Skillrouter Qwen3 0.6B LoRA experimental v0.1

Experimental open-weight LoRA adapter trained on 5,363 weak description-to-skill pairs from the 2026-09-22 skills.sh snapshot. On 672 source-held-out, description-derived queries, recall@1 was 87.5% (BM25 84.8%; frozen Qwen3 71.6%). These queries are not independent user requests. Native Go inference parity and no-match evaluation remain open. See MODEL_CARD.md for the exact base revision, corpus hash, training recipe, license, attribution, and limitations. Archive SHA-256: dbfa4ad9cc0ab89ca5689ae2c4018aef068eb30eba1fc626b396ceb43130257a.