Investigate HTTP output dependence on prefix-cache history #212

Problem

With prefix caching enabled, the 2048-token HTTP fixture changes output depending on prior cache history; without caching concurrent outputs match isolated requests on all four tested models.

Evidence

Qwen3-0.6B, Qwen3-4B and Gemma3-1B each match 8/12 concurrent responses to their prefix-enabled serial baselines. All mismatches are the2048 fixture. First divergence positions1,13,24 respectively. Gemma divergence appears after EOS and must be distinguished from a changed completed answer. Prefix reuse also changes prefill chunk boundaries; numerical effects versus cache corruption are not yet isolated.

Scope

Teacher-force equal token histories and compare cold/partially warm/fully warm prefix paths in BF16 and FP32. Inspect page ownership and cached tensor integrity if differences survive the numerical control. Fix a demonstrated implementation error; do not claim all mismatches are numerical from token similarity alone.

Done when

Controlled evidence identifies the cause for the observed fixtures, regression tests protect any fix, and HTTP cache behavior/limitations are recorded alongside #211/#201.

Owner

Codex investigation; maintainer review.