feat(qwen35): add transactional speculative verifier

[CAICAIIs]

Summary

This PR is the target-verifier slice split out of #626. It makes Qwen3.5 speculative verification reviewable independently from the draft model and serving integration.

It verifies [current token, draft tokens...] in one target forward pass, then commits the longest greedy-matching prefix plus one target token. Qwen3.5 hybrid state is handled as one transaction: paged KV, recurrent state, convolution state, and CUDA Graph slot state.

What changed

• Adds batched target verification and reusable verification buffers.
• Adds commit, partial replay, and rollback for Qwen3.5 hybrid state.
• Uses checked projection calls so launch failures reach rollback.
• Covers accept-all, reject-first, prefix acceptance, KV page boundaries, injected failures, and replay of a graph captured before the transaction.
• Keeps the verifier on one CUDA stream and rejects stream overrides before mutation.

Validation

On one RTX 5090 with Qwen3.5-4B:

• Qwen3.5 feature check and Clippy passed.
• Speculative verifier tests: 11 passed.
• HF golden gate: 2 passed, including the 4097/8192-token fixtures.
• Scheduler E2E: 1 passed.
• Local formatting and diff checks passed.

Scope

This PR does not add the DFlash draft model, scheduler/server opt-in wiring, sampling support, or a performance claim. Those parts remain in follow-up PRs split from #626.

[CAICAIIs]

Follow-up split plan:

• Drafter PR: add DFlash draft-model loading and draft-forward correctness gates on top of this verifier.
• Serving integration PR: wire the opt-in scheduler/server flow, explicit fallback and admission rules, then add same-host benchmark and profile evidence.

Each PR will keep the default Qwen3.5 path unchanged unless DFlash is enabled.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c65ada7ecd

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Allow single-token verify spans at the budget edge

When a speculative step is clamped to the last remaining output token, the verify span can legitimately contain only the current token; the existing speculative scheduler does exactly this with token_ids.truncate(remaining) when remaining == 1 (openinfer-qwen3/src/scheduler/plan.rs:188-198). A one-token target verify should consume the current token and return the model's single posterior token, and the rest of this verifier/commit path already handles span.len() == 1; this guard would instead turn the final speculative step into an error when Qwen3.5 is wired into the same flow.

Useful? React with 👍 / 👎.