Prepare 0.11.0 release

Cargo.toml

@@ -16,7 +16,7 @@ opt-level = 3
 
 [package]
 name = "grill"
-version = "0.10.0"
+version = "0.11.0"
 edition = "2024"
 license = "Apache-2.0"
 publish = false

crates/grill-perf/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "grill-perf"
-version = "0.10.0"
+version = "0.11.0"
 edition = "2024"
 license = "Apache-2.0"
 publish = false

docs/RELEASES.md

@@ -3,6 +3,7 @@
 Release preparation is not publication. No tag or release is created by the
 staging script, pull-request workflow, manual staging dispatch or installed
 smoke helper. Maintainer approval is required before publication.
+Version `0.11.0` is in preparation, not a claim that a public download exists.
 
 Staging artifacts are retained on the Actions run for reviewers with access,
 subject to repository authentication and artifact retention. They are not

docs/performance/INSTALL.md

@@ -16,6 +16,12 @@ earlier published releases and their original pins for historical studies.
 Installing v0.10.0 does not reinterpret old receipts or qualify new backend
 claims.
 
+The speech-server media runner ships from
+[v0.11.0](RELEASE-NOTES.md#version-0110--speech-servers).
+That release is in preparation; the download commands remain pinned to the
+published release above. Use a [source build](#upgrade-rollback-and-source-fallback)
+containing these changes until publication.
+
 Since 0.2.0, packages also carry the claim map and report template under
 `docs/performance/`, public domain examples under `workloads/`, and the explicit
 startup, retention and microbench Python modules under `tools/`. Keep all of

docs/performance/RELEASE-NOTES.md

@@ -1,14 +1,28 @@
 # Performance release notes
 
-## Unreleased
+## Version 0.11.0 — speech servers
 
 - Add a `media-speech-server-v1` runner to `grill-perf media` for an
   already-running OpenAI-style `POST /v1/audio/speech` server: serial requests,
   streamed `first_audio_us`, retained reply bodies and headers rechecked offline,
   a `first_audio_us` gate and a two-sided `duration_ratio` gate against the
   reference arm. See [Speech servers](MEDIA.md#speech-servers).
-- Verification is limited to a loopback fixture server; no live speech server
-  was measured.
+
+Verification and limits of this version:
+
+- The speech runner was tested only against a loopback fixture server in the
+  test suite, not a live speech server.
+- A reply without an `X-Finish-Reason` header has an unknown finish reason: the
+  sample stays eligible and inspect and compare count it as
+  `finish_reason_unknown`. Only servers that send the header get the
+  `finish_not_stop` check.
+- Compatibility check: a WAV audio capture with runner markers made with the
+  published 0.10.0 Linux ARM archive inspected successfully with 0.11.0,
+  preserving its evidence digest, all three eligible measured samples and valid
+  markers; the new speech summaries are null. Three such 0.10.0 captures gave
+  identical gate outcomes under both versions (inconclusive: reference spread).
+  This synthetic check does not qualify every historical capture or
+  cross-version comparison.
 
 ## Version 0.10.0 — audio, video and runner markers
 

licenses/NOTICE-INPUTS.json

@@ -1,8 +1,8 @@
 {
-  "cargo_lock_sha256": "54d15993cdef45480b4b94013d35e241f410c463edcc366ba583db2dd275e7ea",
+  "cargo_lock_sha256": "98e8e77d419daffad29e3fcbb7dd6f5b137d5daa02076451d9791db212976543",
   "manifest_sha256": {
-    "Cargo.toml": "0d759710bc732b11f2f43a91e572090736e06f1dc33ad5104c66b527373d618e",
-    "crates/grill-perf/Cargo.toml": "f8cf20d1e70acb90343fa5abba645b863b9a49d128f9a1e66425a327ccce874a",
+    "Cargo.toml": "d42fce6074b77852935ba2d9b82ea4f419f619a27c0aa8a815fb7bcb50c55a5d",
+    "crates/grill-perf/Cargo.toml": "10efcb4bf626a6bef07acf0bdbd7e508105e15f958e28cd7a000b6b734d82e14",
     "crates/grill-sse/Cargo.toml": "5f6d34630cd32f54aae216f1b5996886e4865343a3d1421e4053ba9c21ccabb1"
   },
   "notice_sha256": {