fix: map dense subprocess KV into hybrid GPU cache groups #258

Problem

Live RNGD-to-GPU Gemma3-1B fails during first KV admission: subprocess prefill returns a dense model-layer tensor, but PagedKV expects one tensor per attention group. The owner exits with a shape mismatch instead of serving the request.

Evidence

Runtime3e16c33, Gemma3-1B BF16, four GPU slots/context9216/page64/cache off; RNGD width8192. Actual first128-token request fails in PDBatcher._admit_prefilled -> PagedKV.write_span: value [26,1,128,256] cannot broadcast to destination [22,1,128,256]. Gemma has22 sliding and4 global layers. Native GPU and grouped Torch producer paths pass existing matrices. Raw failure retained as local201-gemma-matrix-a.json and server log; no completed performance claim.

Scope

Map validated dense model-layer KV spans to receiver groups without changing wire protocol. Preserve existing grouped spans. Reject invalid shape/group input before any partial write. Add focused regression and real tiny Gemma PD coverage, then rerun hardware.

Done when

Dense and grouped imports yield identical per-layer cache contents and decode outputs; invalid input is rejected before writes; existing non-hybrid and grouped paths pass; live Gemma RNGD/GPU reaches and completes its matrix. Numerical parity is a separate measured gate, not presumed fixed.

Owner

Codex; maintainer review. Related #201, #231, #248.