feat(client): pass through sampling-mask logprobs

[faresobeid]

On top of #174's packed completion logprobs, passes prime-rl's sampler logprobs at the sampling-mask ids through renderers.client.generate().

• generate() returns sampling_mask_logprobs = choice.get("sampling_mask_logprobs") next to sampling_mask. prime-rl's /inference/v1/generate sends it packed: base64 float32 {data, shape, dtype}, parallel to the packed sampling_mask ids. Stock vLLM does not send it, and then the key is None.
• Pure pass-through: no request-side change, no parsing or validation (same as sampling_mask), no payload change when absent. verifiers decodes it (feat(v1): optional sampler logprobs on SamplingMask verifiers#2756).

Consumer

prime-rl score centering, PrimeIntellect-ai/prime-rl#3794 (stacked on prime-rl#3900), pins this head. Without that consumer the field is unused.

Stack

Base: #174 (perf/packed-completion-logprobs). Order: #174 -> PrimeIntellect-ai/verifiers#2778 -> prime-rl#3900; then this PR + PrimeIntellect-ai/verifiers#2756 -> prime-rl#3794. Rebased on current main (via #174).

Validation

• uv run pytest tests/test_client.py, ruff check renderers tests, ruff format --check renderers tests.

Note

Pass through sampling_mask_logprobs in client generate result

The async generation client in renderers/client.py now reads the optional sampling_mask_logprobs field from the vLLM generation choice. The server may send packed float32 log probabilities parallel to the kept sampling-mask IDs when score centering is on. Behavioral Change: generate returns None for this value when the response omits the field.

^{Macroscope summarized ecfeef7.}

---

Note

Low Risk
Optional response field pass-through with no parsing or request changes; behavior unchanged when the engine omits the key.

Overview
generate() now surfaces optional sampling_mask_logprobs from the /inference/v1/generate choice, alongside the existing sampling_mask. When prime-rl score centering is enabled, the server can send packed float32 logprobs parallel to the kept mask token ids; stock vLLM omits the field and callers see None.

This is a pass-through only: no new request fields, no decoding or validation (same pattern as sampling_mask). The inline comment documents when the server sends the payload.

^{Reviewed by Cursor Bugbot for commit ecfeef7. Bugbot is set up for automated code reviews on this repo. Configure here.}

[macroscopeapp]

Approvability

Verdict: Approved at 1cd3446

Macroscope's review found this PR approvable — This is a narrowly scoped, backwards-compatible pass-through of optional sampling-mask logprobs from the existing server response. It changes no request behavior, generation logic, or validation and adds no security, schema, or deployment risk.

Notes:

• Macroscope's correctness review did not run, so approvability was decided on eligibility alone.

No code changes detected at ecfeef7. Prior analysis still applies.

^{You can add or adjust custom eligibility rules. Learn more.}