pwiz: Fixed four vendor readers failing from an installer-based install

[chambm]

Summary

Three related pieces, in the order they were found. #4641 has been merged into this branch, so
this is the single PR on top of #4619.

1. Four vendor readers failed from an installer-based install

All pre-existing. None reproduce in a dev or CI build: those keep the vendor DLLs app-local,
and VendorSdkLoader's resolvers prefer app-local over the cache. All four were reproduced on
native Windows with the stock installer before being fixed.

• Three wrong pins in vendor-sdk-pins.json.
  • The Bruker archive claimed Agilent's MHDAC prefixes (BaseCommon, BaseDataAccess,
    BaseError, BaseTof) although it ships exactly two DLLs, baf2sql_c and timsdata.
    FindPin returns the first prefix match and Bruker precedes Agilent, so a BaseCommon
    bind resolved to the Bruker archive: "extracted cleanly but contains no BaseCommon.dll".
  • MIDAC. never matched: prefixes are StartsWith against the assembly simple name, and
    the assembly is plain MIDAC. Agilent then failed one assembly later on "Could not load
    MIDAC".
  • MobilionShim is pwiz-sharp's own shim, not something the Mobilion archive carries, so
    listing it made the staging filter strip a binary nothing could supply.
• MobilionShimNative loads MBI_SDK.dll by full path in a static constructor.
  MobilionShim links MBI_SDK.lib, so it is a static import resolved by the OS loader —
  a path no DllImportResolver sees. The error named the wrapper, not the missing dependency.
• Wiff2LoadContext resolves the wiff2 SDK from the cache when it is not app-local.
  LoadFromAssemblyPath consults no resolver. Same shape as Bruker's
  CompassXtractActivationContext, which calls EnsureExtracted explicitly for the same reason.
• Shimadzu.csproj deploys VC140 app-local, as Agilent.csproj does VC120 and
  Bruker.csproj VC90. A missing MFC140 makes the reader return a file with no spectra
  rather than an error, so any machine carrying the redistributable hides the dependency.

2. A vendor-bundled installer variant (was #4641)

• build.ps1 -WithVendorSdks produces ProteoWizard-WithVendorSdks-Setup-<ver>.exe
  (103.8 MB): the bundled-runtime installer plus every Windows vendor SDK, pre-extracted into
  the layout VendorSdkLoader.EnsureExtracted would have produced, .ok markers included, so
  the loader neither downloads nor extracts.
• Setup.iss installs that cache to %PROGRAMDATA%\ProteoWizard\vendor and stamps
  vendor-cache-root.txt, which is what redirects the loader off its per-user default.
• --with-vendor-sdks through build.bat and tcbuild.bat — opt-in, so a developer build
  does not pay the ~30 s / ~26 MB.
• Installer.Tests extended from 4 fixtures across 3 vendors to 9 across 8, and made to
  accumulate failures rather than abort on the first. This is what found everything in part 1.
• Renamed ProteoWizard-Sharp to ProteoWizard across the installer.

3. A failed installer build was reported as a successful CI build

Found on this PR's own CI run (build 4164785, agent EC2AMAZ-HH9KVFB):

12:02:18  Remove-Item: build.ps1:244  The directory is not empty ... Waters-.../vendor_api
12:02:19  installer build failed; Installer.Tests will skip          <-- WARNING only
12:09:09  Test ignored: Install_PerUser_...   No ProteoWizard-Setup-*.exe found
12:09:09  Test Run Successful.  Total tests: 2   Skipped: 2      -> build SUCCESS

So the installer never got built, both Installer.Tests went Inconclusive, and the build went
green — with the coverage that gates part 1 never actually running.

• build.bat: a failed installer/build.ps1 is now fatal. That branch is only reached
  when ISCC is present (a machine without it takes the else), so a failure there is real.
• build.ps1:244: the vendor_api cleanup retries, and distinguishes what matters. By
  that point everything wanted has been moved out and the directory holds only the x86/mips
  leftovers — for Waters, two empty directories, deepest path 108 chars, no reparse points,
  nothing read-only. Aborting the installer build over that was the wrong trade. It now retries
  5x with backoff; leftover files still throw (the flatten genuinely failed), leftover
  empty directories warn and continue.

The delete failure is agent-specific, not a defect in the flatten: two builds of the same
commit with the same step passed on MacCoss TeamCity Agent 1 and failed only on the EC2
agent, whose volume behaves unlike local NTFS (the same class of difference that makes
std::filesystem::canonical throw there).

Test plan

• pwiz-sharp/build.bat Release --i-agree-to-the-vendor-licenses --with-vendor-sdks — build succeeded, all three installers produced (78.2 / 21.5 / 103.8 MB), every suite green: Agilent 17, Waters 21, Bruker 15 +1, Thermo 15, Sciex 8, Mobilion 2, Shimadzu 2, UIMF 2, UNIFI 44, BiblioSpec 142, Analysis 177, MsData 85, Common 58, Util 36, plus the rest
• installer/build.ps1 -SkipBuild -WithVendorSdks twice in a row — the reported failure was a second-run report
• Cleanup retry branches exercised with blocker processes: transient block recovers (668 ms); a held-open file throws naming the file; permanently-stuck empty directories warn without throwing
• build.bat control flow verified separately — goto out of two nested parenthesized blocks under delayed expansion skips the guard line and propagates the exit code; CRLF endings intact
• Installer.Tests per-user fixture — 9 fixtures across 8 vendors from the real installer artifact
• 43-pair vendor sweep in the wine 11 container with --network none: 43/43 converted, 40/43 count-match (the 3 diffs are ReaderTestConfig::indexRange-truncated references, not regressions)
• ABI/Sciex offline check in the same image: 4/4 converted, 11 count-matches
• Install_PerMachine_DeploysAndConvertsVendorFile is skipped locally — needs elevation

Notes for reviewers

Two things found while investigating, deliberately not fixed here:

• A pre-existing test race. Lockmass_..._Cwt_AppliedAndUnapplied failed once with
  MassLynx combineScan failed (code 5): File Open Error, then passed 3/3 in isolation and in
  a clean full re-run. ATEHLSTLSEK_profile.raw is shared by Analysis.Tests and
  Waters.Tests, Run-Tests-Parallel.ps1 runs suites as concurrent jobs, and the Waters SDK
  writes lmgt.inf inside the .raw. Fixing it means serializing those two suites or giving
  each a private copy of the fixture.
• Shallow-clone pin versions. CI logs Shallow clone enforced, and every vendor pin
  collapses to the build SHA (b10f12fe7fe4 for all of them) where a full clone gives each
  archive its own. Not a correctness bug — the generated .cs and the cache agree within a
  build — but the cache key churns every commit on CI instead of tracking archive content.

Co-Authored-By: Claude noreply@anthropic.com

[Copilot]

🟡 Changes recommended

Installer.Tests’ new “accumulate failures” logic still aborts on non-assert exceptions, and should be adjusted so it consistently reports all vendor failures as intended.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

Fixes multiple vendor-reader failures that only occurred in installer-based installs (where vendor SDK DLLs are resolved from the VendorSdkLoader cache rather than app-local), and improves CI/packaging to reliably build and validate installer artifacts (including an offline-friendly vendor-bundled installer variant).

Changes:

• Correct vendor SDK pin prefixes and adjust vendor SDK extraction/resolution so installed builds reliably load Sciex wiff2 and Mobilion dependencies from the cache.
• Add support for an optional “WithVendorSdks” installer variant that pre-populates the shared vendor cache and stamps a cache-root override for offline/container installs.
• Make installer build failures fatal in CI and improve installer payload staging/cleanup behavior.

File summaries

File	Description
pwiz-sharp/tcbuild.bat	Forces CI builds to include --with-vendor-sdks when invoking build.bat.
pwiz-sharp/pwiz/test/Installer.Tests/InstallerTests.cs	Expands installer smoke coverage to more vendors and reports multiple vendor failures in one run.
pwiz-sharp/pwiz/src/Vendor/Shimadzu/Shimadzu.csproj	Deploys VC140 runtime DLLs app-local for Shimadzu native dependencies.
pwiz-sharp/pwiz/src/Vendor/Sciex/Wiff2LoadContext.cs	Resolves Sciex SDK directory from app-local or VendorSdkLoader cache for installed builds.
pwiz-sharp/pwiz/src/Vendor/Sciex/Sciex.csproj	Adds a direct reference to Vendor.Common to support explicit cache extraction for wiff2 SDK loading.
pwiz-sharp/pwiz/src/Vendor/Mobilion/MobilionShimNative.cs	Ensures MBI_SDK.dll is loaded by full path from the vendor cache before P/Invoke loads MobilionShim.dll.
pwiz-sharp/pwiz/src/Vendor/Mobilion/Mobilion.csproj	Adds a direct reference to Vendor.Common to support explicit cache extraction for Mobilion.
pwiz-sharp/pwiz/src/Vendor/Bruker/Bruker.csproj	Updates comment describing Vendor.Common direct usage (now shared with Sciex).
pwiz-sharp/installer/Setup.iss	Renames product to ProteoWizard and adds an optional vendor-cache install + cache-root stamping for offline installs.
pwiz-sharp/installer/build.ps1	Adds -WithVendorSdks pipeline, improves staging filters (non-Windows payload trimming), and stages msconvert/wiff2 outputs explicitly.
pwiz-sharp/installer/build-linux.sh	Renames Linux artifact names to match ProteoWizard naming.
pwiz-sharp/build/vendor-sdk-pins.json	Fixes incorrect prefixes (Bruker/Agilent/Mobilion) that broke SDK resolution from the cache.
pwiz-sharp/build.bat	Adds --with-vendor-sdks and makes installer build failures fatal to prevent false-green CI runs.

Review details

• Files reviewed: 7/13 changed files
• Comments generated: 2
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.