[JakePre]

What and why

Two commits, in the commit-message form of docs/agents/06-dev-loop.md § 8: host: answer two Switch 2 SDK calls Welcome Tour stops on (the two hooks, with their reference entries naming the game as "no profile yet"), then game: Nintendo Switch 2 Welcome Tour plays on the Mac (the profile, its note, the lists, and the links to its note). Each passes the host tests and test_docs on its own.

Nintendo Switch 2 Welcome Tour (040057C02159A000, update v262144, SDK 20.5.29, NintendoWare's Bezel Engine) is a new game. It now plays on the Mac past its intro video, through the tour, its menus and exhibits, at about 60 fps. It stopped, in turn, on these:

• vi::InitializeWithVrrLfcEnabled at boot: main answers it since 11825d2 (Resident Evil 4), so this PR no longer adds it; the game commit adds Welcome Tour to that entry's Found on.
• Bezel's allocator at boot, out of room in profile_guess.py's 5 GB block (a 4.4 GB alias region): the profile takes the Switch 2's 9 GB layout from Splatoon Raiders and Star Fox (mem_reserve 0x550000000, mem_alias 0x400000000, app_pool 0x240000000). The game takes Bezel's virtual-memory path (Bezel Engine's two memory setups).
• hid::GetJcmsState each frame, and GetMousePointerSpeedScale: main has both since the Joy-Con 2 mouse sensor (413e880), so this PR no longer adds them.
• hid::SetSceneIdForPrivateApp as each scene module starts: the SDK's own connects to the hid service. Nothing reads it, so it's logged.
• oe::QueryApplicationGpuTimeSpan after the first frame (hle_oe.c, main+0x30a83c): the SDK's own locks the gs graphics server's context, as SetGraphicsFrameLengthMode does. It now returns success and a fixed 8 ms, under a 60 fps frame: the host has no GPU time per application, so a game that sizes its picture by the span stays at full size, and a span that is not 0 takes no divide or "not measured" path in some other game's controller (DeathKhan's suggestion below). The game reads the span only on success. Welcome Tour ran the same with 0 and with 8 ms on the Mac: title screen, 16.67 ms frames once loaded, no crash.

Files added:

• the profile (dm/games/040057C02159A000): from profile_guess.py's, with the memory above, the game's name, and v262144 as a placeholder display version (no control data gave one)
• its note, a games-table row and a line in README's Switch 2 list
• reference entries for the new host calls (SetSceneIdForPrivateApp now has an entry of its own); docs_symptoms.py was run

Shaders: 6,766 (the tables were split again on 7 November: main's mesh reader now finds 7 mesh shaders and a task shader of forward_plus.bfsha, with 14 more pairs). On chain 3df44a5a4e0c (9 November), 29 fail to translate (28 on Ampere instruction 0x3ad, one AST form), and 9 lose bindless accesses (429 deleted). The note lists them as open, with the ray tracing NVN calls (unimplemented, [RED] once each at boot).

Controllers. The exhibits use the Joy-Con 2's mouse sensor and HD rumble, which are on main now (#1346, #1349, #1363 and #1426). The note says so.

Checked

• Tests (rebased on main 7479adf, 9 November):

  • make -C dm/host/tests GAME=040057C02159A000 run: all pass.
  • dm/tools/tests (test_docs.py, test_game_profile.py, test_settings.py, test_launcher_settings.py): pass.
  • lint_c.py tidy (LLVM 23.1.5): 0 findings above the baseline.
  • make -k -C dm/gfx/tests GAME=040057C02159A000 run: everything passes but t_render_counter_overlap (1 or 2 of 42 checks, CopyBuffer0) and t_render_enable_queue (9 of 46). Both fail the same way on plain main 7479adf on this M1 Pro, and this PR touches no NVN code.

• Games run: Nintendo Switch 2 Welcome Tour, v262144: about 10 minutes from boot through the intro video to the tour and its exhibits.

  • Frame time p50 16.7 ms, p95 20 ms over 300 frames.
  • The exhibits ran with the controller PRs on top, with a left and a right Joy-Con 2.

• Devices and OS: M1 Pro, macOS 27.2.

• regress.py: I didn't run check.

  • impact names Hades, Breath of the Wild, Pokémon Scarlet and Tears of the Kingdom, through vi::Initialize. That's because the new vi hook sits next to it. vi::Initialize itself is unchanged.
  • The other host changes are hooks no other game's manifest calls.
  • This Mac has no game set up for the harness: regress.py list finds none, since my games are archived off it.

  If you want check on any of these, tell me which.

• Not checked: an IPA (the 16 GB alias region is over iOS's limit, as for Splatoon Raiders).

Docs

• The game note, its games-table row and README's Switch 2 list; reference entries for the new calls, and docs_symptoms.py was run
• No chain tool changed
• No game files, keys or extracted game data. The profile's tables are shader keys and pairs, as other profiles have.

🤖 Generated with Claude Code

[DeathKhan]

A thought on QueryApplicationGpuTimeSpan, from bringing Welcome Tour up on an Android port (DeathKhan/nx2android, also offered to DroidPlay Core as enjoyerofslop/droidplay-core#158):

• What we answer: a fixed time, 8 ms, by default. The host's measured GPU time is answered only when the player turns Dynamic resolution on (DM_DYNAMIC_RESOLUTION=1, as for the games with the DYNAMIC_RESOLUTION capability). There we take the newest completed frame up to 8 back from mtl_gpu_ms, else 8 ms.
• Why: the game reads the span to drive its dynamic resolution. The host's GPU time is what held Ocarina of Time's controller at its floor (frame-rate.md, "Dynamic resolution at its floor while the GPU has time to spare"). A fixed answer under the frame keeps the game at full size, as Dynamic resolution off does elsewhere.
• Compared with 0: 0 reads as "free" too, but a fixed non-zero span is a little less likely to trip a divide or an "unmeasured" path in some other game's controller.
• What we ran: Welcome Tour ran to its title screen on the Odin 3 with what amounted to a fixed 8 ms. The measured reading itself hasn't run with a game.

We've dropped our own takeover of this call, InitializeWithVrrLfcEnabled and SetSceneIdForPrivateApp from #2246, so this PR keeps them. One more small point from DroidPlay Core's review: returning a Result of 0 from InitializeWithVrrLfcEnabled and SetSceneIdForPrivateApp is safe whether the real functions return void or a Result.

[JakePre]

Thanks, switched: the PR now answers a fixed 8 ms (542ef26), for the reason you gave, that a span that isn't 0 is less likely to send another game's controller down a divide or "not measured" path.

What we ran on the Mac (M1 Pro), the same Welcome Tour build with only that line changed: both 0 and 8 ms reach the title screen and settle at 16.67 ms frames (p95 about 17.1 ms), with no crash and the same warnings in the log. The game asks about 5 times a frame, with spans of roughly 0.3–1 M ticks once it's running. Neither run went past the title screen, so whether the game resizes anything in the tour itself is untested.

The measured reading behind Dynamic resolution sounds right as a follow-up of its own; this PR keeps the fixed answer only.