[Spec] Consolidate the verify step into eagle_worker_common.run_eagle_verify

[hnyls2002]

Consolidate the verify step of EAGLEWorkerV2 and MultiLayerEagleWorkerV2 into one eagle_worker_common.run_eagle_verify function. The single-layer eagle verify() (the superset) is the source of truth; both workers' verify() become thin argument-binding wrappers, and multi-layer's stale partial copy is deleted. The two compaction helpers (_finalize_accept_tree_path, _compact_accept_to_front) move along as module-private functions.

The two real differences (preserved verbatim behind switches)

Switch	EAGLE	Multi-layer
metadata_ready_pre_pad	False — relies on eagle_prepare_for_verify marking metadata ready only when the cuda-graph load_batch path ran; the non-graph path re-inits post-pad	True — marks ready pre-pad unconditionally (its pre-existing behavior, kept verbatim)
finalize_tree_path	True — compacts the accepted tree path to the front of each per-req block for topk > 1	False — has never run this compaction (kept verbatim; whether topk > 1 multi-layer should adopt it is left as a follow-up)

What multi-layer gains (each a gated no-op for its configurations)

Feature	No-op evidence
Grammar bitmask (generate_token_bitmask + vocab_mask)	Entire block behind batch.has_grammar; eagle_sample's vocab_mask defaults to None, so the old 3-arg call was already equivalent
Mamba state commit	commit_mamba_states_after_verify early-returns unless the model is mambaish and the backend has the commit hook
c128 draft-state clear	getattr(..., "clear_unaccepted_c128_draft_states", None) is None for other KV caches
NPU mrope position recompute	Gated on _is_npu and model_config.model_is_mrope

Verification

Verification script: https://gist.github.com/hnyls2002/ee26d74ec93ddcd45aa4c485a7c0b996 — runs against the pinned base/head commits:

1. Relocation proof: the committed run_eagle_verify is AST-equivalent to the base commit's eagle verify body plus exactly three declared edits — the finalize_tree_path gate prefix, the gated pre-pad metadata block, and the deletion of the num_tokens_per_req re-stamp (mirroring [Spec] Converge DP-attention spec width scaling onto num_tokens_per_req #31244, which removed it from the base copy after this branch was cut). No other line differs beyond the declared parameter renames.
2. Helper relocation proof: both compaction helpers match their declared moves.

Cross-checks: no test patches any of the moved names; _finalize_accept_tree_path / _compact_accept_to_front have no external callers; StandaloneWorkerV2 / FrozenKVMTPWorkerV2 inherit the eagle wrapper (same switch values as the body they inherited before). compile + ruff clean across speculative/; black reflow AST-equivalence checked.

GPU accuracy validation (eagle + MTP configs) to follow before merge, in addition to the CI reruns.

---

CI States

Latest PR Test (Base): 🚫 Run #29472859914
Latest PR Test (Extra): ❌ Run #29472859780

[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/eagle/test_spec_eagle.py registered/spec/eagle/test_spec_eagle_topk_page.py registered/spec/eagle/test_eagle_constrained_decoding.py registered/spec/eagle/test_eagle_reject_sampling.py registered/spec/test_gemma4_mtp_31b_extra.py registered/spec/test_spec_standalone.py registered/spec/test_frozen_kv_mtp.py registered/unit/spec/test_eagle_worker_v2_topk1_fastpath.py

[github-actions]

Results for /rerun-test registered/spec/eagle/test_spec_eagle.py registered/spec/eagle/test_spec_eagle_topk_page.py registered/spec/eagle/test_eagle_constrained_decoding.py registered/spec/eagle/test_eagle_reject_sampling.py registered/spec/test_gemma4_mtp_31b_extra.py registered/spec/test_spec_standalone.py registered/spec/test_frozen_kv_mtp.py registered/unit/spec/test_eagle_worker_v2_topk1_fastpath.py:

🚀 1-gpu-5090 (3 tests): ✅ View workflow run

cd test/ && python3 registered/spec/eagle/test_spec_eagle.py
cd test/ && python3 registered/spec/eagle/test_spec_eagle_topk_page.py
cd test/ && python3 registered/unit/spec/test_eagle_worker_v2_topk1_fastpath.py

🚀 1-gpu-h100 (3 tests): ✅ View workflow run

cd test/ && python3 registered/spec/eagle/test_eagle_constrained_decoding.py
cd test/ && python3 registered/spec/test_spec_standalone.py
cd test/ && python3 registered/spec/test_frozen_kv_mtp.py

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

cd test/ && python3 registered/spec/eagle/test_eagle_reject_sampling.py
cd test/ && python3 registered/spec/test_gemma4_mtp_31b_extra.py

[hnyls2002]

/tag-and-rerun-ci

[hnyls2002]

/rerun-test registered/models_e2e/test_mimo_v2.py registered/models_e2e/test_deepseek_v3_mtp.py registered/models_e2e/test_dsa_glm52_tp_mtp.py registered/models_e2e/test_qwen3_next_models_mtp.py registered/spec/eagle/test_eagle_dp_attention.py registered/spec/eagle/test_adaptive_speculative.py

[github-actions]

Results for /rerun-test registered/models_e2e/test_mimo_v2.py registered/models_e2e/test_deepseek_v3_mtp.py registered/models_e2e/test_dsa_glm52_tp_mtp.py registered/models_e2e/test_qwen3_next_models_mtp.py registered/spec/eagle/test_eagle_dp_attention.py registered/spec/eagle/test_adaptive_speculative.py:

🚀 8-gpu-h200 (3 tests): ✅ View workflow run

cd test/ && python3 registered/models_e2e/test_mimo_v2.py
cd test/ && python3 registered/models_e2e/test_deepseek_v3_mtp.py
cd test/ && python3 registered/models_e2e/test_dsa_glm52_tp_mtp.py

🚀 4-gpu-h100 (2 tests): ✅ View workflow run

cd test/ && python3 registered/models_e2e/test_qwen3_next_models_mtp.py
cd test/ && python3 registered/spec/eagle/test_eagle_dp_attention.py

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

cd test/ && python3 registered/spec/eagle/test_adaptive_speculative.py