Summary

This PR vendors the pinned RADE Opus snapshot into the repository and updates the RADE build to consume that local snapshot instead of downloading Opus and its neural model at build time.

It also removes the CI and Docker model pre-cache workaround that existed only to reduce flakiness from those network fetches.

What changed

• vendor a prepared RADE-compatible Opus snapshot under third_party/opus-rade/
• switch third_party/radae/cmake/BuildOpus.cmake to build from the vendored archive via CMake instead of downloading upstream Opus during the build
• keep PatchOpusNnet.cmake as a refresh helper for updating the vendored snapshot, not as a normal build step
• update build messaging and README to reflect the new offline RADE dependency path
• remove CI and Docker pre-cache steps that were only there to paper over the Opus/model download dependency

Why

The old RADE path depended on build-time network downloads.

That created two recurring problems:

1. Local agent builds from Codex and Claude needed explicit sandbox approval before they could finish the Opus fetch, and failed downloads wasted time and tokens.
2. GitHub Actions could stall or fail on the Opus/model download path, which made CI slower and less reliable.

Vendoring the exact pinned snapshot keeps behavior deterministic while removing the network dependency from normal builds.

Benefits

• Faster RADE builds: the original download-based full RADE build completed in 67.16s wall-clock when run with unrestricted network access, while the vendored build completed in 53.34s, a 13.82s improvement or about 20.6% faster on the test machine.
• Better reliability: the original path did not complete inside a normal sandboxed agent build and failed in the build_opus download step after 86.11s; the vendored path completed fully offline.
• Easier local agent use: Codex and Claude no longer need sandbox approval just to fetch Opus before they can compile RADE-enabled builds.
• Cleaner GitHub CI: builds no longer depend on opportunistic external downloads or the Docker image model cache workaround.
• Lower token cost and operator overhead: fewer stalled agent runs, fewer retry loops, and fewer approval/download waits means less wasted inference time when builds are kicked off from local coding agents.
• More reproducible builds: RADE now builds from the exact vendored snapshot that the project has chosen to support.

Validation

• cmake -B build-vendor-test -G Ninja -DCMAKE_BUILD_TYPE=RelWithDebInfo
• cmake --build build-vendor-test --target build_opus -j4
• cmake --build build-vendor-test -j4
• cmake -B build-no-rade-test -G Ninja -DCMAKE_BUILD_TYPE=RelWithDebInfo -DENABLE_RADE=OFF
• cmake --build build-no-rade-test -j4
• baseline benchmark from a clean export of HEAD: cmake -B build-bench -G Ninja -DCMAKE_BUILD_TYPE=RelWithDebInfo && cmake --build build-bench -j4
• vendored benchmark from this branch: cmake -B build-bench-current -G Ninja -DCMAKE_BUILD_TYPE=RelWithDebInfo && cmake --build build-bench-current -j4

Tradeoffs

• the repository grows by about 51 MB because the prepared Opus snapshot is now stored locally
• GitHub warns that the archive is above the recommended 50 MB file size threshold, although it is still below the hard file size limit

👨🏼‍💻 Generated with OpenAI Codex (GPT-5.3 Pro) and tested by @jensenpat