[ch-wan]

CP prefill calls the transformer body directly and bypasses the outer model.forward patched by KV canary. This skips the canary pre/post hooks and triggers SimplePhaseChecker: phase mismatch (expect=3, actual=1) in both log and raise modes.

When prefill CP is enabled, also wrap the transformer body's forward with the existing canary hooks. The shared bracket scope keeps decode, which still enters through the outer model, from running those hooks twice. Add a CP=2 regression covering prefill followed by CUDA-graph decode in raise mode with real-KV hashing enabled. The registered test starts the server once; log uses the same hook path and was checked separately.

Validation on two H200 GPUs:

• Reproduced the phase assertion on unmodified main at 66f19f5c46; the same four-request smoke passes with this fix.
• test/registered/mock_model/test_e2e_cp.py: passed in 46.5 seconds (down from 93.1 seconds for two launches), with token-oracle input checking and --kv-canary-real-data all. A separate log smoke also passed; CI retains only raise to avoid repeating server startup.
• Qwen3-0.6B real weights and regular sampling: eight requests passed with CP=2, log, real-KV hashing, and verify-token assertions, including repeated prompts that reuse the prefix cache.
• All applicable pre-commit checks passed.

---

CI States

Latest PR Test (Base): ❌ Run #35664946928
Latest PR Test (Extra): ❌ Run #35664946646
Latest PR Test (AMD ROCm 10): ❌ Run #35664947008

[ch-wan]

/rerun-test test_e2e_cp.py

[github-actions]

Results for /rerun-test test_e2e_cp.py:

🚀 2-gpu-h100 (1 test): ❌ View workflow run

cd test/ && python3 registered/mock_model/test_e2e_cp.py