## avifenesh/memra — v0.94.0…v0.95.0

_18 commits._

### Fixes
- **fix(gemma): the assistant-drafter attach line must NAME the artifact, or the drafter-attach assertion cannot be satisfied** (fda0408)

### Backend
- **boundary: re-pin docs/SERVING.md — the train edited it, and the pins are sha-keyed** (7abc7c9)
- **release: v0.95.0 — sampled-spec stack (version + internal pins)** (46adb2e)
- **merge: lane/sampled-restore-load-guard-20260819 @ 010869175e — SOLO admission for sampled restore, which is what makes the two lanes above safe to ship: the measured crossover sits between c1 and c2 (1.349x at c1, 0.668x at c2, worst 0.477x at c3, 0.813x at c16), so the watermark is SOLO and inert at c1 to 0.1%. Also: boundary-probe default 1.5 -> 4.0, the gemma runbook rewrite, and the MEMRA_DRAFT / MEMRA_MTP_DRAFT seam documentation. Fast-forwarded over the gate-tool branch 8297da4ca7.** (715fae9)
- **merge: lane/sampled-spec-quality-20260819 @ 9e3fc3d154 — three pre-existing sampled-path defects: SIX boundary sites emitted argmax inside sampled streams (measured 1.71% of a customer's tokens drawn from the wrong distribution at c=1, 1 in 58, concentrated in the low-concurrency interactive tier), penalty history was truncated to the burst instead of spanning the session, and restored sessions never republished an extended entry. Penalized-restore refusal lifted as a consequence.** (9254cc3)
- **merge: lane/sampled-hit-spec-20260819 @ 98924898b0 — spec re-arms on SAMPLED prefix-cache hits (v0.93.0 shipped the feature greedy-only by v1 design, and live DE hits are all sampled customer traffic, so the headline was inert for paying traffic); pen_hist seeded from the burst is handled** (48df4e8)
- **gate(mixed): the RESTORED PREFIX LENGTH is part of the program, so a grown-entry pair is not comparable either** (0108691)
- **gate(mixed): the coexistence engagement assertion is POLICY-KEYED, because SOLO made it unsatisfiable** (a1a7c86)
- **guard v3: the watermark is SOLO, because the measured crossover is between c1 and c2** (465cfa5)
- **guard v2: read DEMAND where it is knowable (the HTTP in-flight gauge), and re-evaluate as late as possible** (da1b428)
- **spec: load-aware sampled-restore admission (the 19%-at-c16 door) + a powered boundary probe + the MEMRA_DRAFT seam trap** (6d2f334)
- **gate(mixed): same-path guard on the sampled reproducibility pair** (8297da4)
- **gate(sampled-spec): the three cells the 96GB window needed** (6e8e917)
- **data: full release battery on the sampled-spec-quality lane — ALL GREEN, 0 fail / 0 warn** (9e3fc3d)
- **spec: sampled boundary tokens are SAMPLED, penalty history spans the SESSION, restored sessions republish their boundary** (a6a767a)
- **data: full release battery on the sampled-hit lane — ALL GREEN, 0 fail / 0 warn** (9892489)
- **spec: sampled prefix-cache hits re-arm spec (v0.93.0's headline was inert for API-default traffic)** (fd54815)

### Docs
- **docs: v0.95.0 truth pass — the two surfaces this train made stale** (7e8f705)

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