Summary

• update bridge CI and publish defaults from the old llama.cpp pin to b9016
• include the current MTMD image/source layout required by newer llama.cpp builds
• make CI build and upload the same base + mem64 artifact set that publish emits
• forward native-compatible load options through the JS/C++ bridge ABI (n_seq_max, flash attention, KV cache type, KV-unified, RoPE, split mode, main GPU)

Verification

• built locally against llama.cpp b9016 with WEBGPU_BRIDGE_BUILD_MEM64=1
• node --check js/llama_webgpu_bridge.js
• git diff --check
• source tag v0.1.12 published successfully
• published leehack/llama-web-bridge-assets@v0.1.12 with manifest llama_cpp_tag: b9016 and source commit 4048425c9268e7e9aa330364179bcc567d7d306d