## avifenesh/memra — v0.98.0…v0.99.0

_12 commits._

### Fixes
- **perf(serve): batch the per-row filtered device-sample stats — sampled c8 agg 478 -> ~666 tok/s** (ecf9cab)
- **perf(serve): message-boundary prefix seed — cold chat prompts capture at the first-message end** (afbe56f)
- **perf(moe): admit NVFP4 experts to the f16g grouped prefill GEMM — pp14715 4.63x (2331->10785 tok/s)** (4ec1cf2)
- **perf(moe): NVFP4 CSR owner-scan gate_up kernel (lane/moebatch-q35moe)** (9e4d158)

### Backend
- **release: v0.99.0 — ornith15 realistic-cell campaign (NVFP4 grouped prefill 4.63x, message-boundary prefix seed, batched filtered device sampling); version + internal pins** (73d79a3)
- **data: perf-ci rows for the moebatch merge gate (--perf-quick, 0 fail 0 warn)** (257774f)
- **Merge lane/moebatch-q35moe-20260821: ornith15 realistic-cell campaign — prefill 4.63x (NVFP4 f16g admission), message-boundary prefix seed, batched filtered device sampling, CSR-NVFP4 decode dedup** (92b5125)
- **data: two-rig pre-merge battery record — both rigs ALL GREEN** (d4c78e5)
- **data: moebatch increments 3-4 receipts + cachecell scoreboard (session 34.4->20.9s, c8 151->662)** (dd315d2)
- **data: moebatch increment 2 receipts — prime anatomy (moe 88.6%), NVFP4 f16g admission 4.63x pp, gates green** (ada1a37)
- **data: moebatch increment 1 receipts — CSR-NVFP4 byte-identical (0 mismatch under =2 across run-spec), B=8 +6.8%, serve single-stream 249 = vLLM parity; aggregate increments sized** (7463990)

### Docs
- **docs: FLAGS rows for MEMRA_PRIME_ANATOMY and MEMRA_DBB_SAMP diagnostics** (ee3be83)

_Recap by [Repo Wrapped](https://repowrapped.com/gh/avifenesh/memra?utm_source=github-action)._