.NET 10 migration validation - 19 October 2026¶

Documentation index · Development

This record covers the migration from main 4c9a85b (.NET 8, DS4Windows 5.0.34) in PR #56. It records automated evidence and maintainer acceptance for the migration; it does not establish compatibility with every controller or Windows edition.

Build inputs and scope¶

• Pin SDK 10.0.402, which publishes runtime 10.0.13. Verify the SDK archive's SHA-512 against Microsoft's release metadata before use. Keep exact SDK selection and locked dependency restore.
• Retarget all 14 C# projects, including installer hosts, tests and diagnostic tools. Preserve the application's windows10.0.19041.0 API baseline.
• Keep x64 packages self-contained. Keep the existing workstation/concurrent GC settings, controller protocols, polling policy, VIIPER 0.1.12 and USB/IP 0.9.8.1. No driver update is introduced by this migration.
• Retain direct application/library versions. Refresh all affected dependency locks for the new framework. .NET 10 prunes transitive framework packages; installer single-file tooling follows the SDK's ILLink 10.0.13.
• Remove unused .NET 8 installation checks and obsolete download notices. Record the running runtime version in both saved and copyable startup logs.

Compatibility review¶

• .NET 10 adds Windows Forms ContextMenu and MenuItem types. Profile Editor explicitly selects its existing WPF controls to avoid ambiguous names; menu behaviour is unchanged. See Microsoft's breaking changes.
• Match Marshal.QueryInterface calls to the current in Guid signature; retain the existing native WinRT identity/lifetime test.
• WPF now rejects some invalid dynamic-resource types and empty grid definitions. Build and exercise the real light/dark resources and layout tests.
• The app remains a normal multi-file publish. The installer hosts remain single-file/self-extracting; review the changed native-library search rules and exercise the built bundle's offline layout in addition to MSI lifecycle checks on an isolated CI machine.
• .NET 10 changes XmlSerializer's handling of obsolete properties. The profile DTOs have no such attributes; retain the full profile/settings round-trip suite as the gate.
• Both initial hosted CI runs exposed 24 bytes across 20,000 calls in the completed Switch 2 USB startup transaction allocation test, despite the first local full suite passing. A cold focused run reproduced it. A no-GC measurement scope alone still failed, so this was not resolved as GC accounting noise. JIT disassembly showed a Tier1-OSR transition inside the measured loop, with inlined Monitor paths. Separate diagnostic child processes disabling tiering, OSR or PGO each passed. Warm and measure the same optimized, non-inlined test-driver helper so that transition does not first occur in the measured loop. Retain the 1,000-call warmup, 20,000-call measurement and exact zero-byte assertion. Production methods, runtime/JIT/GC settings and controller behaviour are unchanged. Diagnostic runs are attribution evidence, not release gates; ordinary default-runtime focused and full-suite runs remain required.
• The subsequent full run also reproduced a 24-byte failure in the Bluetooth pacer saturated failure/reset storage test. A cold focused run reproduced it with a Tier1-OSR transition; separate no-OSR/no-PGO diagnostic processes passed. Optimizing only its cycle helper did not fix this case. Compile the saturated measuring test driver before its warmup instead, preventing OSR inside its counters. Keep the original warmup size, measured count, ownership checks and exact zero-byte limit. Its ordinary focused run then passed alongside the startup transaction and measurement-scope positive-control tests.
• A later full run exposed the same 24-byte cold-window observation in the pacer's completion-lease cycle gate. Its isolated default-runtime run passed, so the full-suite event has no captured object-allocation trace. Use the same pre-optimized measuring-driver policy for that warmed test too, retaining its zero-byte limit and all ownership assertions. The final full-suite and independent CI runs must pass with normal production runtime settings.
• The parallel branch run of a509bf3 reported 98,352 counter bytes in the framed acknowledgement receiver's zero-allocation test, while the PR run and both fresh local diagnostic processes passed its zero/positive rows. No object trace was captured for that hosted event; it is not attributed as a new production allocation. Apply the same pre-optimized measuring-driver policy to that warmed gate, retaining its existing no-GC scope, exact zero check and real-allocation positive control. Require fresh full CI on the final revision.

• The next PR runner reported 3,840 counter bytes in the nominal speaker-frame gate while the parallel branch suite passed. That gate had only two warmup calls and omitted the bounded scope already used by other resampler gates. Pre-optimize its measuring driver and wrap only the measured counter pair in the existing strict scope. Add a non-inlined real heap-allocation positive control on the same workload. Keep the original 1,000 measured frames and zero-byte assertion; do not claim an object trace for the original CI event.

• Hosted run 39291077961 failed the paced native-source fixture's exact PCM comparison (expected 29, received synthesized zero), while the parallel PR run passed. The fixture refilled both media inputs on its native-source thread despite claiming continuous independent media. No production failure trace was captured. An added pause extending beyond the 16-frame buffer deterministically failed the old fixture's continuous-media prerequisite. Feed synthetic media from its existing HID-submission callback instead, so native-source scheduling cannot interrupt a refill between the two inputs. The new pause case then passes, as do the original 100/200 Hz cases. Retain exact PCM/zero-boundary comparisons, command ordering and backlog limits; additionally require a successful final drain. No production code changes.

• The follow-up branch runner 39298056566 passed cadence but measured 24 bytes in the warmed native-busy retry allocation gate; the PR suite passed. A cold local run reproduced the 24-byte failure, while a separate diagnostic child process with OSR disabled passed both zero and positive-control rows. Apply the existing pre-optimized measuring-driver policy to this test method as well. Preserve its 2,000 warmup calls, 20,000 measured retries, bounded measurement scope, exact zero-byte assertion and real-allocation control. No object-allocation trace was captured and no production JIT setting changes.

Automated evidence¶

Local Windows 11 x64 checks:

The existing CI workflow uses global.json and performs its own clean restore, both complete test suites, portable packaging, translation probing, installer layout, MSI install/repair/uninstall and upgrade/profile-preservation checks. Consult the migration PR's checks for their current status. CI runs on Windows Server 2022; passing it does not replace Windows 10/11 desktop testing. The test-results artifact also includes the sidebar's light/dark scale renders.

Desktop and hardware acceptance¶

On 19 October 2026, the maintainer reported that the updated PR #56 test build was working correctly and approved merging and releasing it as 5.0.35. No separate Windows 10 edition/build result was supplied; do not infer exhaustive Windows 10 or hardware coverage from that acceptance. Both final hosted workflows passed, including installer lifecycle and legacy upgrade/profile preservation checks.

The following checklist remains available for future regression testing of the .NET 10 package. Earlier testing of 5.0.34's .NET 8 build alone does not qualify this migration.

1. Close the existing app, games and VIIPER. Use a separate extracted test package and a copy of your profiles. Confirm Log > Copy Text shows .NET runtime: .NET 10....
2. On Windows 11 x64 and an explicitly recorded Windows 10 x64 edition/build, open Settings, Profile Editor, Log and Changelog in light and dark themes. Check resizing, links, numeric controls and theme switching. In Settings > Appearance > Sidebar icon, select each of the five icons, switch themes and move the window between differently scaled displays. Confirm the 44 × 44 layout stays sharp, then restart and verify the choice. Existing settings without a choice should start with Blue Tile, first in the list; the tray icon setting remains separate.
3. Load an existing profile, change a harmless setting, save and reopen it. Confirm mappings, gyro and polling selections remain correct.
4. With a DualSense over USB, emulate DualSense and use Test Heavy and Test Light. Check native vibration/triggers in DEATHLOOP and vibration in GTA V Enhanced. Repeat over Bluetooth, including disconnect/reconnect.
5. Check controller speaker output and Audio Haptics with a familiar sound; try both system and application capture, plus custom frequency filters. Confirm starting/stopping capture and exiting the game remain responsive.
6. On a disposable test machine/VM, install, upgrade, repair and uninstall the test installer. Confirm it needs no separate .NET runtime and preserves saved profiles. Avoid uninstall experiments on a working setup.

Record OS edition/build, transport, controller, test-build identity and the full copied log with the results. Windows 10 support limits are described in Development.