Expose Torch PD producer cache and chunk controls for reproducible serving #241

Problem

Torch PD always enables producer prefix caching, even when decoder prefix caching is off. The serving CLI cannot request a genuinely uncached producer for reproducibility or controlled benchmarks.

Evidence

#240: Qwen3-0.6B GPU PD BF16 concurrent output matches serial8/12 with35072producer cache-hit tokens. A process-local cache-off control yields12/12 with zero producer hits; FP32 with cache enabled also yields12/12. The default may remain useful for throughput, but its policy must be explicit and controllable.

Scope

Add an explicit opt-out flag for the in-process Torch prefill producer, preserving the existing default and decoder policy. Forward it through the HTTP validation harness. Add actual tiny-model factory tests verifying repeated requests reuse prefixes by default and do not reuse them when disabled. No automatic accuracy/performance claim and no change to vendor/subprocess engines.

Done when

The public CLI reproduces the cache-off GPU control, existing defaults remain intact, and tests verify actual prefill cache behavior.

Owner

Codex implementation; maintainer review/merge.

Measured scope update

The producer chunk default2048 also differs from standalone chunk256. Qwen3-0.6B BF16 cache-off PD still differs from native at2K; matching producer chunk256 restores all3 native/PD serial comparisons. Expose --pd-prefill-chunk with the existing default2048 and reject nonpositive sizes before loading a model. Verify actual chunk boundaries with a tiny real model. This keeps producer reproducibility settings explicit without changing defaults.