Add MTP anchored ngram speculation

[i386]

Summary

Adds a clean Skippy protocol break for mandatory client-open stage streams and implements server-native n-gram speculative proposals on the embedded stage-0 OpenAI path. The new path can use GLM native MTP as the first-token anchor and n-gram matching as a cheap span extender before distributed batched verification.

This PR is based on current main from a clean worktree. It does not include the discarded SPD/shard sidecar work.

What changed

• Replaces the old server-ready/client-ready compatibility handshake with mandatory SOPN client-open framing and bumps the stage state version to 7.
• Threads --openai-ngram-speculative, --openai-ngram-size, and --openai-ngram-min-match through serve-binary, embedded OpenAI args, and host-runtime Skippy resolution.
• Adds the native native-mtp+ngram proposal path and telemetry labels for proposal source and MTP anchor verification.
• Updates skippy-bench, skippy-correctness, and skippy-prompt callers to use the new stage-open protocol.
• Adds a distributed skippy-bench run --driver open-ai-chat path that starts stage0 embedded OpenAI serving and drives /v1/chat/completions, so MTP+n-gram can be measured against the Improve GLM MTP parity and batched verify sampling #858 rows.
• Adds explicit benchmark flags for native MTP sideband and OpenAI speculation controls instead of relying on ad hoc shell env.

Validation

Clean-main branch validation:

• cargo fmt --all --check
• cargo test -p skippy-protocol --lib
• cargo test -p skippy-server --lib
• cargo test -p skippy-bench
• cargo test -p skippy-correctness
• cargo test -p skippy-prompt
• cargo test -p mesh-llm-host-runtime --lib inference::skippy
• cargo clippy -p skippy-protocol --all-targets -- -D warnings
• cargo clippy -p skippy-server --all-targets -- -D warnings
• cargo clippy -p skippy-bench --all-targets -- -D warnings
• cargo clippy -p skippy-correctness --all-targets -- -D warnings
• cargo clippy -p skippy-prompt --all-targets -- -D warnings
• cargo clippy -p mesh-llm-host-runtime --all-targets -- -D warnings
• cargo clippy -p mesh-llm --all-targets -- -D warnings
• cargo check -p mesh-llm

Benchmark path

The new benchmark row for this PR should be run with:

skippy-bench run \
  --driver open-ai-chat \
  --stage-native-mtp \
  --openai-speculative-window <N> \
  --openai-ngram-speculative \
  ...

Run it against the same GLM 4.7 MTP split/model/prompt corpus used for #858. Generated-token tok/s in driver-result.json is the hard metric; acceptance remains diagnostic telemetry.

Notes

This PR intentionally breaks Skippy wire compatibility and removes the old optional compatibility path. Lab tok/s results still need to be collected on this branch before moving the PR out of draft.

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 6302844e-e1f7-44a9-bcdd-a5a3d8c118a9

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch jd/jianyang-mtp-ngram

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}