fix(zmq): enforce wire constraints on TokenSpeed direct-backend workers

[slin1237]

Problem

TokenSpeed defaults grammar_backend to none and then drops structured-output constraints without complaint. The gateway does its part correctly — apply_tokenspeed_constraint maps all four constraint kinds onto the wire — so the request arrives intact at an engine that quietly ignores it.

The user-visible result is that tool_choice=required, tool_choice={function} and response_format=json_schema return freeform text with a 200. Output logprobs default off the same way, so logprobs=true returns null.

This is a production hole, not a test-harness one. The gRPC servicer path passes --grammar-backend xgrammar and --enable-output-logprobs explicitly, with a comment naming this exact failure mode; the smg serve ZMQ launcher never did. Anyone running TokenSpeed over ZMQ today gets silently unenforced constraints.

The e2e lane is just where it surfaced: on run 31243936303 the e2e-1gpu-chat-zmq (tokenspeed) lane logged grammar_backend='none' at startup and produced 44 failures — every one a grammar test (38 function-calling required/specific, 4 json_schema, 2 regex). tool_choice=auto passed throughout, so parsing was never the problem.

Change

TokenspeedWorkerLauncher now supplies both flags as defaults rather than launcher-owned flags, via a new WorkerLauncher._backend_arg_defaults helper. The distinction matters: _filter_backend_args strips flags SMG owns outright, whereas these are engine defaults SMG needs for the wire to behave as the API promises but an operator may legitimately override. Pass --grammar-backend llguidance and you keep yours; the other defaults stay.

The e2e harness delegates to this launcher (_build_tokenspeed_zmq_cmd), so the ZMQ lane picks the fix up with no harness change.

What this does not fix

Failing loud would be better than defaulting. The frontend cannot do that today: EngineCoreReadyResponse carries no grammar-backend field, so it cannot distinguish an unenforceable constraint from an enforceable one. That needs a handshake addition on the TokenSpeed side (#48 territory), and is worth doing — a silently-ignored constraint is a correctness bug that looks like a model quality problem.

Test Plan

• pytest tests/test_serve.py -k Tokenspeed → 11 passed, including three new cases: defaults are applied, an operator override wins and leaves other defaults intact, and the --flag=value form is honored.
• ruff check / ruff format --check clean.
• The real proof is the e2e-1gpu-chat-zmq (tokenspeed) lane on this PR: those 44 failures should clear.

Checklist

• cargo +nightly fmt passes (no Rust changes)
• cargo clippy --all-targets --all-features -- -D warnings passes (no Rust changes)
• (Optional) Documentation updated
• (Optional) Please join us on Slack #sig-smg to discuss, review, and merge PRs

[coderabbitai]

📝 Walkthrough

Summary by CodeRabbit

• New Features

  • TokenSpeed workers now automatically enable grammar constraints and output log probabilities by default.
  • Backend settings supplied by operators take precedence over these defaults, including --key=value options.

• Bug Fixes

  • Prevented launcher defaults from overriding explicitly configured backend arguments.

Walkthrough

TokenSpeed workers now receive default xgrammar and output-logprob backend arguments. The launcher preserves explicit flag and --key=value overrides. Tests cover defaults and override behavior.

Changes

TokenSpeed backend defaults

Layer / File(s)	Summary
Backend default resolution bindings/python/src/smg/serve.py	WorkerLauncher._backend_arg_defaults adds missing backend flag-token sequences and preserves supplied flags.
TokenSpeed command integration and validation bindings/python/src/smg/serve.py, bindings/python/tests/test_serve.py	TokenSpeed adds --grammar-backend xgrammar and --enable-output-logprobs by default. Tests cover explicit values and equals-form arguments.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Possibly related PRs

• smg-project/smg#460: Adds the backend-argument forwarding that this change extends with defaults and override handling.

Suggested reviewers: catherinesue, key4ng

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly identifies the TokenSpeed direct-backend wire-constraint fix, which matches the main change.
Description check	✅ Passed	The description clearly explains the TokenSpeed constraint and logprob issue, implementation, rationale, and test coverage.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch zmq/tokenspeed-grammar-backend

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

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

[claude]

👋 The PR description doesn't fully follow
PULL_REQUEST_TEMPLATE.md:

• Missing header: ## Description
• Missing header: ### Problem (has ## Problem instead — should be a ### subsection under ## Description)
• Missing header: ### Solution (has ## Change instead — should be ### Solution under ## Description)
• Missing header: ## Changes (has ## Change instead — should be ## Changes, plural)

Please update the PR description so reviewers have the context they need.

[claude]

Clean, well-motivated fix. The _backend_arg_defaults helper correctly separates overridable defaults from launcher-owned flags stripped by _filter_backend_args. The three test cases cover the key scenarios (defaults applied, operator override wins, equals-form honored). No issues found.

[coderabbitai]

🧹 Nitpick comments (1)

bindings/python/tests/test_serve.py (1)

873-881: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

🟡 Nit Use a non-default value for the override test.

After the default changes to llguidance, pass none here. This verifies that a different operator value is preserved and that the default is not duplicated.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@bindings/python/tests/test_serve.py` around lines 873 - 881, Update
test_build_zmq_command_defaults_yield_to_the_operator to pass the non-default
grammar backend value "none" in the build_command arguments, then assert that
"none" is preserved while confirming --grammar-backend appears only once and the
other default remains.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Nitpick comments:
In `@bindings/python/tests/test_serve.py`:
- Around line 873-881: Update
test_build_zmq_command_defaults_yield_to_the_operator to pass the non-default
grammar backend value "none" in the build_command arguments, then assert that
"none" is preserved while confirming --grammar-backend appears only once and the
other default remains.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: bb5576de-30f1-4021-8103-0652c25e6c7e

📥 Commits

Reviewing files that changed from the base of the PR and between c3a8287 and f3659de.

📒 Files selected for processing (2)

• bindings/python/src/smg/serve.py
• bindings/python/tests/test_serve.py