## avifenesh/memra — v0.96.0…v0.97.0

_11 commits._

### Fixes
- **perf(spec): lift the Qwen35Moe legacy-replay pin (lane/draftcost-moe)** (0428403)
- **perf(spec): ROUND-STREAM verify arm for the qwen35 serving class (lane/draftcost-moe, 2b)** (b0ba237)
- **perf(spec): graph-draft admission for resident-MoE heads (lane/draftcost-moe)** (d2b2fb0)
- **perf(spec): admit Qwen35Moe to the t-parallel verify arm (lane/draftcost-moe)** (209fba9)

### Backend
- **release: v0.97.0 — MoE speculative serving (version + internal pins)** (18cfac7)
- **merge lane/draftcost-moe-20260820 — MoE spec serving: t-parallel verify + graph-draft admission + replay pin lift** (bc11e10)
- **data: draftcost-moe increment 2 sealed — Qwen35Moe replay pin lifted, qualified (canary 8/8 x3 arms, long-cell identical @15.5k, run-spec PASS both arms); serve spec+pmin0.3 now BEATS plain +9-14%** (7198c90)
- **data: draftcost-moe — masked head halves the draft step (head read 270->43us), embedded stays right arm until round host-tax dies** (bf617a1)
- **data: draftcost-moe increment 2a qualified (graph draft engages for resident-MoE head, zero eager prints, run-spec PASS, acceptance identical); 2b/2c sized** (680ea2a)
- **data: draftcost-moe receipts — rowwise verify was the wall (5.6ms/tok), tparallel admission qualified (canary 8/8, run-spec PASS both arms, K=8 +92%); increment 2 = round-stream arm + graph-draft MoE admission** (4d7b1b0)
- **probe: MEMRA_SPEC_ANATOMY=1 — eager draft-step phase timers (glue/attn/ffn/head, sync-bounded, 128-step cumulative print); diagnostic seam for the draftcost lane** (bd209c5)

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