Pokémon Scarlet

Base game 1.0.0 (v0), SDK 13.3.5, Game Freak's Trinity engine on NVN. Playable on the Mac: language and character selection, the academy opening, the native Bink movie, getting up in the bedroom, walking downstairs and through the living room and kitchen, camera control, dialogue and save/reload. The unsigned 1.0.0 IPA is built (~7.0 GB); iPhone testing is pending. Only 1.0.0 is registered, and the profile's shader tables come from v0. When the NSP folder also contains updates, build the supported version with scripts/build_ipa_from_nsp.py <NSP folder> -k <prod.keys> -p v0 (add --platform mac for a Mac build). The builder otherwise selects the newest supplied version. The supplied 4.0.0 update is damaged and remains unsupported (see Files and builds). Pokémon Violet is its twin: the same sdk and subsdk0 builds and the same shader tables, with a main of its own.

Memory

• Resolved: startup aborted in StandardAllocator::Initialize. nninitStartup (main+0x12ded10) calls the helper at main+0x12df040, which initializes the allocator at main+0x43853d0 with (nullptr, 0x300000000, true). A probe on sdk+0x324aa8 confirmed the 12 GB request; the abort's backtrace ends at sdk+0x324b7c, main+0x12ded4c. This is the same compact-address-space limit as Legends: Z-A. A build-gated cut at main+0x12df074 first reduced the virtual reservation to 4 GB. Current: the host's StandardAllocator cap, no patches.py (StandardAllocators over 12 to 62 GB): the gate gives the 12 GB request 4.06 GB at alias +0.06 GB. Checked 2026-10-30 on 1.0.0 without the cut, headless from a fresh save with the route below for 480 s, to frame 9,900 in the opening (the get-up prompt comes at about 10,000; screenshots of the Paldea narration and the flyover): the alias region was used up to +4.12 GB of 4.38 GB, no region failed, and the only [RED] line was the conservative-raster diagnostic. Physical commits remain bounded by the application pool.
• Resolved: the 4 GB request still failed with 1 GB placement alignment. The existing 62 MB node-area cuts apply because SDK build 3644a676… is nirvanA Initiative's. Its allocation helper (sdk+0x37c2d0) rounds regions of at least 1 GB up to a 1 GB boundary, losing enough space to prevent the 4 GB reservation from fitting. The additional cut at sdk+0x37c33c used 32 MB alignment, as for the equivalent SDK build in Sparks of Hope; the SDK-wide rules now give it 2 MB, as below. Decompilation confirms region size and page guards are unchanged. The same startup check fails with only the main cut and reaches nnMain with both. Since Subnautica: Below Zero (010014C011146000) the build aligns its 4 MB, 32 MB and 1 GB classes to 2 MB, as SDK 20 does: since 2026-10-30 from the SDK-wide alignment rule (the row was removed).
• Resolved: freed startup pages stayed mapped past the application pool. Startup probes all allocatable memory, frees it, and leaves those pages in the SDK's cache. Without mem_map_limit, a physical-map census reaches 0xd1f41000 bytes in 8 seconds, above the 0xCD500000 pool even before other process memory is counted. With the existing limit enabled, one 512 KB map fails at a full pool, the SDK unmaps its 0xc8a40000-byte cached block and retries. The official rewriter outputs and rebuilt host pass the startup check with this setting.

Shaders

The base game's Trinity packs and main contain 30,142 unique shaders: 14,999 vertex, 15,057 pixel and 86 compute, with 24,487 variation pairs. There are 40 shaders embedded in main. The existing Trinity reader and the game's own Oodle decoder extract them. Chain 157e1b96c8da, profile inputs 6b273274b0eb: all 30,142 compile for Mac and iPhone, with no undecoded bindings. The complete build supplies the stamped c1_full.tsv and cb_extent.tsv; no shader override is needed.

The 4.0.0 split aborts on 195 malformed active Trinity packs in the damaged input described below. Full 4.0.0 shader coverage and an update IPA have not been produced. Its partial startup used the base shader libraries and reported missing update shader keys.

Pipeline preparation depth

Resolved: the game calls nvnDeviceSetDepthMode with near = 0 and uses main0_z vertex entries. The profile's depth_mode zero makes the offline prebuild prepare those same bases. The first full regression exposed the missing setting: its tool prebuild compiled 24,481 bases for near = -w with zero failures, then the game opened those archives but rebuilt the pairs for near = 0 before presenting a frame, missing the route's deadline. The corrected profile then passes the unchanged 180-second route: first frame in 1.9 seconds, frame 5,324, 29.9 fps, p50/p99 33.28/38.54 ms and no dropped frames under Metal API validation. The shader input digest is unchanged. This is the existing per-depth-mode entry-point mechanism.

Mac routes and validation

• A fresh save reaches the bedroom with DM_KBD=Taylor and DM_KEYS="f300:A,f600:LUp+LLeft~0.3,f660:A,f1000:A,f1300:A*90/90". This selects English and the blonde girl, enters the name, confirms the character and advances the academy text. Let the opening movie finish. The get-up prompt appears around frame 10,000; LDown~2 starts getting up.
• The bedroom save continues with DM_KEYS="f300:A*5/150". The final presses can open the bag-and-hat dialogue; close it with A before walking. The left stick moves, the right stick turns the camera, and A interacts. The iOS profile starts with the full 3D control layout for these controls.
• smoke.tsv uses an isolated, fixed copy of the save just after getting up at <NSP folder>/0100A3D008C5C000-work/smoke-save. Make that copy before running the regression harness; the save stays outside the repo. regress.py capture records its frame 3000 for regress.py replay: 1 of 8 replays equals the live frame, noise 7 pixels (2026-10-30, chain f3ceca86cac5; 4 of 8 and 20 pixels, 7 of 8 and 18 in earlier captures). The capture takes 53 s (a frame capture runs under the standard thread policy, resolved; 1,178 s before). The first captures aborted the replay (a replay's own objects take serials above a capture's, resolved). Texture IDs 1726 to 1786 in the bedroom stay registered after the game finalizes their textures' memory pools (each texture has a pool of its own) and reuses the pool objects' memory. The capture read texture data there as pool addresses and once stopped incomplete (a captured texture ID whose memory pool was finalized, resolved).
• Initial gameplay validation, 2026-10-27, chain 7fbca6dd477d: the complete Mac shaders render the character, academy, movie and bedroom. A fresh 10-minute run and a 20-minute reload run under Metal API validation reached gameplay without a GPU fault, missing shader or validation error. The reload exercised walking, camera rotation and dialogue. A further reload walks down the stairs and around the lower-floor living room. Its longer run met the compiler-cache error below; a fresh 230-second replay reaches the kitchen and renders Mom with only the conservative-raster diagnostic. In the stationary bedroom, frame times stay around 33.3 ms (30 fps), resident memory settles around 3.0 GB and GPU allocations around 1.37 GB; this is an indoor smoke observation, not an outdoor benchmark.
• Release route, chain 157e1b96c8da: all 30,142 Mac shaders compile; the full bedroom route passes under Metal API validation with the numbers above. The repeat using current main's RomFS image mount also passes: first frame in 1.6 seconds, frame 5,345, 29.9 fps, p50/p99 33.26/39.02 ms, peak footprint 4,329 MB and no dropped frames. usage.tsv combines this full run with the shader features. Its two stationary screenshots show the character, room and bag-and-hat dialogue. The sole red diagnostic is conservative raster dilation. A preceding run had repeated display-link stalls and missed the frame gate; the unchanged repeat passed. The cause of that run's display-link stalls was not isolated.
• Build and checks: the unsigned IPA contains 30,142 primary iOS shader libraries (41,025 including variants), a 6,678,553,818-byte RomFS image and the expected main/SDK cuts. All ZIP CRCs pass. The Python suite passes 449 tests (3 skipped), as do the host and full Metal graphics suites; pinned lint/types report no new findings. The final comparison of 75065df74762 against 7b00b952abc1 covers 14 available local builds: 13 pass, with 192,870 shader records, unchanged MSL and no new shader or interface failures. ANIMAL WELL's full runtime comparison holds 59.9 fps, with screenshots and frame times within the harness tolerances. Balatro is an unsupported OpenGL ERROR, so the whole report is not labelled PASS. BotW's static checks pass; its unavailable fixed Zora save was explicitly skipped for runtime. Final Fantasy Tactics passed the earlier static comparison (78cb26d91373 against 987759c2a218), but its local build is no longer available for the final sweep. Detailed logs and the exact coverage record stay in <NSP folder>/0100A3D008C5C000-work/release-1.0.0/; the final report is reports/full-final-2/report.md there. All four pre-push CI jobs pass on the tested production source. The subsequent frontend, Mac-launcher and split-progress integration leaves runtime, shader-chain and iOS sources unchanged; 86 affected tests and pinned lint pass, and the generated iOS launcher settings exactly match the verified IPA.

Pipeline cache ownership

• Current: a longer indoor run reported 20 pipeline/transform-feedback creation or archive-write failures naming a missing com.apple.gpuarchiver/gpuarchiver-db9690/PersistentState. The log does not identify the process that removed it. The subsequent 230-second replay with folder-attribution probes reaches the kitchen without this error; new-library and live-base-after-flush single-process checks also pass. These results do not prove the original failure cannot recur.
• Resolved: compiler-cache ownership defect. The investigation reproduces deletion of a foreign PersistentState in both an isolated fixture and the real-Metal archive test. The latter fails before the shared fix and passes 73 checks plus separate-process compaction/restart checks afterward. See Metal's build folders for the mechanism and fix; these tests do not identify the original run's deleting process. The initial chain's rebuilt host's 450-second replay (frame 13,200) reaches Mom's kitchen dialogue and commits two saves under Metal API validation, with no missing-cache or pipeline-creation error; conservative raster dilation remains the sole red diagnostic. It ends at the harness's intentional deadline.

Files and builds

• Current: the base NSP passes extraction integrity checks: 24 RomFS files, 6,369.2 MB. Modules are main (ca44e7dd…) and sdk (3644a676…). The host supplies the imports from Nintendo's movie module subsdk0.

• The complete, CRC-checked 1.0.0 IPA remains preserved at <NSP folder>/PokemonScarlet-1.0.0.ipa.

• Blocked: supplied update 4.0.0 (v786432) has erased payload bytes. Independent checks trace 2,633 invalid 16 KB data blocks to three all-zero ciphertext ranges in the original NSP: 2,629 blocks in arc/data.trpfs and four in audio/wwise/bin/Switch/DEMO.bnk. The ranges below use absolute NSP offsets, with exclusive ends; they total 41,615,360 bytes (39.6875 MiB).

  Start	End	Bytes
  0x12c220000	0x12d000000	14,548,992
  0x12d300000	0x12e000000	13,631,488
  0x12e330000	0x12f000000	13,434,880

  The program NCA's signed CNMT expects SHA-256 7e5b9bbf…; its actual hash is 8ccebbb9…. The metadata and program NCA header signatures are valid. The existing external checker's SignaturePassed result covers those signatures, not a recomputation of the payload hashes: signed headers and their hash declarations can survive while the ciphertext they describe is erased. Expected data-block hashes are nonzero, neighboring valid blocks match independently, and relocation reconstruction reproduces all 2,633 failures. This proves source-content damage rather than an extractor-only failure; the event that erased the bytes remains unknown.

  The damage intersects 318 active Trinity packs, with 195 malformed pack metadata records aborting the full shader split. Base data reproduces the signed expected hash for only three blocks; 2,630 remain unrecovered. No original or extracted game bytes were changed. The two affected files were scanned completely at data-block level; the other files passed the earlier extractor checks. Full shader content cannot be certified from this copy. Detailed evidence is in the work folder's update4/integrity/final-evidence.json.

  User-authorized --allow-damaged extraction writes all five ExeFS and 30 merged RomFS files (9,465.4 MiB of RomFS) under extract-4.0.0-retry-*/unverified/, retaining UNVERIFIED.txt and extraction results. Its code builds, main 421c5411… and sdk bb560bdf… (SDK 13.3.13), were probed with experimental cuts and a Bluetooth LE adapter outside this release. That partial startup reached the language screen through frame 3,475 using base shaders; it was not gameplay validation. The investigation is preserved under the work folder's update4/ and release-1.0.0/before-scope-trim/. No 4.0.0 build or cuts are registered in this profile. An intact update source is required for its full shader build and gameplay validation.

• The supplied DLC is a licence-only NSP, with no data to extract.

Open

• Conservative rasterization is unimplemented. The enable setter is ignored and the dilation setter reports [RED] NVN unimplemented function called. Its visual contribution has not been isolated; the diagnostic has not been suppressed.
• Outdoor exploration and battles have not been validated yet.
• iPhone gameplay, device memory and touch controls are untested. Audio runs through both Wwise output queues on the Mac without sustained underruns; listening on the iPhone is pending.

Found on this game

• StandardAllocators over 12 to 62 GB of address space.
• 1 GB alignment of regions of 1 GB and more.
• Maps past the application pool fail as on Horizon.
• A replay's own objects take serials above a capture's.
• A captured texture ID whose memory pool was finalized.
• A frame capture runs under the standard thread policy.
• A capture refused after a change to the reflection structs.
• Conservative rasterization remains unimplemented.
• Metal's build folders left by pipeline archive flushes.