release: version 0.13.0

CHANGELOG.md

@@ -2,6 +2,21 @@
 
 All notable changes to the SUCO distributed compilation system will be documented in this file.
 
+## [0.13.0] - 2026-08-04
+
+### Added
+- **Remote preprocessing — the client no longer runs `-E` (#42).** Instead of expanding a translation unit locally and shipping 1–3 MB of preprocessed text, the client now runs the compiler's cheap `-MM` dependency scan, packs the project headers into a content-addressed bundle, and sends the **raw** source plus that bundle; the *worker* preprocesses. Measured: **−42 % client CPU** on a header-heavy Linux corpus and **3.19× wall-clock** on a real 101-TU Windows build (20.78 s → 6.52 s). The win scales with header weight — heavy templates and umbrella headers benefit most, trivial TUs barely at all.
+
+  **This is on by default.** Byte-identity was the gate and was verified on both platforms before flipping it: 240/240 objects byte-identical to a native compile across `-O0/-O2/-O3` on Linux; on a real Windows→Linux PE/COFF cross-compile, 13/15 identical with the two exceptions fully explained (`.text` machine code byte-identical, sizes equal, identical to the normal grid object, 100 % deterministic across runs — the delta is path-string normalisation, the same normalisation the grid already applied). A 101-TU Windows project builds, links and runs. Translation units that cannot be preprocessed remotely fall back to the old path automatically: anything using `__DATE__`/`__TIME__`/`__TIMESTAMP__` (in the source *or* in a project header — the worker's clock would differ), C++20 modules, and MSVC. Both paths are pinned in CI, so the classic path stays exercised. Escape hatch: `SUCO_REMOTE_PREPROCESS=0`.
+- **The worker can sandbox the compiler itself (#43),** opt-in via `SUCO_SANDBOX=1`. A compile is untrusted execution, so it now runs with a read-only filesystem, no network, and **only its own job directory writable** — the object and the compiler's own intermediates (`TMPDIR`) live inside that directory, so two concurrent jobs on a worker cannot see or clobber each other's scratch. Prefers `bwrap` when installed, falls back to `unshare(1)`; both were verified to produce byte-identical objects, because namespaces do not change compiler output. Off by default, and a no-op when off.
+- **A VS Code extension** (`extension/vscode/`) showing grid status in the status bar (workers online, cache-hit rate, active jobs), a one-click toggle that injects `CMAKE_CXX_COMPILER_LAUNCHER=suco-cl++` into CMake Tools, and a shortcut to the dashboard. No runtime dependencies. The coordinator host defaults to localhost and otherwise follows `SUCO_COORDINATOR_HOST`, so an already-configured machine needs no setup. A Visual Studio 2022 counterpart is scaffolded (`extension/visualstudio/`) but has **not** yet been loaded in a live VS instance.
+
+### Fixed
+- **A Windows worker could not run any dispatched job at all.** Three POSIX assumptions killed it before it ever compiled: the toolchain was unpacked with `tar -I zstd`, which is GNU tar syntax that the **bsdtar** shipped as `tar.exe` on Windows rejects outright; the received toolchain archive was written to a hardcoded `/tmp/…` that does not exist there; and the header-cache directory fell back to `/tmp/.cache/…` whenever `HOME` was unset — the normal case on Windows. All three are now platform-correct (bsdtar auto-detects zstd, so plain `-xf` is right; `std::filesystem::temp_directory_path()`; `LOCALAPPDATA`/`USERPROFILE`). Found by actually running a Windows worker and reporting that the log never showed a job.
+- **A path-escape hole in the header-bundle unpack guard, specific to Windows.** The guard rejected absolute paths and `..`, which is not enough: `"C:foo.h"` is drive-*relative*, so `is_absolute()` is false, yet `dest / "C:foo.h"` resolves to `"C:foo.h"` — `std::filesystem` discards the destination when the right-hand side carries a different root name — writing outside the job directory. UNC paths and root-directory paths escape the same way. Bundle entries are now validated on the raw string, so a Linux worker and a Windows worker accept or reject the same bundle identically.
+- **Remote-preprocessed cross-compiles used the wrong compiler.** The new dispatch path took the *local* compiler name rather than the resolved target-qualified one, so a MinGW target would have been built with the host `g++`. It now mirrors the existing dispatch path's compiler resolution.
+- **`make_ppa_source.sh` read the wrong version** — it matched the first `VERSION x.y.z` in `CMakeLists.txt`, which is `cmake_minimum_required(VERSION 3.15)`, so PPA source packages were versioned `3.15` instead of the project version.
+
 ## [0.12.0] - 2026-07-26
 
 ### Added

CMakeLists.txt

@@ -1,5 +1,5 @@
 cmake_minimum_required(VERSION 3.15)
-project(suco VERSION 0.12.0 LANGUAGES C CXX)
+project(suco VERSION 0.13.0 LANGUAGES C CXX)
 
 set(CMAKE_CXX_STANDARD 20)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)