v1.0.30

Fixes the Dismount-WindowsImage -Save lock that broke ISO creation — at its real root cause. v1.0.29 (now retracted) misdiagnosed the WIM-commit failure as transient host interference and bolted on a retry; the lock is actually an in-process .NET registry-provider handle held by the build's own PowerShell process, which a retry can never clear. Offline-hive value enumeration now uses reg.exe exclusively (never the HKLM:\z* provider), so no hive handle is held at dismount time — the reg.exe-only pattern of upstream tiny11builder and Microsoft's offline-servicing docs. Pester 521/0 (plus the admin-gated Synthetic harness).

Fixed

• The Standard build no longer fails at Dismount-WindowsImage -Save with "The process cannot access the file because it is being used by another process." Root cause: Invoke-RegistryPatternZeroAction read the loaded offline NTUSER hive through the .NET PowerShell registry provider (Test-Path / Get-Item on HKLM:\z…). The provider caches an in-process RegistryKey handle to the hive's backing file inside the mount; the kernel keeps a hive loaded while any key handle is open, and reg unload has no force flag — so the handle survived into the dismount and locked install.wim. v1.0.29's retry could not help: the handle is held by the same long-lived process for the whole retry window, and GC (the only thing that would release it) is non-deterministic and never fires during the backoff. The latent bug has existed since v1.0.3 — the provider read predates the v1.0.28 idiom tweak and was byte-identical in the v1.0.27 "revert to v1.0.6" — surfacing intermittently with GC timing and deterministically once a build actually reached the commit step. This corrects the v1.0.26–v1.0.29 "a transient WIM dismount/save file-lock is the leading suspect" framing: it was never environmental.
• Stale-hive recovery (Clear-Tiny11StaleHives) no longer opens a provider handle either. It detected stranded HKLM\z* hives with Test-Path HKLM:\z…, which — on the exact stranded-hive path it exists to recover — opened a fresh in-process handle. It now probes via reg.exe (Test-Tiny11HiveLoaded).

Changed

• All offline-hive registry access in the build process is now reg.exe-only. New Tiny11.Hives helpers: Get-Tiny11RegValueNames (value-name enumeration via reg query, replacing Get-Item … | Select-Object -ExpandProperty Property), Test-Tiny11HiveLoaded (load check via reg query), and Invoke-Tiny11RegExe (a %SystemRoot%\System32\reg.exe absolute-path wrapper — child process, handle closed on exit). Invoke-RegistryPatternZeroAction is rewritten onto Get-Tiny11RegValueNames. No change to a successful build's output image — the same value names are matched and the same reg add … /d 0 writes are issued.
• v1.0.29's WIM-integrity gate and -Save retry are retained as defense-in-depth (harmless, and they still guard a genuinely bad save or rare external interference), but they are no longer the mechanism — the reg.exe fix removes the in-process cause at the source.
• Offline-servicing binaries now resolve from %SystemRoot%\System32 by absolute path, not via %PATH% (Local-Dependencies-Only). reg.exe (Hives Invoke-RegCommand + Core's inline tweak writes), dism.exe (Worker cleanup/export, Core appx enumeration/removal, Actions.ProvisionedAppx, and the Start-CoreProcess wrapper that backs dism/takeown/icacls), and robocopy.exe (Worker ISO copy) are all OS-intrinsic servicing tools that can't be vendored, so they're pinned to the System32 path — blocking a PATH-hijacked stand-in. New Tiny11.Hives resolvers Get-Tiny11RegExePath / Get-Tiny11DismExePath. (The post-boot cleanup script's reg.exe runs on the target machine — a separate context — and is unchanged.)

Added

• Regression guard tests/Tiny11.OfflineHive.NoProvider.Drift.Tests.ps1 — a static scan that fails if any src module reintroduces an HK*:\z offline-hive provider-drive reference. Plus reg.exe-only unit coverage for Get-Tiny11RegValueNames, Test-Tiny11HiveLoaded, and the rewritten Invoke-RegistryPatternZeroAction. (The v1.0.29 synthetic-WIM harness never loaded a registry hive, so it structurally could not have caught this class.)

Notes

• Process fix: v1.0.29 shipped without a real end-to-end ISO build — only the content-agnostic synthetic-WIM test ran in CI — which is exactly why a real-build regression slipped through. v1.0.30 is gated on an actual install.wim-loading build (load hives → pattern-zero → unload → dismount-save) before release.
• v1.0.29 was retracted: its GitHub release and Velopack feed assets were deleted, so v1.0.28 remains the offered latest until v1.0.30 ships. The signed v1.0.29 tag is retained — the repo's tag-protection ruleset blocks tag deletion.