[lalalune]

Overview

The architecture test target fails to compile at 56e70ff because its assistant fixture calls the removed pre-norm embedding APIs. Use the current NextN setter/getter and update the comment to match the API. This repairs the test build used by the native dependency pinned in elizaOS/eliza#36270.

The follow-up also closes the outer condition in the installed ggml CMake package config with endif() and classifies Kokoro as a no-RoPE architecture. These repair the package-consumer configure failure and the compiler switch error exposed by CI.

Validation

• Original source fails syntax compilation on the two removed identifiers; repaired source passes with the same compiler flags.
• Release CPU-only CMake build of test-llama-archs passes on macOS arm64.
• test-llama-archs -a llama -s 1 passes dense and MoE CPU comparison/roundtrip cases. Tensor split cases skip because this build has no non-CPU backend.
• Gemma 4 and Gemma 4 assistant runtime validation remains unavailable: the existing fixture explicitly skips both for missing ISWA cache parameters. The Gemma invocation runs zero cases, not a passing model test.
• Follow-up: local build and installation pass, and examples/simple-cmake-pkg configures and links against that installation.
• Follow-up: llama-model.cpp fails with -Werror=switch before the Kokoro case and passes with the same compiler flags afterward.
• Full hardware CI and Android execution were not run locally. Fresh CI is running for the follow-up commit.

Requirements

• Contributing guidelines reviewed by Codex. The maintainer explicitly authorized this commit, push and PR after acknowledging the repository's automated-submission policy and stated ban risk.
• AI usage disclosure: YES. Codex identified the compile failure, updated two API calls, one comment, the CMake conditional terminator and the Kokoro switch case, and ran the listed checks. The submitting maintainer owns review and maintenance of these changes.