diff --git a/CITATION.cff b/CITATION.cff
index 087da5e0..5387b078 100644
--- a/CITATION.cff
+++ b/CITATION.cff
@@ -11,7 +11,7 @@ authors:
     given-names: Silvio
 repository-code: "https://github.com/alpsaur/MinkowskiEngine"
 url: "https://github.com/alpsaur/MinkowskiEngine"
-version: "0.5.7"
+version: "0.5.8"
 date-released: "2026-07-17"
 license: MIT
 keywords:
diff --git a/MinkowskiEngine/__init__.py b/MinkowskiEngine/__init__.py
index 8f8ba7e1..b324a5e9 100644
--- a/MinkowskiEngine/__init__.py
+++ b/MinkowskiEngine/__init__.py
@@ -22,7 +22,7 @@
 # Please cite "4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural
 # Networks", CVPR'19 (https://arxiv.org/abs/1904.08755) if you use any part
 # of the code.
-__version__ = "0.5.7"
+__version__ = "0.5.8"
 
 import os
 import sys
diff --git a/README.md b/README.md
index bf31ebe9..096c8e03 100644
--- a/README.md
+++ b/README.md
@@ -24,13 +24,13 @@
    - See the [performance guide](docs/performance.md): combined with TF32, a real training pipeline measured ~2x faster per step at ~25% less memory vs stock-ME fp32
 3. **It's tested and installable**: [prebuilt wheels](https://github.com/alpsaur/MinkowskiEngine/releases/latest) (one `pip install`, no compile), CI on every change, a repaired test suite, working Docker, [hosted docs](https://alpsaur.github.io/MinkowskiEngine/), and built-in `torch.profiler` ranges (`ME::*`) for diagnosing your own workloads.
 
-### Quick Install (v0.5.7)
+### Quick Install (v0.5.8)
 
 **Prebuilt wheels** (fastest, no compilation; CUDA 12.8 / sm_120, built against **torch 2.9.x cu128**, Python 3.10–3.12):
 
 ```bash
 # pick the wheel matching your Python from the release page
-pip install https://github.com/alpsaur/MinkowskiEngine/releases/download/v0.5.7/minkowskiengine-0.5.7-cp312-cp312-linux_x86_64.whl
+pip install https://github.com/alpsaur/MinkowskiEngine/releases/download/v0.5.8/minkowskiengine-0.5.8-cp312-cp312-linux_x86_64.whl
 ```
 
 All wheels: [Releases](https://github.com/alpsaur/MinkowskiEngine/releases/latest). The wheel dynamically links your torch install, so it needs a **torch 2.9.x + cu128** runtime; for other torch lines, build from source instead.
@@ -50,6 +50,14 @@ docker build -t minkowski_engine docker
 
 The build no longer wipes `build/` on every run, `python setup.py build_ext --inplace` works out of the box (a `MinkowskiEngineBackend/` namespace stub is included), and a minimal `pyproject.toml` is provided for PEP 517 installs. `FORCE_CUDA=1` (env var) forces a CUDA build on machines without a visible GPU.
 
+### New in v0.5.8
+
+**Correct interpolation at sparse boundaries.** `MinkowskiInterpolation` / `features_at_coordinates` renormalizes weights over the lattice corners that actually exist, fixing the long-standing upstream bug ([#477](https://github.com/NVIDIA/MinkowskiEngine/issues/477)) where queries next to valid data decayed toward zero. Exact-on-voxel and full-neighborhood results are bit-identical to before; only previously-broken boundary queries change.
+
+**Opt-in deterministic convolution.** `ME.set_deterministic(True)` sorts conv inputs into a canonical coordinate order so outputs are reproducible run-to-run regardless of input row ordering (upstream [#554](https://github.com/NVIDIA/MinkowskiEngine/issues/554)). Off by default; costs one sort per conv.
+
+Also: regression tests locking in empty-pruning behavior ([#579](https://github.com/NVIDIA/MinkowskiEngine/issues/579)) and TensorField backprop ([#395](https://github.com/NVIDIA/MinkowskiEngine/issues/395)); a new [operations cookbook](docs/operations.md) (memory, DDP, torch.compile, determinism).
+
 ### New in v0.5.7
 
 **Fused + vectorized gather/scatter.** The copy-GEMM convolution path used to launch tiny gather/scatter kernels once per kernel offset (27x per layer for a 3^3 kernel); profiling a real sparse U-Net showed these copies eating **56% of total GPU time** across ~358k launches per 50 steps. They are now a single vectorized gather and a single vectorized scatter-accumulate per direction (16/8/4-byte chunks per thread), and the backward input-gradient GEMM is reoriented so its scatter is coalesced. Measured on the same workload: copy-kernel GPU time **5.9s → 1.7s**, kernel launches **358k → 14k**, total GPU time per step **−38%**. Enabled by default; `ME_FUSED_COPY=0` restores the legacy path, `ME_FUSED_COPY_MAX_MB` caps staging memory (default 2048).
@@ -91,6 +99,8 @@ The Minkowski Engine is an auto-differentiation library for sparse tensors. It s
 
 ## News
 
+- 2026-07 **v0.5.8: correctness release.** Boundary interpolation fixed (upstream #477, open since 2022), opt-in deterministic convolution (#554), regression tests for #579/#395, and an operations cookbook (memory, DDP, torch.compile).
+
 - 2026-07 **v0.5.7: fused gather/scatter, 38% less GPU time per training step.** Profiling showed the per-kernel-offset copy kernels were the engine's biggest cost (56% of GPU time); they are now fused and vectorized. On by default with an opt-out (`ME_FUSED_COPY=0`).
 - 2026-07 **v0.5.6 released: [prebuilt wheels](https://github.com/alpsaur/MinkowskiEngine/releases/tag/v0.5.6), half precision, faster training.** One-line install for Blackwell GPUs (no more compiling from source). fp16/bf16 now work across all ops with tensor-core GEMMs and `torch.autocast`; in our training runs this cut peak GPU memory by ~20%. An opt-in `ME_LAZY_SYNC=1` flag removes redundant GPU synchronizations for ~7–11% faster training steps.
 - 2026-07 **v0.5.5, project infrastructure**: CI (build + tests on every change), working `pip install` / Docker / [docs](https://alpsaur.github.io/MinkowskiEngine/), a repaired test suite, and packaging/build-system fixes throughout.
diff --git a/tests/conftest.py b/tests/conftest.py
index fab71b0f..dfc55dbf 100644
--- a/tests/conftest.py
+++ b/tests/conftest.py
@@ -120,6 +120,12 @@ def _gradcheck_compat(*args, **kwargs):
         "non-differentiable extras (kernel maps) that modern gradcheck's "
         "numerical-jacobian filtering no longer aligns with analytical outputs"
     ),
+    "tests/python/interpolation.py::TestInterpolation::test_gpu": (
+        "same torch 2.x gradcheck IndexError as ::test (non-differentiable "
+        "extras); pre-existing, unrelated to the renormalization fix -- "
+        "verified failing identically on the pre-fix build. Gradient "
+        "correctness is covered by interpolation_renorm.py gradcheck tests"
+    ),
 }