release: 0.71.2

CHANGELOG.md

@@ -7,6 +7,51 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.71.2] - 2026-10-12
+
+A patch release on 0.71.1: five new RustSec advisories against quinn-proto are fixed by the upgrade, `apr-cli` builds again without the `inference` feature, the silicon-coverage read survives transient 5xx answers and a push run's token, a batch of cargo-free guards moves from the PR path to the nightly LAB lane, and 28 dark source files whose code already lived in a compiled file are deleted.
+
+### Security
+
+- fix(deps): quinn-proto 0.11.15 -> 0.11.18 for RUSTSEC-2026-0338..0342 (#5236)
+- fix(security): interim ignore for RUSTSEC-2024-0449 (unsafe-libyaml unmaintained) — no patched release exists (Refs #5205) (#5206)
+
+### Fixed
+
+- fix(apr-cli): build without the inference feature (#4972) (#5167)
+- fix(serve): size minimal llama fixture weights as [out, in] rows (Refs #4985) (#5186)
+- fix(serve): ALiBi slope doc says 2^(-8(h+1)/n); delete the dark layers/a_li_bi.rs (#4700) (#5168)
+- fix(silicon-coverage): retry a 5xx read (cap 3, failure stays not_measured); a 403 body is not a runner (#5220)
+- fix(silicon-coverage): push runs read with the job token too; probe runners only where the listing is readable (#5237) (#5239)
+- test(train): the models row test pins the row #5142 made, a dispatch is a night (#5166)
+
+### Changed
+
+- LAB guard batch: #5025 #5047 #5010 #5011 #5048 #5034 #5064 run nightly, off the PR path (#5223)
+- chore(src): delete 27 dark src files whose code a compiled file holds (#4700 batch 1) (#5163)
+
+### Known issues
+- **Qwen3 MoE (Qwen3-30B-A3B, Qwen3-Coder-30B-A3B) on GB10 is not claimed for 0.70 (#3715); planned for 0.70.1.**
+  On GB10 (sm_121) every verb (run, chat, serve, code) times out at 600 s on the 20k-token rung. The 0.69.1
+  pre-load refusal is gone in 0.70.0, so these models now load and run, but too slowly to be claimed. The
+  release gate does not owe the qwen3moe x GB10 cells (`cells.declaimed` in
+  `contracts/model-capability-ladder-v1.yaml`), and the ladder gate is RED if any rung or long-rung
+  representative still claims them. Withdrawn, not waived: lambda holds both files and still owes every
+  qwen3moe cell.
+- **GB10 dense models: prompt processing unbatched; planned for 0.70.1 (#4590).** Widens the Qwen3 entry below to every dense model on GB10 (operator ruling 2026-09-28
+  16:15Z): the serial prefill is the sm_12x default for every non-hybrid arch, so dense Qwen2.5 on GB10
+  is de-claimed too (`cells.declaimed`: qwen2 and qwen3 on gx10). Output is correct; only prefill speed
+  is affected. Withdrawn, not waived: lambda still owes every dense cell, gx10 still owes every Qwen3.5 cell.
+- **Qwen3 on GB10 (sm_121, and the sm_12x family) is correct but very slow, and is not claimed for
+  0.70 (#4590).** Dense Qwen3 on sm_12x prefills one token at a time, the sm_12x default, because the
+  generic batched prefill corrupts the KV cache there (FALSIFY-CPU-GPU-009). Measured on 0.70.0
+  (8d021f61e) at an 8k think-on prompt: Qwen3-1.7B 967 s, Qwen3-8B 2163 s, against 63-80 s for
+  Qwen3.5-9B on the same host (Qwen3.5 has its own batched prefill, #3596). `apr` prints a
+  `[KNOWN-ISSUE #4590]` warning when it loads Qwen3 on sm_12x. The release gate does not owe the
+  Qwen3 x GB10 cells (`cells.declaimed` in `contracts/model-capability-ladder-v1.yaml`); every other
+  host still owes every Qwen3 cell. Planned for 0.70.1: native sm_121 SASS, then a re-test of batched
+  prefill on GB10.
+
 ## [0.71.1] - 2026-10-12
 
 A patch release on 0.71.0: the published aprender-core crate compiles again, a tag push runs the full workspace, the CLI-tools bin has its surface row, the CRUX judge's two new definitions are back at grade B, and under the standing release policy the pre-publish dogfood (R5) leaves the release path the way readiness (R8) did, with CRUX smoke (R7) the release gate.