[r2d3]

ASIMD (asimd/search/minmax.h):

• OP_SIGN_SEARCH_CORE: remove stray x32 bookkeeping from the x16 block. It made
  sign_min/sign_max ignore up to 15 trailing samples.
• sign_minmax: clear v1 rather than v2 when count == 0, so max returns 0.
• sign_minmax: use separate min and max candidates in the final reduction.
  The max lanes were replaced by the min lanes, giving a wrong max.
• sign_minmax: declare v4-v7 as clobbered.

Generic (generic/search.h):

• sign_minmax compared the signed value with amax; it now uses the absolute value.

Tests (utest/search):

• sign_minmax failed only if min AND max were both wrong; it now fails if either is.
• same for minmax (min AND max should be wrong for the test to fail)
• Add sizes that exercise the x16 block with short tails.

Tested on Apple M1 Max, macOS 26.6: the new sign_* tests fail before the fix and
pass after it, and the full utest suite passes.

Not tested on other architecture. Test correction could expose bug on these architecture.

[r2d3]

What is the active branch for this repo ? master, devel, ... ?

[sadko4u]

Hello!

Thanks for the contribution. Merged changes to the devel branch: ae96fa9
Just sad to see this patch coming exactly after 1.2.37 release of LSP Plugins.

Please consider further PRs to be based on the devel branch.