares the current metric against a value derived from the same current configuration, so both sides are always equal and the guard never fires.

Anything else?

The segment stores no record of the metric it was built with, so the searcher has no way to detect the mismatch today. A fix needs to persist the build-time metric in the segment metadata and compare against that.

Are you willing to submit a PR?

• I'm willing to submit a PR!