tree 4b4742756cd78b51c23fb015158b1e5d0fb5e7d4
parent d6a163cbe4dbebef25202bbf71fd29b190215e60
author Marc Kupietz <kupietz@ids-mannheim.de> 1785391794 +0900
committer Marc Kupietz <kupietz@ids-mannheim.de> 1785391794 +0900

Fix collocator lookups for merged models and guard word ranks

w2v.merge was unusable: the server died at startup with "Undefined subroutine
&main::mergeVectors", because mergeVectors was not exported.

With that fixed, count based collocators and precomputed profiles were looked
up with unshifted ranks. The collocator db and the profiles belong to the
primary model, which mergeVectors puts at [merged_end, words), so its ranks
have to be shifted back. The adjustment existed but was dead code, as it read
$IDS::DeReKoVecs::Read::mergedEnd while the value was assigned to
$main::mergedEnd. Instead of wiring a second copy of the value, the merge
offset is now read from the C layer, which is the only place that knows it.

Out of range ranks reach libcollocatordb straight from the request and crash
the process there, so a single GET /getClassicCollocators?w=<beyond vocabulary>
was enough to kill a worker. Ranks are validated before use now, both for the
collocator db and for the model matrix in getVecs, and out of range ids yield
an empty result.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Change-Id: I26ef508e3682cbe00df27d4de5588d1f1493c701