[DeathKhan]

What and why

Nintendo Switch 2 Welcome Tour (040057C02159A000, SDK 20.5.29) stops at start-up on SDK calls that talk to a system service whose session the host never opens, and on a red stub that leaves a size query unanswered. #1425 (its profile) takes over three of those calls: vi::InitializeWithVrrLfcEnabled, hid::SetSceneIdForPrivateApp and oe::QueryApplicationGpuTimeSpan. This PR has the others, with no overlap with #1425.

They come from DeathKhan/nx2android (cbeb5ccd6, eea4efef9, 87772ede9), an Android port of nxnative, as reviewed into DroidPlay Core: enjoyerofslop/droidplay-core#158 and #157. Its maintainer asked for this matching PR, since these files are shared.

The system calls (hle_vi.c, hle_hid.c, hle_oe.c):

• vi::QueryDisplayCurrentRefreshRate: answers 60 Hz, the display's own, as one 32-bit Hz word and a bool. An answer of two 64-bit words overwrote the caller's saved frame pointer (main+0x5e3e20).
• vi::QueryDisplayAvailableRefreshRates: lists 60 Hz alone.
• vi::SetDisplay120HzRefreshRateCapability: is logged.
• hid::GetGlGrButtonConfigState: answers no GL / GR assigned. Its SDK body aborts in sf (0xe15, sdk+0x65cfdc).
• gs::sdk::CreateNativeMemory / DestroyNativeMemory: the handle is the range's address, and nothing is registered. Both are only seen as imports so far, never called, so the 8-byte handle isn't checked yet.

nvnDeviceCalculateAccelerationStructureMemoryRequirements writes its three sizes:

• The bug: the game sums the structure, build-scratch and update-scratch sizes into the memory pool it makes for its acceleration structures (query at main+0x6130cc, the sum at main+0xfc2cc / +0x11bf1c), and doesn't zero them first. The red stub wrote nothing, so on the Android host the sum was stack garbage: a 4 GB pool, and the game ran out of memory in its loader 16 frames in. On the Mac the result depends on what the stack holds.
• The fix: it now answers 64 KB for each, with a red line once.

Tests: t_oe checks the native memory handle.

History: an earlier version also took over #1425's three calls. They were dropped in 3cf027f, so the two PRs don't collide.

Checked

• Tests, on an Apple-silicon Mac (macOS 27, Apple clang), built from this branch with GAME 0100606026F12000:
  • t_oe: 54 checks, 0 failed.
  • t_hid: 209 checks, 0 failed.
  • Built with the Makefile's own compile lines, from an export of dm/host and the game's profile. Not run: the whole make -C dm/host/tests run.
  • hle_vi.c, hle_hid.c and hle_oe.c pass clang -fsyntax-only -Wall -Wextra. hle_nvn.c's only warnings are existing ones on other lines.
• Tool tests: dm/tools/tests test_docs, test_env_switches, test_lint_c and test_architecture pass. docs_symptoms.py was run.
• Games run: none on the Mac host.
  • On DeathKhan/nx2android (Android, Ayn Odin 3, Adreno 830): Welcome Tour reached its title screen with these calls, game: Nintendo Switch 2 Welcome Tour plays on the Mac #1425's three and the acceleration-structure answer.
  • regress.py check: not run. These hooks are reached only by Switch 2 games that call them.
• Devices and OS: the Mac above, for the unit tests.

Docs

• The finding is in its docs/reference/services.md and docs/reference/nvn-objects.md entries, and python3 dm/tools/docs_symptoms.py was run
• No game files, keys or data extracted from a game in the change or the attachments

🤖 Generated with Claude Code

[DeathKhan]

Updated in 3cf027f, a new commit on the same branch:

• Dropped: InitializeWithVrrLfcEnabled, SetSceneIdForPrivateApp and QueryApplicationGpuTimeSpan, with the GPU time's weak stand-ins, its frame-rate.md line and its addition to the DM_DYNAMIC_RESOLUTION row. game: Nintendo Switch 2 Welcome Tour plays on the Mac #1425 takes over the same three calls with the same C names, so whichever landed second would not build. This PR now has no overlap with game: Nintendo Switch 2 Welcome Tour plays on the Mac #1425.
• Kept: the refresh rates, the 120 Hz capability, GL / GR, native memory (with its never-called note) and the acceleration-structure sizes commit.
• t_oe: keeps only its native memory check.
• Title and body: updated.

Checked, re-run on an Apple-silicon Mac from this branch:

• t_oe: 54 checks, 0 failed.
• t_hid: 209 checks, 0 failed.
• clang -fsyntax-only -Wall -Wextra is clean on the three changed host files.
• test_docs, test_env_switches, test_lint_c and test_architecture pass.

Not run: the full make run, any game.

[DeathKhan]

The one failing check (t_save_sync, 1 of 106) is not from this PR and is not flaky: it fails the same way on main. The check at dm/host/tests/t_save_sync.c (main, line 346) expects cdevice=Mac, but on the GitHub macOS runners the device model is VirtualMac2,1, so strstr(line, "cdevice=Mac") never matches:

✗ both changed: the report says conflict, with the cloud's device (... cdevice=VirtualMac2,1 sametitle=1 conflict=1 pull=0)
== save sync: 106 checks, 1 failed ==

Same line in both jobs (Host unit tests, Sanitizers) of this PR's run 43218207071 and of main's run https://github.com/seedds/nxnative/actions/runs/43326823244 (8d44aaf, 2026-11-12); everything else in the conflict report is as expected (conflict=1). This PR does not touch save sync (its files are hle_vi.c, hle_nvn.c, hle_hid.c, hle_oe.c, t_oe.c and docs), so merging main would not change the result. Matching the model on a real Mac and on the runner (for example checking that the cdevice= value contains Mac) would fix it on main.

[DeathKhan]

The host and sanitizer failures here are main's own t_save_sync check (cdevice=Mac vs the runner's VirtualMac2,1); #2608 fixes the test.