Implementation, installed development acceptance and release qualification are separate. This page records the current development evidence; the v1 release profile defines the reviewed binary release boundary. A source update does not update an installed SDK.
| Host | Verified development baseline | Remaining boundary |
|---|---|---|
| Linux x86_64 | Clean 9fa354d SDK: all 590 source CTests pass without skips; 1,184 installed consumer routes, twelve PostgreSQL routes, 146 debugger checks, execution checkpoints, twenty GTK desktop/input cases and all 88 editor groups pass. |
Qualified in a private consumer environment. The reviewed Ubuntu release toolchain and shared deployment permissions require their own gates. |
| Windows x86_64 | Clean 9fa354d SDK: 588 of 590 source CTests pass, with two headless display skips subsequently exercised successfully in Nathan's desktop. Relocated SDK passes all 1,184 consumer routes, twelve PostgreSQL routes, 146 debugger checks, execution checkpoints, twenty Win32 desktop/input cases, four desktop DAP sessions and all 88 editor groups. |
The candidate needed a recorded, scoped read/execute grant on nillib. The permanent packaging correction separately passes real cross-account access tests; older SDKs do not acquire it automatically. Bare-host installation and release qualification remain open. |
| macOS 15 Intel x86_64 | Relocated 1efbe66 SDK passes all 1,184 consumer routes, twelve PostgreSQL routes, 146 debugger checks, execution checkpoints, twenty installed AppKit desktop/input cases, four desktop DAP sessions and all 88 editor groups. Current 59ffa97 source fixtures pass all 22 application/tooling/packaging CTests. |
The installed editor needed a diagnostic rerun after a project-build timeout; it passed unchanged deadlines, with the original cause unconfirmed. The original full source report retains four failures resolved by later fixtures and one ELF-only execution skip. Shared deployment and release qualification remain open. |
| macOS 15 Apple Silicon arm64 | Relocated 1efbe66 SDK passes all 1,184 consumer routes, twelve PostgreSQL routes, 146 debugger checks, execution checkpoints, twenty installed AppKit desktop/input cases, four desktop DAP sessions and all 88 editor groups. Current 59ffa97 source fixtures also pass all 22 application/tooling/packaging CTests. |
The original full source report retains four failures and one ELF-only execution skip; subsequent hosted checks resolve the four fixture failures. Shared deployment and release qualification remain open. |
The SDKs contain eleven root commands, eighteen compiled standard-library
modules, their source, native runtimes and HTML documentation. The installed
consumer campaign covers both nilstd and nillib: 888 retained execution
results are independently replayed from raw reports, while 296 object/link
routes are checked by the pinned campaign driver. Each private candidate retains
267 payload files plus its manifest through relocation to a path containing
spaces and Unicode. Mac executables are checked as Mach-O for their expected
architecture. These candidates have not replaced accepted SDKs or activated a
release.
The Linux editor result uses VS Code 1.144.1 in an isolated X11 session and empty profile. Both current Mac application campaigns use VS Code 1.146.0. Each passes 88 distinct groups, all nine test modules, the full-suite completion sentinel and no exclusions. Editor results using build-tree tools are distinguished from results using an installed SDK.
Installed desktop/editor coverage is complete on both Mac architectures: ARM run 43146638765 and Intel run 43149252918. The first Intel editor attempt timed out after 25 groups. With task lifecycle diagnostics added, all 88 groups pass using unchanged SDK bytes and deadlines; the Ny VM project group takes 161.171 seconds of its 180-second allowance. All 31 observed task lifecycles complete. This establishes successful installed execution, not a root-cause fix for the earlier timeout or a performance promise.
Exact revisions, evidence hashes, host preparation, interrupted attempts and failure follow-ups are retained in the current parity notebook. Earlier baselines remain in the Windows/Linux closeout and macOS bring-up notebook.
The original full Mac LLVM-ON suites at 1efbe66 each record 583 passes, four
failures and one skip from 588 registrations. ARM failures are the test-only
AppKit input injection, two manual native-link fixtures missing framework
arguments, and a first project build exceeding its fixture deadline. Intel's
AppKit checks pass; its fourth failure is instead the first tool-discovery
project run exceeding a 60-second child deadline.
Subsequent test fixtures isolate AppKit injection from host compilation, pass
the runtime's required framework arguments, and give real Darwin project runs
bounded allowances that reflect toolchain fingerprinting work. Explicit short
process/error deadlines remain unchanged. The complete 21-contract campaign at
eca43f1 passes on both architectures, including project build/run and tool
discovery, with four desktop DAP sessions and all 88 editor groups
(run 43141540789).
These results supplement the unchanged SDK; the original failed reports remain
failed. The ELF-only freestanding executable fixture remains an explicit
format-specific skip. Actual Mach-O archives separately pass import validation.
The next campaign at 59ffa97 adds mixed-source/native SDK packaging, including
the destination-access regression: all 22 selected CTests, four desktop DAP
sessions and all 88 editor groups pass per Mac architecture
(run 43148956133).
The exact pinned thirteen-test packaging invocation passes; its original JUnit
output retains only eight complete result lines because CTest truncates passing
output at 1,024 bytes. Future runs retain a bounded 64 KiB per SDK/project test.
Windows's former transient native output-replacement failure is corrected at
9fa354d: compiler and linker share bounded Windows sharing/lock retries and
preserve existing output on persistent failure. Strict MSVC controls exercise
real held handles, read-only output, Unicode paths and exact byte preservation;
the former single-attempt implementation fails the transient-release control.
The complete source suite then has only the two headless display skips above.
LLVM-OFF builds exercise frontend/interpreter behavior and rejection of disabled native operations. Their passing tests do not establish native execution.
Both Mac architectures pass native incremental reuse and invalidation. Compiler identity hashes actual loaded files and exact shared-cache backing files; native linker probes collect 8,321 inputs. Unmodeled dependency profiles rebuild rather than reuse stale output. SQL runtime loader paths are scoped to execution processes so they do not change ordinary compiler/linker identity.
Bounded x86_64 and aarch64 native inline assembly are described in the assembly contract. Thin Mach-O64 x86_64 and arm64 objects/executables support bounded context inspection, opt-in annotation embedding and dead-strip retention. See the container contract for supported forms.
ELF, COFF and Mach-O freestanding runtime archives are checked for malformed or mixed object profiles and unexpected imports. Actual MSVC no-CRT controls link and execute with only the explicitly selected host exit import. This does not claim a cross-platform bare-metal application profile.
Managed and owned native lifetime tests pass on both Macs. The owned workloads allocate and free 36,000 guarded buffers per memory mode; withholding frees reaches the 96 MiB retained-payload bound. Darwin measures intercepted malloc-family payloads, excluding allocator metadata and mappings. Linux's address-space limit and Windows Job limits retain their platform-specific metrics; these are not identical whole-process memory measurements.
The UI provider is GTK4 on Linux, Win32 on Windows and AppKit on macOS. Linux needs an external GTK4 runtime. Display/input tests need an interactive window server; headless skips do not count as passes. Source AppKit campaigns pass provider, lifecycle, layout, focus, sensitivity, Unicode entry and real interpreter/native desktop/input contracts on both architectures. Test input uses each owning application's responders, without claiming general desktop automation or cross-process accessibility support.
nilinterpreter --ai provides opt-in local control through Windows named pipes
or Unix sockets. Pause, resume, instruction stepping and bounded inspection use
the real VM execution thread. --ai-wait starts paused before main.
vm.snapshot and vm.restore provide reusable same-process execution
checkpoints for paused deterministic VM programs: execution position, frames,
values, heap identities and admitted runtime state. Open resources, external
effects, native/JIT execution, cross-process persistence and code hot reload
remain outside this profile. See NIL-0006.
DAP keeps VM execution, inspection and UI cleanup on the process main thread.
Passing interpreter debugging does not establish native debugging.
Sockets, secure randomness and text-byte APIs pass the same six interpreter contracts on Linux, Windows and both Macs without skips. These additions require the interpreter on every host because the native text ABI does not yet preserve arbitrary embedded NUL bytes. File timestamps use native filesystem metadata across interpreter and native routes; text/binary NIL, JIT and AOT checks pass on all four hosts. PostgreSQL acceptance uses an isolated owned server and an external client runtime; it does not bundle or qualify every provider version.
The SDK builder honors CMAKE_BUILD_PARALLEL_LEVEL. SDK publication checks
cover staged layout, relocation, retained payload bytes and modes, and bounded
replacement. Cross-account Windows use additionally exposed a binary-library
access defect: the private temporary compilation directory's DACL survived its
move into the SDK. Packaging now copies validated, completed outputs into fresh
destination entries, inheriting normal SDK access without exposing private
compilation or changing machine ACLs. Linux tests cover restrictive umasks,
unchanged bytes and failed-copy rollback; a separate limited Windows account
reads all 102 published library files while remaining unable to read private
staging. These checks do not retroactively change older SDK permissions or
qualify arbitrary shared installations.
The release gates still apply before publishing new binaries. Current development coverage does not expand the reviewed Ubuntu release profile, establish a bare-host Windows or macOS installation, or imply a stable v1 release.