chore(release): v0.1.1 coarse-shard NLL MT + safe GPU init

CHANGELOG.md

@@ -12,6 +12,16 @@
   unavailable or unsupported GPU work falls back to fixed contiguous Rust CPU
   workers instead of performing respondent score aggregation in Python.
 
+## [0.1.1] - 2026-07-31
+
+### Fixed
+- Coarse fixed-shard Rust multithreading for the JML `neg_loglik_and_grad` hot path (`thread::scope` person shards, N≥256) with bit-identical reduction vs single-thread (PR #371).
+- wgpu GPU init uses `Backends::PRIMARY` only (no GL/EGL), so sandboxes with broken `/dev/dri` soft-fail to the f64 CPU path instead of SIGSEGV (PR #371).
+- Paper-grounded simple-structure MLS2PLM formula contract reaffirmed (Kang & Jeon 2025; Jeon et al. 2021); no formula drift.
+
+### Changed
+- Unit test forces multi-worker NLL shards and compares all gradient blocks to the single-thread reference.
+
 ## Unreleased
 
 ### Changed

crates/fast-mlsirm-py/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "fast-mlsirm-py"
-version = "0.1.0"
+version = "0.1.1"
 edition = "2021"
 license = "MIT"
 description = "PyO3 bindings for the fast-mlsirm Rust core."

crates/mlsirm-core/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "mlsirm-core"
-version = "0.1.0"
+version = "0.1.1"
 edition = "2021"
 license = "MIT"
 description = "Core likelihood and gradient routines for fast-mlsirm."

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "maturin"
 
 [project]
 name = "fast-mlsirm"
-version = "0.1.0"
+version = "0.1.1"
 description = "Fast simulation, fitting, and recovery diagnostics for MLSIRM/MLS2PLM models."
 readme = "README.md"
 requires-python = ">=3.10"