fix(build): clean up Clang compiler warnings

[trigger2k20]

Clean up Clang compiler warnings

Branch and commit

• Branch: fix/clang-compiler-warnings
• Commit: e79aee8
• Commit message: fix(build): clean up Clang compiler warnings
• Base used for development: fd93c41aa6ae2c7d1cc1a1e2d6b79dbe6d341708 (v1.10.2)

Summary

This change removes Apple Clang compiler warnings from the portable CPU/test build without globally suppressing warning classes.

The cleanup covers GLM-5.3, Qwen3.8 Edge adapter initialization, the expert-store registry/test initialization, and DeepSeek V4 CPU/GPU conditional compilation. The changes are intended to be warning cleanup only; no performance claim is made.

Changes

GLM-5.3

• Explicitly initializes the Mat.vk field to NULL in the three affected matrix initializers.
• Keeps hc_mult declarations only in code paths where the value is actually used, avoiding unused-variable diagnostics in alternate/Segment builds.

Qwen3.8

• Converts the ColiEdgeAdapter initializer to designated initialization.
• Initializes the reserved_fn array explicitly.
• Removes the scalar-brace/missing-field diagnostics and makes the adapter initializer less dependent on struct field order.

Expert store

• Explicitly initializes the additional ColiExpertStore field in test_expert_store_ops.c.
• Makes COLI_ESR_EXPORT feature-detect externally_visible instead of unconditionally applying an attribute unsupported by Apple Clang.
• Retains the used attribute on GCC/Clang builds.

DeepSeek V4

• Restricts gpu_batch, cols, and gpu_compute state to GPU-tier compilation where those values are consumed.
• Removes unused scales/qdq scratch allocations and their size accounting.
• Keeps the CPU-only path free of GPU-only unused-variable diagnostics.

Compiler diagnostics addressed

The original Apple Clang build/test output contained warnings in these groups:

1. Three glm53.c missing field 'vk' initializer diagnostics.
2. One GLM-5.3 Segment-build unused variable 'H' diagnostic.
3. Qwen3.8 Edge adapter braces around scalar initializer.
4. Qwen3.8 Edge adapter missing field 'reserved_fn' initializer.
5. Expert-store test missing field 'gpu' initializer.
6. Four expert_store_registry.c unknown attribute 'externally_visible' ignored diagnostics.
7. DeepSeek V4 unused variable 'gpu_batch'.
8. DeepSeek V4 unused variable 'cols'.
9. DeepSeek V4 unused variable 'scales'.
10. DeepSeek V4 unused variable 'qdq'.
11. DeepSeek V4 variable 'gpu_compute' set but not used.

Expected runtime/test diagnostics containing the word warning were not treated as compiler warnings.

Validation

Final validation on macOS/Apple Clang:

Ran 722 tests in 54.468s

OK (skipped=39)

=== COMPILER WARNINGS ===
ZERO COMPILER WARNINGS

The compiler-warning check matched source-location compiler diagnostics separately from expected runtime/test messages.

git diff --check also passed.

Scope

Files changed:

• c/deepseek_v4.c
• c/expert_store_registry.c
• c/glm53.c
• c/qwen38.c
• c/tests/test_expert_store_ops.c

Final commit diff:

5 files changed, 43 insertions(+), 21 deletions(-)

Notes

• No warning class is disabled globally.
• No throughput/performance improvement is claimed.
• GPU-only DeepSeek variables remain available in GPU-tier builds; the cleanup avoids defining them in CPU-only builds where they are not consumed.
• The registry attribute handling uses feature detection so Apple Clang can omit externally_visible while compilers supporting it retain the attribute.

[JustVugg]

Closing and reopening only to make CI build against current dev: the one red job (Windows UCRT64) is tests/test_qwen36_tier_fill_wait.c: implicit declaration of setenv, which is a base-side break that #1459 fixed at 06:37 UTC, and this run was created at 23:00 UTC the night before. Nothing in your branch is at fault. Everything else in the diff I have already reviewed: the scales/qdq scratch is genuinely unused in the indexer batch function, __has_attribute(externally_visible) is the right guard for Clang, and the trailing NULL initialises the vk field dev already warns about. If the fresh run is green it goes in.

[trigger2k20]

Thanks for checking this and for confirming the UCRT64 failure was base-side.

The fresh runs against current dev are now both green:

• CI qwen38: the MTP head's layer on the device beside the full chain #2200 — passed
• check fix(benchmarks): propagate UX engine failures #2180 — passed