PR is ready - routed-experts vLLM pin rebuilt from v0.21.0 + revert #42434 + PR39568

This branch moves prime-rl to the routed-experts response path used by current renderers/verifiers main, without the verifier-side renderer bypass. The x86_64 vLLM pin now points at a PrimeIntellect v0.5.0 release asset built from upstream vLLM releases/v0.21.0 plus upstream revert vllm-project/vllm#42434 and upstream PR vllm-project/vllm#39568.

The previous PR39568-only wheel still contained the PR39917 validation that rejected enable_return_routed_experts with async scheduling during inference startup. This new wheel removes that validation via revert #42434 while keeping the PR39568 routed-experts transfer path.

• keep routed-experts data opaque through verifiers during token truncation; prime-rl decodes at the orchestrator boundary
• preserve this branch's routed-experts source of truth: RoutedExperts(data, shape, dtype), explicit dtype maps, and _pack_routed_experts / _unpack_routed_experts for multi-turn stitching
• update trainer packing/loading to slice, append, pad, and reconstruct the RoutedExperts transport struct with torch.frombuffer
• pin vllm-router to 0.1.25 for the matching raw-uint8 schema and add pybase64
• pin x86_64 vLLM to vllm-0.21.0+cu129.r42434.pr39568.a106aa6-cp38-abi3-manylinux_2_24_x86_64.whl
• update deps/renderers to main 3ae276c (renderers-v0.1.8.dev25), including routed-experts sidecar parsing and fastokens>=0.2.0
• update deps/verifiers to main 521d436c (v0.1.15.dev8-1-g521d436c), including the routed-experts response sidecar support and verifier helper cleanup
• add the fastokens exclude-newer exemption and lock fastokens==0.2.0 to satisfy current renderers main
• disable vLLM async scheduling only for the NIXL routed-experts capture path, where async scheduling leaves placeholder sampled-token state during capture
• fix rendered multi-node orchestrator args to use the student client config keys

Related PRs

• Router payload format: Feat: downstream perf improvements - bump .25 router#34
• Router release trigger: release: v0.1.25 router#35
• Renderers routed-experts sidecar parse: Handle routed experts as response sidecar renderers#54
• Verifiers routed-experts response sidecar: Support routed experts response sidecar verifiers#1423
• Verifiers cleanup against main: Clean routed experts response path verifiers#1433
• vLLM routed-experts upstream: [RFC] Replace shared-memory routed experts with ModelRunnerOutput transfer and HTTP support vllm-project/vllm#39568
• vLLM PR39917 revert: Revert "[Core] Replace routing replay with device cache and async D2H pipeline" (#39917) vllm-project/vllm#42434

Verification

• uv sync --all-extras --locked
• uv lock --locked
• uv run ruff check .
• uv run ruff format --check --config=pyproject.toml
• PYTEST_OUTPUT_DIR=/tmp/outputs uv run pytest tests/unit -m "not gpu" - 454 passed, 65 deselected, 35 warnings
• Wheel smoke import: vllm.__version__ == "0.21.0+cu129.r42434.pr39568.a106aa6", ModelRunnerOutput.routed_experts present, _validate_return_routed_experts absent, and the old async-scheduling validation string absent.
• Uploaded the new r42434.pr39568.a106aa6 vLLM wheel to the prime-rl v0.5.0 release and removed the stale PR39568-only wheel asset that failed inference startup.