Document source: 5a7a75c321ea; code reference: 5a7a75c321ea. zh_CN. Source language: en; requested: en; actual: en; translation: original.

Source documentation does not establish software or physical hardware qualification. See the support and qualification records for their exact scope.

Build system maintenance contract

Nexus uses one authored TOML assembly, a validated immutable configuration IR, static instances and typed factory, and one CMake software target graph. Ninja executes the declared graph. Products own application sources and composition.

The 48-item execution ledger preserves the requested exit conditions. Implementation, executed software verification, environment distribution and physical acceptance have distinct evidence. Historical source qualification does not qualify changed build tools.

Configuration publication is a single-writer transaction. It protects every authored and fact input before changing owned output. A rejected or interrupted attempt cannot expose the previous configuration as current. Successful identical generation preserves file content and timestamps so ordinary reconfiguration does not force unrelated compilation. A configuration digest is not a compiler or generator-source identity.

CMake owns source lists and target dependencies. The resolver emits selected facts and its configuration-code dependencies, never a parallel software build graph. ABI, optimization, diagnostics and firmware link policy have explicit target boundaries. The firmware target guarantees its declared delivery outputs and resource check. External application warnings are a consumer policy.

Developer presets and native CMake/CTest remain the normal build interface. Focused tests are development feedback. Commit gates and formal qualifications execute their complete declared suites with fresh reports. All runners preserve actual command exits, bounded concurrency, deadlines and process cleanup.

Preparation may retrieve dependencies. Formal reproducibility uses two fresh network-disabled builds with the same verified source SDK and recorded environment and toolchain bytes. Compiler caches improve development only; they cannot prove fresh-build reproducibility. Durable evidence stores retain original records, member identities and failed attempts without changing their historical status.

Behavioral fixes use actual RED then GREEN execution. Python tools use unittest; C/C++ contracts use the pinned GoogleTest/GoogleMock host framework. CMake graph and policy defects require real small consumer builds and artifact observations. Firmware stays C11 and contains no host test runtime. Normal commit hooks and independent review apply to this maintenance work.

The following support matrix records observed host tools and declared locks. The final clean-source software gates and eight exact offline candidates are recorded in the execution record. Host observations retain their own scope; source qualification uses the recorded code commit and never follows a later documentation HEAD automatically.

Host or execution path	Maintained boundary	Current evidence
Linux x86_64 development host	Configuration, Ninja, Native/ARM consumers, offline tools and evidence transport	Actual local execution; Linux 6.18.44 and glibc 2.41 observed
Linux descendant cancellation	Bounded /proc inspection, group/descendant freeze, kill and terminal-state observation	Actual nested-process and hostile signal-handler regressions
Linux/amd64 OCI artifact	Sealed manifest, config and layer bytes; local offline load	Actual export, independent archive restore and Docker load/inspect
Windows development paths	Bootstrap selects .venv/Scripts; CMake and tool paths remain explicit	Integration not executed in this maintenance run
macOS development paths	Python/CMake tooling entry points; host-specific compiler setup	Integration not executed in this maintenance run
Other hosts or cross-host ARM distributions	Require a separately reviewed tool lock and execution matrix	Not established by Linux archive identities

Linux descendant cleanup depends on Linux process observation. Generic process termination paths on other hosts do not establish the same nested-process acceptance. Windows and macOS path handling is an implementation boundary, not an integrated test result.

Tool	Declared requirement or lock	Actual local observation
Python	Minimum 3.11; local lock 3.12.14	CPython 3.12.14
CMake	Minimum 3.31; tool lock 3.31.6	3.31.6
Ninja	Tool lock 1.13.2; single configuration	1.13.2.git.kitware.jobserver-pipe-1
Native compiler	GNU 14.2.0 in the environment lock	GCC/G++ 14.2.0, Debian 14.2.0-19
ARM compiler	GNU Arm 14.3.rel1, Linux x86_64 archive lock	GCC 14.3.1, build dated 20250623
clang-format	14.0.6	14.0.6
pre-commit	4.3.0	4.3.0
Optional ccache	4.14.1, Linux x86_64 static distribution	Actual binary and cache hit/miss validation
Git	Source and submodule identity tooling	Host 2.52.0; recorded OCI artifact 2.47.3
GoogleTest/GoogleMock	1.16.0, pinned commit and archive SHA-256	Host-only framework; final full-source execution separately recorded

Ninja single-configuration builds are maintained. Ninja Multi-Config is explicitly rejected; other generators have no implied acceptance. C11 remains the firmware language. Native test targets enable C++17 without linking that test runtime into firmware.

The Linux x86_64 CP312 offline tool closure contains 15 hash-locked wheels, including the Setuptools build backend required by isolated local pre-commit preparation. A fresh CP312 virtual environment completed an actual --no-index --require-hashes installation and real isolated hook initialization. The complete isolated tool domains retain their actual recorded source scopes; the final ordinary commit gate separately executes the current source. CP311 wheel bytes were checked against their 15-wheel lock; CP311 installation was not executed. The minimum Python version and wheel inventory do not substitute for that missing installation result.

The recorded OCI artifact uses Python 3.12.15 and Git 2.47.3, distinct from the current host versions. Its manifest is e936b188b9e30cf37ce109a5407019c91fe2dbc62fb10afceb3f2abbfc823ac8 and its image config is a1ba03c16d424aef88dad6e5b2ead2892734a0bf909dedc23d8855fa4861fbe0. The actual ARM toolchain tree contains 7,190 files and 1,106,755,364 bytes with tree SHA-256 44d2fe14e401730074302ad87bdef9762a6cfe12de4bebc6853d181fd51c74fd. Toolchain and OCI transport executions retain their own tool-source identities; their results do not transfer source qualification to later tool changes.

Fixed base/frontend digests, Debian snapshot selection, five APT package versions and hash-required Python inputs are declared in the recipe lock. Reconstruction of the observed OCI bytes has not executed successfully: recipe_reproducible remains false. Verified artifact reuse and declared recipe inputs therefore remain separate acceptance states.

The execution ledger keeps all 48 requested requirements and exit conditions unchanged. Its updated paths bind the implementation, original RED/GREEN records and the final-source software gates in the execution record and result index. Historical measurements retain their source and exact selections instead of inheriting a later HEAD.

Automatic developer provenance binds a unique immutable command-attempt receipt and checks its command and working directory. Complete current contracts and actual candidate plan execution establish this association; a shared latest report name cannot substitute for an exact attempt.

BLD-18 retains full SDK byte checks at each consumption boundary after measurement, with no parent marker or persistent verification bypass. BLD-31 retains common static libraries after measuring declared versus actually compiled units and exact selected-image costs. Their decision_verified states record measured choices rather than speculative source splitting.

BLD-38 records actual fixed-checker candidate plans and genuine failed-checker execution. The original orchestration exit condition remains unchanged. The maintained 72-stage software plan did not run through one complete plan invocation; individually executed full gates are recorded separately. BLD-39 validates dependency-upgrade evidence contracts, without claiming an actual vendor dependency upgrade.

BLD-40 covers actual CP312 offline installation and isolated hook preparation; CP311 installation remains unexecuted. BLD-41 covers actual whole-toolchain and OCI transport and verified reuse of the same bytes. BLD-42 remains declared_recipe_verified_reconstruction_unverified, with recipe_reproducible=false. BLD-44 covers actual new-source archived proof and independent archive-only verification. Local archive retention and a durable hard-link name do not establish off-site storage, publication or a release.

The host observations retain their recorded source scopes in /workspace/nexus-build-maintenance-20261012/documentation/host-tool-support-observations.json. Current locked closure and final-source receipts are indexed in the delivery result index. The ledger’s configuration, target, workflow, delivery and transport reports retain their raw executions, failure attempts and source scopes. Physical HIL remains not_executed throughout this build-system maintenance work.