fix(moe-shards): guard empty q4 dense FFN + wire --metal (closes #151)

[deem0n]

Summary

Two fixes for larql run --moe-shards on pure-MoE models (e.g. Gemma-4 26B A4B), reported in #151:

1. Empty interleaved_q4k.bin panic guard. resolve_ffn_weights panicked with range end index 3345408 out of range for slice of length 0 when the dense FFN mmap was 0 bytes. Pure-MoE vindexes ship no dense FFN tensor, so the fallback branch sliced into an empty mmap. Now guarded with q4_ffn_mmap.is_empty(): returns empty QuantWeight stubs — patch_pipeline_layers_for_remote_moe overwrites them downstream and moe_fn supersedes the dense FFN path during decode, so the stubs are never read.

2. --metal not wired into run_with_moe_shards. Always used default_backend() instead of honouring the CLI flag. Mirrors the pattern PR fix: wire --metal into remote FFN path, add post-FFN norms, flush stdout (cherry-pick of #115) #122 already applied to run_with_remote_ffn: explicit metal_backend() with CPU fallback on init failure, and a clear error when the gpu feature isn't compiled in.

Both fixes were needed to get Gemma-4 26B A4B running with --moe-shards.

Files changed

• crates/larql-compute/src/pipeline_layer.rs — is_empty() guard + regression test
• crates/larql-cli/src/commands/primary/run_cmd.rs — new metal: bool param on run_with_moe_shards, mirroring run_with_remote_ffn's backend init, threaded from args.metal at the call site

Net diff: 2 files, +60 / -1.

Test plan

Local verification on macOS, branched off current chrishayuk/larql:main (post-#145):

• cargo check -p larql-compute -p larql-cli --lib --tests — clean
• cargo clippy -p larql-compute -p larql-cli --lib --tests --no-deps -- -D warnings — clean
• cargo fmt -p larql-compute -p larql-cli -- --check — clean
• cargo test -p larql-compute --lib — 657 passed (includes new resolve_ffn_weights_returns_empty_stubs_when_q4_ffn_mmap_is_empty regression)
• cargo test -p larql-cli --tests — 3 passed (1 ignored, model-heavy)

CI verification (all green — 15/15)

• test - ubuntu-latest (inference + cli) — pass (1m27s, 8m17s)
• test - macos-14 (inference + cli) — pass (5m18s, 2m49s)
• test - windows-latest (inference + cli) — pass (12m23s, 12m26s)
• test · ubuntu-latest (compute) — pass (3m20s)
• test · macos-14 (compute) — pass (6m28s, 1m1s)
• test · windows-latest (compute) — pass (10m50s)
• coverage - ubuntu × 2 — pass (6m46s, 6m29s)
• coverage · ubuntu — pass (51s)
• bench — pass (9m27s)
• verify — pass (3m48s)

🤖 Generated with Claude Code