Summary

• RLX migration: replaced burn/cubecl/gpu-fft/llama-cpp stack with the RLX inference runtime (rlx, rlx-models, rlx-umap, rlx-fft-metal, rlx-metal) across skill-llm, skill-router, skill-eeg, skill-screenshots, skill-daemon-state, and skill-daemon
• Dependency cleanup: removed unused cubek-matmul patch (burn-era Metal fix, no longer in crate graph); cleaned up stale deny.toml entries (sleepfm, winreg@0.56.0, cubek.git, tribev2-rs)
• Clippy / fmt / deny fixes: boxed large enum variant (RlxTextEmbedding), fixed neutts/llm/tts unused-macro warnings, suppressed CLIP float precision lint, removed dead save_user_settings function, fixed doctest (rust,ignore → no_run)
• Reembed pipeline fix: scan for work before loading the encoder — empty skill_dir now short-circuits to done in <1ms instead of blocking 30s on a HuggingFace download attempt

Test plan

• cargo clippy --locked --workspace — clean
• cargo deny check — clean
• cargo test -p skill-eeg — 120 unit + 5 integration tests pass
• cargo test -p skill-router --features gpu -- umap_e2e --include-ignored — UMAP Metal E2E (small/medium/large) pass
• cargo test -p skill-daemon -- routes::settings::tests::trigger_reembed — reembed tests pass in <1s
• bash scripts/test-daemon-e2e.sh — 19/19 service lifecycle tests pass

🤖 Generated with Claude Code