# Clean Windows release qualification — 2026-10-10 The clean detached checkout builds the published source `54619ce189c4280e729544b21a95503dc814fa0d` with the native MSVC 14.44 owner factory and the default release profile. No source or release optimization flag changed. The factory invocation reports Bun.version `1.4.4`, the exact source revision, and its built bun-profile.exe path. The CLI banner retains the source's `1.4.4-canary.1` channel; this is a candidate, not a new stable tag. The Windows bun.exe is 152247296 bytes. SHA256: `08f2b2a880d8dc7d4a72b4923d4adb3a0387b46dd266ad180eac4d5a3ec8e391`. The checkout remains clean after build and qualification. ## Native tests Nine release runtime suites pass 102 tests, with 27 platform skips and zero failures. Coverage includes filesystem watchers, terminal subprocesses, timer races and parallel worker startup/reporting. The complete selected gate, including all test/internal files, passes 995 tests, with 60 skips, zero failures, six snapshots and 146629 assertions across 101 files. Release requires `--expose-internals` for internal test APIs, as specified by ModuleLoader.rs. The initial invocation omitted that flag: 960 tests passed and five files failed to load bun:internal-for-testing. The final invocation reruns every selected file with the supported flag; no source, assertion or deadline was changed. ## GPU and Python The same clean release executes the existing Buv GPU qualification against the canonical aphrody_ffi provider. Physical RTX 4070 selects Dx12; WGSL and CUDA independently return all 37 exact expected values. Driver enumeration reports NVIDIA 616.92, CUDA 13.4, compute capability 8.9 and 46 multiprocessors. The provider SHA256 remains `351e930b6b7fe81f4eca8f405c59e0123e9430e320fb51a5d9782ef6ab7bd6ee`. The published PyCUDA test passes one test and six assertions through the shared Buv/PyJS Python host. It uses the existing isolated Python environment, provider and CUDA toolkit. This proves native Windows execution; it does not prove zero-copy buffers, a shared CUDA context, Linux GPU activation or musl ABI. ## Plugin The release's agent-plugin doctor recognizes the installed Codex and Claude plugin `1.4.4+2fb86b44ed79` without a Bun-version mismatch. Plan-before-apply installation with `--update --json` returns `up to date`, both targets and zero actions. Installed content remains the 313-file plugin tree previously verified against the embedded payload. Skills are used in this workflow; actual provider activation or newly exposed MCP tools is not inferred. ## Build resources and remaining gates The first build stopped on a libc build-script rustc probe routed through an inherited sccache wrapper, with socket timeout 10060. The successful retry clears RUSTC_WRAPPER and CARGO_BUILD_RUSTC_WRAPPER only in its process. The native factory already invokes rustc directly; global cache settings remain. Before the build, an exact-file plan reclaimed 9557 old Rust metadata files from the resolved physical canonical debug/deps directory. Every file was older than one day, matched the hashed .rmeta name, had Rust magic, and retained its planned size/time. A canonical Cargo mutex and idle compiler check guarded application. No recursive removal was used. Executables, DLLs, PDBs, rlibs, release outputs, source, protected stores and global sccache were preserved. Measured free space rose from 2753650688 to 15665561600 bytes. Receipts are local under tmp/windows-clean-release-\*. The artifact qualification is Windows x64 only. Full release publication, other architectures, packages, performance equivalence, Linux hardware and production activation remain open. The separate strict Windows runtime gate retains 53 diagnostics outside the last owned batches; GNU x64/ARM64 strict gates pass. This document does not close those plans or change the production runtime. ## Next adapter and host readiness follow-up The clean release scope check exits zero. An exact list of the five files in `test/integration/next-bun` passes 82 tests and 319 assertions, zero failures. The initial directory prefix also matched next-bun-app and next-bun-pages; that pass was explicitly stopped because its dependency installation was not yet confined to a local registry. Its partial passes are not counted as completion of the application integration gate. Actual Next application and Pages Router gates remain open pending a controlled fixture install. Fresh native Infra observations confirm DBFR still selects 1.4.3-aphrody.3; the isolated stripped Linux candidate retains its recorded SHA256. VPS retains the owner Linux checkout at 4e3ddd04da1, with only the already published Linux test projection changes, whose exact file checksum matches origin/main. The catalogued VPS reclaim was planned before application, with no active Ninja or rustc process. Infra reports exit zero. Measured filesystem available space rose from 3450236928 to 6331330560 bytes. Its reported 9.87 GiB logical reclaim includes tmpfs and is not the actual disk gain. Existing Docker containers, production runtime and DBFR services were not activated or changed. The next Linux release build still needs its own source/capacity/factory plan. ## Offline real Next applications BUN_TEST_NEXT_CACHE_DIR selects a preseeded fixture cache. The helper uses `--offline --frozen-lockfile --cache-dir --config=`; the temporary install config prevents the fixtures' cache=false setting from disabling the prepared cache. Missing cached packages fail without fetching. No global Bun configuration, fixture assertion or test deadline changes. Without the variable, the existing fixture behavior is retained. Provisioning runs explicitly outside the tests: copy each fixture's package.json and bun.lock into an isolated preparation directory, then run the owner release factory with `install --cwd --cache-dir --frozen-lockfile --ignore-scripts`. Prepare the App fixture 16.1.6 and Pages overlays 16.4.0 and 16.5.0-canary.4. Set BUN_TEST_NEXT_CACHE_DIR to that cache for the application gate. This is preparation, not an internet-dependent test. The release engine at 54619ce189c, with the owned helper patch, passes all 12 tests and 251 assertions across the exact Pages and App fixture files, zero failures. Pages HTML/API parity passes all three Next versions. App Router passes Turbopack, webpack, next-bun without Node on PATH, and Bun.build through @aphrody/next-bun. Coverage includes server actions, prerendered HTML, RSC payloads, PostCSS execution and client chunk traversal. This replaces the earlier interrupted application-gate evidence. The helper's scoped TypeScript dependency closure, Prettier and oxlint gates all exit zero. The initial offline invocation failed before builds because a separate --config value became a dependency positional; the supported equals form fixes the invocation. No runtime workaround or assertion reduction was introduced. Receipts: tmp/next-owner-offline-plan.json, next-owner-seed-\*.log and windows-clean-next-offline-apps-final.log/.exit. ## Strict historical performance comparison The current Windows release was measured against the retained base executable 620b50f6abea3413a30235c5885bfe8cbffd592d (1.4.3). The candidate is 54619ce189c (1.4.4). Forty measured runs and five warmups use the existing thresholds, hyperfine and --strict. No threshold or runtime feature was removed. All startup, build/install, RSS, import and local HTTP blocking thresholds pass. Startup RSS increases by 816 KiB and builtins RSS by 728 KiB, inside the existing 1 MiB / 2 MiB limits. The strict gate exits one on binary size alone: 152247296 versus 96548352 bytes, ratio 1.577 and delta 55698944 bytes. PE section inspection attributes 44121356 additional bytes to .text and 10573708 to .rdata. This is executable code/data, not a removable debug section. The compared source versions differ: these measurements do not isolate the fork delta at an identical upstream source pin. A current comparable baseline and the size investigation remain required; no complete performance closure or release approval is inferred. Receipts: tmp/perf-owner-latest-plan.json, tmp/perf/owner-54619ce-release/perf-report.json and tmp/perf-owner-pe-sections.log. ## Latest Linux release refresh The VPS owner checkout advances cleanly to published 988d73b430a after checking that its sole Linux test modification matches the target blob byte for byte. Other worktrees and DBFR candidates remain untouched. The existing release uses 4179013632 measured bytes. With manifests, toolchain, build scripts and vendor sources unchanged, an incremental plan budgets an entire replacement release directory plus 2 GB scratch. Available space at the initial observation is 6456922112 bytes; the runner checks the 6179013632-byte minimum before applying. The tmux owner-linux-latest runner acquires the canonical heavy factory lock nonblocking, uses four jobs and nice 10, then invokes the unchanged owner release factory and its private populated cache. Its subsequent gate includes all internal tests with --expose-internals and the six established Linux suites. The runner, Ninja PID, clean source pin and building phase were observed live. This is an active build, not a successful artifact or deployment receipt. Plan: tmp/vps-owner-latest-release-plan.json. Remote phase/PID/exit receipts are in the existing owner-linux-qualification directory; original receipts remain. ## Common-pin internal qualification (2026-10-10) Both native owner factories report Bun 1.4.4 at 988d73b430a3c47cd61d6f8d440777407f77555c. The Linux gate initially reported 1119 pass, 19 skip and three failures. Four scoped corrections complete the fake toolchain host, match the installer's Windows Terminal executable name, accept the provenance suffix in the LTO cache key, and retain APK numeric comparison behavior without implicit fallthrough. The gzip fixture now supplies an owned Uint8Array buffer. No assertion, skip policy or runtime feature was removed. The patched Linux rerun reports 22 pass in the three affected files, then 1122 pass, 19 skip, zero failures and 147818 assertions across 98 files. Windows reports 893 pass, 33 skip, zero failures and 146170 assertions across 92 internal files. Both use the freshly built release through the owner build-then-exec factory with --expose-internals. Scoped TypeScript, Prettier and oxlint exit zero. Receipts: tmp/linux-internal-fixtures-plan.json, tmp/linux-internal-fixtures-{tsc,format,lint}-final._, and tmp/windows-common-pin-internals._; Linux native-fixtures.log and native-complete.log remain in the VPS owner-linux-qualification directory. The refreshed Linux stripped artifact SHA256 is 8bb9105153ce3742d14f9c04247782c5013d050813b00eae880d041ae9ceb8e3; the profile artifact is 12bcb75989497280e3ca2dab3fc51c580dca9965d7bac78317eb6492274365bc. These checksums identify build outputs, not an activated deployment. The retained Windows 54619ce executable and PDB share CodeView GUID 8EAF22F1-5F73-D4BD-4C4C-44205044422E and age 1. Its PDB attributes 25021543 text bytes to 79 uv modules. The old baseline PDB is unavailable; this measures current footprint and does not attribute the size regression. Performance size closure, physical Linux NVIDIA qualification and deployment remain open. ## Refreshed GPU and plugin qualification (2026-10-10) The owner Windows release factory at 988d73b430a repeats the physical RTX 4070 qualification. wgpu selects Dx12 and WGSL returns the 37 expected doubled values, exactly matching the CUDA result. NVRTC and the CUDA driver load; the device reports compute capability 8.9, 46 SMs and driver API 13.4. The independent PyCUDA case passes through the shared Buv/PyJS Python host: one test, six assertions, zero failures. This does not prove shared CUDA contexts or zero-copy buffers. Receipts: tmp/windows-common-pin-gpu._ and tmp/windows-common-pin-pycuda._ (both exit zero). The same factory's agent-plugin doctor exits zero and reports current: true for the installed Codex and Claude payload 1.4.4+2fb86b44ed79, hash 2fb86b44ed790069182d6cb9920cb9955b59b1305e81aa9829415cdbeec9ba08. The installed bun-build skill was read and used for factory routing. This proves installation/content compatibility; callable provider MCP tools and provider session activation remain separate qualification requirements. Receipt: tmp/windows-common-pin-plugin-doctor.\*. The refreshed stripped Windows executable is 152247296 bytes, SHA256 2782577addbdd6c51199f9e7e51ffa1645ab0331f6f035f039444e88a1152fe7. The existing strict size threshold remains unchanged and unresolved.