[Spec] Publish DFLASH verify read-done event for fine-grained WAR barrier

[hnyls2002]

What

Publish the WAR read_done event from the DFLASH target-verify forward so the scheduler's global _apply_war_barrier takes the fine-grained wait_event(read_done) fast path instead of the coarse wait_stream(forward_stream) fallback.

Why it is valid

The target-verify cuda graph builds its page table in load_batch (pre-replay) and then replays reading only that static snapshot. So the forward finishes reading the shared req_to_token / SWA buffers before replay — read_done is recorded at exactly that point, so it is a sound WAR signal.

WAR ordering

The over-alloc write to req_to_token (plan stream) stays ordered after the previous forward's read through this chain:

read_done  (recorded on the forward stream, after req_to_token is read)
   │  decode_cuda_graph_runner: war_fastpath_read_done_event = read_done
   ▼
schedule_stream.wait_event(read_done)         scheduler.py:1521  (_apply_war_barrier, called at loop top :1578)
   │
   ▼
plan_stream.wait_stream(schedule_stream)      dflash_info_v2.py:188  (prepare_for_decode)
   │
   ▼
over-alloc writes req_to_token (plan stream)  dflash_info_v2.py:218

prepare_for_decode already does plan_stream.wait_stream(schedule_stream) (originally to see the scheduler's filter/merge writes), so the plan-stream over-alloc write inherits the read_done dependency transitively — no explicit plan-stream wait is added. The barrier now waits only until req_to_token has been read (start of the verify forward) instead of the whole forward, recovering the overlap window the coarse fallback serialized.

Dependencies

Builds on #29556 (merged) — removed verify_done and the dflash _war_barrier_enabled opt-out, routing dflash through _apply_war_barrier. Independent of #29343 (fa3).

---

CI States

Latest PR Test (Base): 🚫 Run #28342455012
Latest PR Test (Extra): ❌ Run #28342454942

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[hnyls2002]

/rerun-test registered/spec/dflash/test_dflash.py registered/cuda_graph/piecewise/test_pcg_with_speculative_decoding_dflash.py registered/spec/test_gemma4_dflash_31b_extra.py (latest main)

[github-actions]

Results for /rerun-test registered/spec/dflash/test_dflash.py registered/cuda_graph/piecewise/test_pcg_with_speculative_decoding_dflash.py registered/spec/test_gemma4_dflash_31b_extra.py:

🚀 1-gpu-5090 (2 tests): ✅ View workflow run

cd test/ && python3 registered/spec/dflash/test_dflash.py
cd test/ && python3 registered/cuda_graph/piecewise/test_pcg_with_speculative_decoding_dflash.py

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

cd test/ && python3 registered/spec/test_gemma4_dflash_31b_extra.py

[github-actions]

Results for /rerun-test registered/spec/dflash/test_dflash.py registered/cuda_graph/piecewise/test_pcg_with_speculative_decoding_dflash.py registered/spec/test_gemma4_dflash_31b_extra.py (latest main):

🚀 1-gpu-5090 (2 tests): ✅ View workflow run

cd test/ && python3 registered/spec/dflash/test_dflash.py
cd test/ && python3 registered/cuda_graph/piecewise/test_pcg_with_speculative_decoding_dflash.py

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

cd test/ && python3 registered/spec/test_gemma4_dflash_31b_extra.py

⛔ (latest: No test file found matching (latest under test/registered/ or python/sglang/multimodal_gen/test/.

⛔ main): No test file found matching main) under test/registered/ or python/sglang/multimodal_gen/test/.

[hnyls2002]

/tag-and-rerun-ci