fix: keep bounded sampling masks sparse

[mikasenghaas]

Summary

• keep sampling masks with at most 512 IDs sparse during device-to-host transfer
• preserve vLLM's scheduler and response formats
• fall back to vLLM's full-vocabulary bitset for wider masks
• restore the bounded sampling-mask optimization from feat: top-p/top-k train sampling with sampling replay #3235 on current main

Verification

• uv run --no-sync pytest -q tests/unit/inference (3 passed)
• uv run --no-sync ruff check src/prime_rl/inference/patches.py
• uv run --no-sync ruff format --check src/prime_rl/inference/patches.py

The paired live-training benchmark ran 10 steps per revision with the same
Qwen3-0.6B model, Reverse Text environment, V2 runner, sampling replay
(top_p = 0.95, bounded to top_k = 512), nodes, and warm caches. main ran
first and this PR ran immediately afterward. Request- and token-normalized
metrics account for the different trajectories sampled during live training.

Metric	main	This PR	Change
Complete 10-step orchestrator loop	58.3 s	52.9 s	-9.3%
Mean vLLM request inference time	890 ms	354 ms	-60.2% / 2.51x faster
vLLM decode time per generated token	9.72 ms	3.95 ms	-59.4% / 2.46x faster
Mean agent model-call time, steps 2-10	837 ms	391 ms	-53.2% / 2.14x faster

• main W&B run
• PR W&B run

A targeted batch-384 sampling-mask benchmark reduced device-to-host data from
6.94 MiB to 0.75 MiB and CPU mask materialization from 211 ms to 1.20 ms. This
isolates the regression to vLLM's dense vocabulary-width bitset conversion;
the HTTP payload, router, environment server, and trainer are downstream of
that work.

---

Note

Medium Risk
Monkey-patches core vLLM v1 sampler/mask conversion on every worker; incorrect sparse logic could break sampling replay or mask payloads, though behavior is gated to bounded top-k and falls back to upstream.

Overview
Restores a sparse path for vLLM sampling masks when return_sampling_mask is on and effective top_k is at most 512, avoiding the dense full-vocabulary bitset that regressed device-to-host transfer and CPU materialization.

A new monkey_patch_bounded_sampling_mask() is wired into apply_shared_vllm_patches(). It wraps Sampler.__call__ (via a ContextVar) and replaces SamplingMaskTensors.from_logits to build SparseSamplingMaskTensors from a bounded topk over logits, with to_cpu_nonblocking / tolists still emitting vLLM’s SamplingMaskLists so scheduler and response formats stay unchanged. Wider masks or disabled sparse mode fall back to stock vLLM behavior; the patch is idempotent via _prime_rl_uses_sparse_sampling_masks.

^{Reviewed by Cursor Bugbot for commit ae64d7f. Bugbot is set up for automated code reviews on this repo. Configure here.}