chore(release): sync v0.2.0-alpha.1 to develop

CHANGELOG.md

@@ -7,6 +7,11 @@ and published versions will follow [Semantic Versioning](https://semver.org/).
 
 ## [Unreleased]
 
+## [0.2.0-alpha.1] - 2026-09-23
+
+This alpha adds host-orchestrated GPU-selected 2-opt search. The existing PTX
+and CUDA kernel ABI are unchanged; no search-level speedup is claimed.
+
 ### Added
 
 - Host-orchestrated GPU-selected 2-opt search for one route or every route in a
@@ -49,6 +54,7 @@ and kernel compiler workspace are not published.
 - The supported input model is CVRP; Solomon time windows are parsed but not
   enforced by the current construction or local-search routines.
 
-[Unreleased]: https://github.com/pedrozaz/vrp-gpu/compare/v0.1.1-alpha...develop
+[Unreleased]: https://github.com/pedrozaz/vrp-gpu/compare/v0.2.0-alpha.1...develop
+[0.2.0-alpha.1]: https://github.com/pedrozaz/vrp-gpu/compare/v0.1.1-alpha...v0.2.0-alpha.1
 [0.1.1-alpha]: https://github.com/pedrozaz/vrp-gpu/compare/v0.1.0-alpha.1...v0.1.1-alpha
 [0.1.0-alpha.1]: https://github.com/pedrozaz/vrp-gpu/releases/tag/v0.1.0-alpha.1

Cargo.toml

@@ -8,15 +8,15 @@ members = [
 exclude = ["crates/vrp-gpu-kernel"]
 
 [workspace.package]
-version = "0.1.1-alpha"
+version = "0.2.0-alpha.1"
 edition = "2024"
 rust-version = "1.88"
 authors = ["Guilherme Pedroza"]
 license = "Apache-2.0"
 repository = "https://github.com/pedrozaz/vrp-gpu"
 
 [workspace.dependencies]
-vrp-gpu = { version = "0.1.1-alpha", path = "crates/vrp-gpu" }
+vrp-gpu = { version = "0.2.0-alpha.1", path = "crates/vrp-gpu" }
 
 [workspace.lints.rust]
 missing_docs = "warn"

README.md

@@ -5,14 +5,14 @@
 
 VRP-GPU is an experimental Rust library for capacitated vehicle routing (CVRP).
 It parses Solomon-style instances, builds routes with a greedy nearest-neighbor
-heuristic, and provides CPU 2-opt local search. The published `gpu` feature
-evaluates CUDA 2-opt candidates and selects one improving move; the development
-branch also has host-orchestrated GPU-selected search. Neither is a GPU-only
-end-to-end solver. The CPU implementation is the correctness reference.
+heuristic, and provides CPU 2-opt local search. The optional `gpu` feature
+evaluates CUDA 2-opt candidates and supports host-orchestrated, GPU-selected
+search over routes and solutions. It is not a GPU-only end-to-end solver. The
+CPU implementation is the correctness reference.
 
-> **Published alpha:** [`vrp-gpu 0.1.1-alpha`](https://crates.io/crates/vrp-gpu/0.1.1-alpha).
+> **Published alpha:** [`vrp-gpu 0.2.0-alpha.1`](https://crates.io/crates/vrp-gpu/0.2.0-alpha.1).
 > Public APIs may change before the first stable release. See the
-> [API documentation](https://docs.rs/vrp-gpu/0.1.1-alpha/vrp_gpu/) for this
+> [API documentation](https://docs.rs/vrp-gpu/0.2.0-alpha.1/vrp_gpu/) for this
 > published version.
 
 ## Workspace
@@ -34,7 +34,7 @@ To use the published alpha, specify its prerelease version:
 
 ```toml
 [dependencies]
-vrp-gpu = "0.1.1-alpha"
+vrp-gpu = "0.2.0-alpha.1"
 ```
 
 To try changes that have not been published, pin a tested repository commit:
@@ -44,22 +44,25 @@ To try changes that have not been published, pin a tested repository commit:
 vrp-gpu = { git = "https://github.com/pedrozaz/vrp-gpu", rev = "<tested-commit>" }
 ```
 
-Enable CUDA candidate evaluation with the opt-in `gpu` feature:
+Enable CUDA candidate evaluation and GPU-selected search with the opt-in
+`gpu` feature:
 
 ```toml
 [dependencies]
-vrp-gpu = { version = "0.1.1-alpha", features = ["gpu"] }
+vrp-gpu = { version = "0.2.0-alpha.1", features = ["gpu"] }
 ```
 
 The default feature set is CPU-only. See the
 [crate-specific README](crates/vrp-gpu/README.md) for a minimal example and the
 public feature contract. The current internal CLI prints a placeholder message;
 use the library API for actual routing work.
 
-The CUDA API currently creates a context and transfers data for each nontrivial
-call. In the [recorded CPU/GPU validation](docs/benchmarks/2026-09-22-cpu-gpu.md),
-its end-to-end latency was higher than the CPU reference for every measured
-route with at least two customers. No GPU speedup is claimed for this alpha.
+Each nontrivial CUDA selection currently creates a context and transfers data;
+iterative search repeats that overhead. The
+[recorded CPU/GPU validation](docs/benchmarks/2026-09-22-cpu-gpu.md) measured
+the earlier single-move API, which was slower than the CPU reference for every
+measured route with at least two customers. Search-level performance has not
+been measured. No GPU speedup is claimed for this alpha.
 
 ## Compatibility
 

crates/vrp-gpu/Cargo.toml

@@ -6,7 +6,7 @@ rust-version.workspace = true
 license.workspace = true
 authors.workspace = true
 repository.workspace = true
-description = "Experimental CVRP library with CPU 2-opt and optional CUDA move evaluation"
+description = "Experimental CVRP library with CPU 2-opt and optional CUDA-selected local search"
 documentation = "https://docs.rs/vrp-gpu"
 readme = "README.md"
 keywords = ["vrp", "cuda", "gpu", "routing", "optimization"]

crates/vrp-gpu/README.md

@@ -2,15 +2,14 @@
 
 `vrp-gpu` provides CVRP data structures, CPU reference heuristics and optional
 CUDA evaluation of 2-opt moves. It parses Solomon-style text, constructs routes
-with a greedy nearest-neighbor heuristic, and improves routes on the CPU.
-The published GPU API computes candidate deltas or selects a single move. The
-development branch also provides host-orchestrated GPU-selected 2-opt search
-for routes and solutions; this new API is not in the published `0.1.1-alpha`.
+with a greedy nearest-neighbor heuristic, and improves routes on the CPU. The
+GPU API computes candidate deltas, selects individual moves, or runs
+host-orchestrated GPU-selected 2-opt search for routes and solutions.
 
 The current alpha is
-[`0.1.1-alpha`](https://crates.io/crates/vrp-gpu/0.1.1-alpha).
+[`0.2.0-alpha.1`](https://crates.io/crates/vrp-gpu/0.2.0-alpha.1).
 The API may change before a stable release. See the
-[documentation for that version](https://docs.rs/vrp-gpu/0.1.1-alpha/vrp_gpu/).
+[documentation for that version](https://docs.rs/vrp-gpu/0.2.0-alpha.1/vrp_gpu/).
 
 ## Features
 
@@ -28,7 +27,7 @@ To use the published alpha, specify its prerelease version:
 
 ```toml
 [dependencies]
-vrp-gpu = "0.1.1-alpha"
+vrp-gpu = "0.2.0-alpha.1"
 ```
 
 For changes not yet published, pin a tested Git revision:
@@ -81,7 +80,7 @@ To use the optional GPU API in the published alpha, enable the feature:
 
 ```toml
 [dependencies]
-vrp-gpu = { version = "0.1.1-alpha", features = ["gpu"] }
+vrp-gpu = { version = "0.2.0-alpha.1", features = ["gpu"] }
 ```
 
 `gpu::evaluate_two_opt_deltas` returns a row-major `n × n` matrix with computed
@@ -94,9 +93,9 @@ invalid input, size overflow or CUDA driver failure. Empty and singleton routes
 return without creating a CUDA context. A returned move is only a proposal:
 apply it with `cpu::apply_two_opt` if your application accepts it.
 
-## Development-only GPU search API
+## GPU search API
 
-On the development branch, `gpu::two_opt_route` repeatedly selects and applies
+`gpu::two_opt_route` repeatedly selects and applies
 best-improvement moves to one route. `gpu::two_opt` does the same independently
 for each route in a solution. Both return `GpuSearchReport` with accepted move
 count and distance improvement computed by `f64` accumulation of the `f32`
@@ -105,7 +104,7 @@ validation, CUDA, or selected-move consistency error leaves the input
 unchanged. A solution search does not itself establish fleet, capacity, or
 customer-coverage feasibility; check `Solution::is_feasible` when needed.
 
-The following example uses the development-branch search API. It compiles
+The following example uses the search API. It compiles
 without running in rustdoc because execution requires a compatible NVIDIA GPU
 and driver:
 
@@ -179,8 +178,9 @@ each iteration; there is no reusable device session. The checked-in artifact
 targets `sm_120` and has been
 hardware-validated on an NVIDIA GeForce RTX 5060 Ti. Other devices have no
 compatibility claim. The [recorded validation](https://github.com/pedrozaz/vrp-gpu/blob/develop/docs/benchmarks/2026-09-22-cpu-gpu.md)
-found this end-to-end GPU API slower than the CPU reference for all measured
-nontrivial routes; no speedup is claimed. See the [input and algorithm guide](https://github.com/pedrozaz/vrp-gpu/blob/develop/docs/user-guide.md)
+found the single-move GPU API slower than the CPU reference for all measured
+nontrivial routes. Search-level latency has not been measured; no speedup is
+claimed. See the [input and algorithm guide](https://github.com/pedrozaz/vrp-gpu/blob/develop/docs/user-guide.md)
 for exact model and numerical contracts.
 
 ## Minimum supported Rust version