nxnative-easy/dm/games/0100000011D90000/README.md

Pokémon Brilliant Diamond (0100000011D90000)

Update 1.3.0 (v393216) over the base game. Unity 2019.4.27f1 IL2CPP, SDK 11.4.3, NVN, Wwise and FMOD. Modules main, subsdk0 (NintendoSDK_movie) and sdk; the latter two are the builds used by Metroid Dread. Plays on the Mac at 60 fps with 1.3.0. An unsigned IPA is built; iOS is untested.

What it needs

Setting	Where	Why
SDK node area 1 GB → 62 MB	dm/tools/sdk_patches.py, build 2d369d99…, sdk+0x34e0c8 / +0x34e164	StandardAllocator::Initialize aborts before nnMain without it; see Memory
5 GB block: mem_reserve 0x140000000, mem_alias 0x118000000, mem_stack 0x10000000	game.tsv	virtual address memory: StandardAllocator and Unity reserve regions independently
app_pool 0x80000000 (2 GB)	game.tsv	1 GB exhausts the game's malloc while loading; the hardware pool leaves little alias space beside StandardAllocator
sys_resource 0x1000000	game.tsv	the NPDM's 16 MB system resource selects virtual address memory
fps_target 60, present_interval display	game.tsv, build-gated engine limiter	raises Unity's own cap and the interval-2 present cap; see Frame rate

No game-specific code cuts. DM_FPS_TARGET=30 DM_PRESENT_INTERVAL=game restores the original 30 fps.

Memory

• Resolved: abort before nnMain. nninitStartup (main+0x2c1cf0) queries pool minus used memory, then calls StandardAllocator::Initialize(nullptr, size) at main+0x2c1d5c. It never sets a heap on the virtual-address-memory path. With the hardware pool and the original 1 GB node area, CentralHeap::Init fails: Break at sdk+0x39fbd4, backtrace sdk+0x145038, sdk+0x145198, sdk+0x2f98e4, main+0x2c1d60.
• The SDK's AddressRegionManager constructor (sdk+0x34e0b8) loads the node-area limit at +0x34e0c8, and registers its occupied length at +0x34e164. Both are mov w8/w9,#0x3fe00000. The same two cuts as Engage's SDK shrink them to 62 MB. A repeat boot then reaches nnMain without a Break.
• With the original 3.21 GB application pool and the node-area cut, mapped regions already reach +3.90 GB of the 4.375 GB alias region during startup; only 0.50 GB is committed in the first 12 seconds.
• Resolved: 1 GB pool exhausted during loading. Both a windowed run with audio and a headless silent run crash as commits reach 1.00–1.01 GB, on null reads at main+0x2605cc (IL2CPP) and main+0x578624. A 2 GB pool passes this load and reaches the language menu and opening dialogue. At that point the alias high-water mark is +2.75 GB and 1.01 GB is committed. Over the 30-minute opening / Twinleaf Town run the high-water mark stayed at +2.75 GB, with up to 1.50 GB committed. The house settled near 2.24 GB resident / 0.87 GB Metal; outdoors near 2.50 GB / 1.01 GB, with scene-related steps rather than sustained growth.

Files and shaders

• RomFS: 18,022 files, 4.71 GB, including 14,052 UnityFS bundles (3.83 GB), Wwise banks and streams, and 18 WebM movies. The player's managed metadata is Data/Managed/Metadata/global-metadata.dat.
• The split finds 14,441 stage shaders (7,247 vertex, 7,173 pixel, 20 compute and one geometry), including 11 GLSLC outputs in main. All 14,441 compile for Mac and iOS, with no unresolved bindings or unfillable driver-bank reads; 47 shaders read c2 as data.
• Resolved: missing skinning compute shader, black 3D behind the dialogue. Stripped ComputeShader objects in Data/Resources/unity default resources have type hash 1a95529ac2eb1a97cd9d2a4356bd722f. unityfs.py now has Unity 2019.4.27f1's release layout, checked against AssetRipper's dump: 116 nodes with matching alignment flags, both objects read to their exact ends. Unlike newer Unity versions, each named kernel holds its code directly. Internal-Skinning contributes 16 kernels and Internal-BlendShape three; the missing runtime key 2c1e06378edbd813 matches Internal-Skinning#main.0 exactly. With these kernels the professor, character selection, household and outdoor character models render correctly, with no missing shader in the tested route.
• The rewriter handles six offset jump tables and one indirect call through x18 in main, with no additional rewriter mechanism.

Frame rate

• Resolved: gameplay returns to 30 fps after 60 fps menus. The game's IL2CPP initialization loads 30 at main+0x2bc13f8 and calls the Application.set_targetFrameRate wrapper at +0x26a0240. The engine's icall registration at +0xe34e60 resolves to +0xdc8e70 → +0x5fa640, which stores the integer at main+0x4afc868; its getter at +0x5fa630 reads the same field. The host's integer-FPS limiter row is gated on main build 94ceae32… and keeps it at fps_target, even when the game resets it.
• Display pacing alone leaves that 30 fps engine cap. The profile also uses present_interval display, because the game asks for interval 2 independently. A headless Mac run with both changes reports 60.0 fps outdoors (frame-time median about 16.5 ms). DM_WINDOW=0 still creates a window because the Mac front end tests whether the variable exists; timing comparisons omit it and use DM_PRESENT_LINK=0 to avoid occluded-window display-link stalls.
• Current: movement and collision comparison passed. Separate launches from the same save, at 30 and 60 fps, received LDown~0.5, LRight~1, LUp~1, with screenshots before and after each hold. Static-background phase correlation found offsets of (0, 0), (0, 1), (4, 1) and (-1, 4) pixels respectively: the same route and wall contact, within frame/input sampling differences. This checks that walking does not double in speed; it does not establish correctness of every later physics system. Captures are run/shots/rate{30,60}-{start,down,right,wall}*.png in the build folder.

Mac testing

Build with scripts/build_ipa_from_nsp.py <NSP folder> -k <prod.keys> --platform mac; its final line runs the game with the correct paths. Saves are under <NSP folder>/0100000011D90000_build/run/save. Use DM_CTL=<file> for input and screenshots, and DM_KBD=<name> for scripted name entry. Logs, screenshots and analysis stay in the NSP folder.

Testing

• 2026-10-26, Mac: a fresh save through language selection, Rowan's introduction, character and name entry, the player's room, downstairs dialogue, Twinleaf Town movement and the rival encounter. The 30-minute run had no red line; its final SIGALRM was the test timer. Both audio outputs reported zero underruns.
• An extended reload run reached the rival's upstairs room, Route 201 and Verity Lakefront. It continued for over 90 minutes at 30 fps, with about 2.68 GB resident / 1.10 GB Metal in the lakefront scene. After about 58 minutes the audio watchdog intermittently reported device callback stalls and restarted the AudioUnits; each recovered, with no game hang. The cause of those host-device stalls is unconfirmed.
• Autosave created SaveData.bin and Backup.bin (979,108 bytes each); a separate launch reloaded the player outside the house correctly.
• Reloading exercises the VP8 logos and opening movies. The host now decodes the real movies without errors and displays their pictures, with no depth-test override. A 106-second sample decoded 3,182 packets at about 1 ms per packet.
• The final rebuilt host passed a three-minute boot with real audio muted: VP8 decoded, 30 fps, both outputs reporting zero underruns and no red line. SIGALRM ended the timed run.
• The 60 fps update passed separate 30/60 movement comparisons and indoor/outdoor transitions with autosave. The host's integer cap changed 30 → 60 as intended and outdoor logs report 60.0 fps. The extended 60 fps route run had one 211 ms Core Audio callback stall; the existing watchdog restarted the AudioUnit and playback recovered, as in the earlier 30 fps run.
• The final unsigned 60 fps IPA is 4.6 GB (4,965,541,636 bytes), version 1.3.0, bundle ID org.nxnative.0100000011d90000; it contains no code signature. Its 21,392 Metal libraries cover all 14,441 stage shaders and their required variants. The packaged executable contains the new Unity limiter and its exact build gate.
• PR #190 independently found the shared SDK cut and the same 19 built-in compute shaders on 1.0.0. Its shader tables describe that build; this profile uses the 1.3.0 split. Its AudioOut-stop / FMOD finding is included as a general fix, with a guard preserving buffers queued before Start; the 1.3.0 run passed character selection before that fix.
• Host tests pass, including synthetic VP8 / VP9 NV12 and RGBA decoding, AudioOut stop / queue stress and integer / float / millisecond / vsync frame-limit regressions. The graphics Python suite passes (41 tests, two skipped); Metal depth / stencil clear, stencil-view, read-only depth and upload-budget regressions pass. Rewriter patch, chain and documentation checks pass.

Found on this game

• SDK 11.4.3's instance of the SDK's address-region node area, with the cut keyed by the shared SDK build ID.
• Legacy kernels held directly in Unity ComputeShader objects.
• VP8 decoding through libvpx.
• Zero initialization of depth and stencil textures.
• AudioOut stop releases pending buffers, reported on 1.0.0 in PR #190.
• Unity's integer target frame rate.

Open

• FPSLocker Warehouse flags physics issues above 30 fps for this exact build. A short movement check cannot rule out every physics issue in later areas.
• First battle and later areas remain untested.
• User testing on iOS remains required.
• Recheck the intermittent Core Audio callback stalls during a long run on the final host build.