nxnative review: what Napple can take (review only, 2026-10-25)¶

Branch claude/nxnative-review-2026-10-25-titles, session nxnative-review (Mac). Nothing was implemented, built or run. No code was copied from nxnative; every claim below cites a file I read, with line numbers where I have them. Paths: NX: = /Users/neto/Documents/Napple/nxnative/ (read-only, commit 1a93086, MIT, NX:LICENSE, Copyright 2026 seedds). Unprefixed paths are in Napple/ of this repo at main 9ac63a8. ACR: / PERF: / MK8: = other batch branches read through git show origin/<branch>.

Status words: read (I read both sides), inferred (reasoned from one side, not confirmed on the other), verify (needs a run before acting).

0. Context that decides how to read everything¶

Same lineage, different HLE level. nxnative and Napple descend from the same idea (MacNativeGames / d2r-mobile: run the game's A64 natively, rewrite what macOS forbids, NVN on Metal). They differ below it. nxnative hooks SDK functions by symbol (NX:dm/host/hle_*.c, HLE_FORCE(...), e.g. NX:dm/host/hle_sf.c:33) and runs the SDK's own audio renderer and movie decoder on the CPU. Napple answers IPC as the real services would (docs/runtime-services.md, runtime/src/services/). So nxnative's service and audio fixes are findings about what games do, not transplantable code. Their GPU layer is a closer analogue of gfx/, but their shaders are translated offline by Ryujinx's translator (SPIR-V, SPIRV-Cross, metallib; NX:dm/gfx/tools/dmsh_build.py, NX:dm/third_party/ryujinx/), ours by an own Maxwell-to-MSL translator (gfx/src/shader/maxwell_translate.c, 2044 lines) at run time. Shader findings therefore map by behaviour, not by patch.
Nothing from nxnative is in Napple's pipeline yet. git grep -il nxnative on main and on the three batch branches only hits MacNativeGames/ (a separate snapshot: MacNativeGames/README.md:258, docs/how-it-works.md:84: its NSP common-ticket import follows nxnative's approach) and my own orchestration file. git log --all --grep=nxnative finds only my orchestrate commit. No integrator file or change note cites it. The batch sessions found several of the same bugs independently (marked overlap below). So "already in progress" below means same fix landing from another route, never "being ingested from nxnative".
MK8D is two different programs. Evidence from our own logs (session notes, ios-sim-mk8d-base-a32.log vs ios-sim-mk8d-latest-update.log): base v0 (b5c39b5b…) is AArch32 (a32: loader lines), but the Testing Material UPDATE v1376256 (fe941ed5…) loads as ordinary 64-bit modules (loader: mapped main … relocs=149788, no a32). nxnative has no AArch32 support (no hit for aarch32/armv7/elf32/thumb in NX:dm/tools, NX:dm/host) and only ever runs the merged 4.0.0 update (NX:dm/games/0100152000022000/game.tsv build v1441792 4.0.0 main=2c336a9b…, "only the merged 4.0.0 copy is extracted"). Consequence: base+update+DLC MK8D needs no A32 recompiler; the A32 work only matters for a base-only build. The update path's first blocker in our runtime is FAIL: IPC command time:u:300 not implemented (last line of the update log; see S1).

(a) Summary and ranked top 15¶

Short answers to the five questions:

ACNH black frame. nxnative's AC (3.0.3, same build line) is black unless the profile key render_cond on is set: AC wraps clears in SetRenderEnableConditional, and "four clear-only passes wiped the composite before the window copy" (NX:dm/games/01006F8002326000/README.md, "What it needs"). We record the condition and then draw as if always enabled (gfx/src/api_cmd.c:435-437, gfx/src/mtl_replay.m:1557). Our timestamp reports already write counter 0 plus time (gfx/src/mtl_replay.m:225-229), which is the semantic their evaluation relies on. This is the first thing to try. It may not be the whole story: ac-render's symptom is a culled final quad (ACR change note, "Next blocker"), whose cause nxnative does not explain (upper-left origin, winding code is equal on both sides: NX:dm/gfx/nvn/nvn_mtl.m:7366 vs gfx/src/mtl_replay.m:747-748). Second suspect from their docs: the quad's data comes from an SSBO and a UBO (driver table words c0), see G6/G12.
SMO visual bugs. nxnative has no note on exploded skinned geometry or the cap brim. Their SMO fixes that touch us: c0+0x610 largest point size (fog; G6), position invariance (we already have it, gfx/src/mtl_shader.m:121), nvnTextureCompare channel-mapping fix (HUD). Our integer-vertex-input handling is ahead (attr kind keyed per runtime format: gfx/src/shader/maxwell_translate.c:245,1900), so the repacked R8_UINT bone-index stream in PERF's finding (fa3fa316…) is not explained by their "integer vertex inputs" entry. Best use of nxnative here is as a reference translator oracle (E1).
First-use stutter. They avoid it with (i) shaders compiled offline into metallibs, (ii) a shipped/recorded prewarm manifest of pipeline keys (pso_manifest.tsv, pso_user.tsv: NX:docs/development.md:281-293), (iii) Metal 4 base pipelines specialized in microseconds instead of 3-5 ms builds, with an on-disk archive (NX:docs/development.md:296-308), (iv) a look-ahead that compiles the next pipelines of the submission in the background (pso_predict_rest, NX:docs/reference/shaders.md "The precompile look-ahead restores the scissor"), (v) normalized keys so duplicates collapse (NX:dm/gfx/nvn/nvn_pso_skey.h:22). We already persist keys, prewarm and use a binary archive (gfx/src/mtl_shader.m:1-60); missing are (ii) a seed list for a fresh install, (iv), (iii) and (v).
AArch32. Not handled by nxnative; see section 0.3.
Where they are more correct / where we are ahead: see (c) and the "ahead" table in (d).

Top 15, "take this first" (ranked by benefit to the four titles per effort)¶

#	Item (row id)	Titles	Effort
1	Evaluate conditional rendering, key `render_cond` (G1)	ACNH, MK8D, TOTK	S-M
2	`time:u` command 300 and the rest of the MK8D 64-bit boot path; decide base-only A32 is optional (S1, section 0.3)	MK8D	S
3	Graphics driver-segment words `c0+0x610` (max point size 2047), `+0x644` (base instance) (G6)	SMO, instanced draws everywhere	S
4	Per-pass dump diagnostics (`PASSDUMP`/`TEXDUMP`/`HILITE`/`DEPTHOVR` style) to pin the AC culled quad and SMO skin bug (T4)	ACNH, SMO	M
5	Vertex outputs start at (0,0,0,1), not only `pos` (G7)	all	S
6	Same colour texture at two slots: leave the duplicate unattached (G3)	agl titles (SMO, AC, MK8D use agl G-buffers)	S
7	Attachment larger than the pass: clear in its own pass first (G2)	all	S
8	Compressed render target through an uncompressed shadow (G4; queued as backlog "copy-back for compressed RTs")	MK8D	M
9	`SignalToAddress` type 2 with count <= 0 must write value-2 (K1, our bug)	all	S
10	Seed prewarm manifest + key normalization + background look-ahead (G11)	all	M-L
11	Cross-check rewriter jump-table shapes and exclusive-window cases against their list (R1, R2)	ACNH (63 MB text), updates	M
12	Present path hardening: commit residency before present, `QueueFinish` does not wait for display, present rescue (G10)	windowed Mac, iOS	M
13	MK8D service stand-ins: six default Miis with valid CharInfo, MiiModel system-data mount, device vs account save as two saves (S2, S4)	MK8D	M
14	Symptom-indexed reference docs and per-subsystem entries (T1)	process	S-M
15	CI workflow for host + CPU graphics + tool tests + iOS compile smoke (T2)	all	M

(b) Full table¶

Licence codes: NX nxnative's own MIT; D2R d2r-mobile MIT (per NX:README.md Acknowledgements); RYU Ryujinx MIT (NX:dm/third_party/ryujinx/LICENSE.txt); DEKO deko3d zlib (ported in NX:dm/gfx/nvn/nvn_tex.c:1-8); ATM Atmosphere GPLv2, YUZU GPLv2: reference only, never copy; MNG MacNativeGames MIT (NX:dm/third_party/macnativegames/LICENSE); LH LibHac (version 0.19.0 in NX:dm/tools/nx_extract/nx_extract.csproj:10; licence not stated in nxnative's tree, believed MIT upstream, verify before any reuse); SPV/GLS SPIRV-Cross, glslang (pinned by tool fingerprint NX:dm/gfx/tools/dmsh_chain.lock, not vendored, licences not verified here). Rule for every row: take the finding and design, write our own code, cite the nxnative doc in the commit. If any nxnative text or code is ever copied, add it to THIRD_PARTY_NOTICES.md (today it lists Ryujinx.Audio, MacNativeGames, LASRecomp, libopus, libvpx; no nxnative).

Owners: ac-render owns runtime/src/services, kernel waits in runtime/src, gfx/src/api_*.c, gfx/src/tex/; mk8d-recompiler owns runtime/src/a32jit, crates/napple-rewrite A32; ios-sim owns frontend/src/ios, runtime/src/ios_*, tools/build_ios*; backlog smo-gfx owns gfx/src/mtl_*.m, gfx/src/shader/. Proposed owners below say which new session, so that none touches another's files; rows whose files an active session owns say "after merges".

GPU: passes, clears, state¶

Id	nxnative (file)	Ours (file)	Benefit	Eff / risk	Lic	In progress	Owner
G1	Conditional render evaluated at replay: `NX:docs/reference/rendering.md:551-560`; code `NX:dm/gfx/nvn/nvn_mtl.m:539-575` (counter at rec vs rec+16, timestamp report = counter 0, samples-passed condition keeps drawing), `:1299`, skips at `:8228,8245`; key in `NX:dm/games/01006F8002326000/game.tsv`, `0100152000022000/game.tsv`, TOTK	recorded `gfx/src/api_cmd.c:435-437`; ignored `gfx/src/mtl_replay.m:1557`; timestamp reports right `:225-229`	ACNH black frames, MK8D race clears, TOTK	S-M / SMO: their note says evaluating hurt SMO (`NX:dm/games/0100000000010000/README.md`, "Rendering"), so gate per profile or only evaluate deterministic timestamp conditions	NX (design only)	no	new `gfx-cond` session after smo-gfx (same files `mtl_replay.m`, `api_cmd.c`)
G2	Clear of attachment larger than pass: own pass first: `NX:docs/reference/rendering.md:206-213`, `rp_begin_ex_` `NX:dm/gfx/nvn/nvn_mtl.m:2707`	load-action clears folded when `clear_covers_target()` `gfx/src/mtl_replay.m:328-350,504` (size mismatch handling verify)	stale depth/colour outside small passes (agl small targets)	S / low	NX	no	smo-gfx
G3	Same texture at two colour slots: unattach duplicate: `NX:docs/reference/rendering.md:185-191`, `rt_dup_slot` `NX:dm/gfx/nvn/nvn_mtl.m:2673,2725`	no equivalent found (grep for dup/same texture in `mtl_replay.m`: none)	G-buffer corruption on TBDR (scattered squares); agl titles bind such slots	S / low	NX	no	smo-gfx
G4	Compressed RT via shadow: `NX:docs/reference/rendering.md:170-177`, `bcrt_attach/bcrt_flush` `NX:dm/gfx/nvn/nvn_mtl.m:2571-2603`; depth-via-colour shadow `shrt_block` `:2603`	backlog item "copy-back for compressed render targets (MK8D)" `orchestrations/backlog.md`	MK8D	M / medium	NX	queued, not started	smo-gfx
G5	Layered passes for layer-writing VS: `NX:docs/reference/rendering.md:199-205`	no `renderTargetArrayLength`/layered path found in `gfx/src/mtl_replay.m`	LUT/cube passes (not seen on our 4 titles)	M / low	NX	no	backlog
G6	Driver segment `c0+0x610` = 2047 (max point size, SMO `alPointSpriteShader`), `+0x644` base instance: `NX:docs/reference/shaders.md:505-519`, `NX:dm/gfx/nvn/nvn_mtl.m:5234-5238`, `:4990`; layout table `NX:docs/agents/04-graphics.md` section 3	graphics `c0` carries only SSBO tables at 0x110/0x510 `gfx/src/mtl_replay.m:790-796`; compute words 0x200/0x210/0x310 are present `:801-806` (same layout as theirs)	SMO Cap fog never parts; any shader using `gl_BaseInstance`	S / low	NX	no	smo-gfx
G7	Vertex outputs init (0,0,0,1): `NX:docs/reference/shaders.md:323-329`, `vs_init` `NX:dm/gfx/tools/dmsh_build.py:580`	only `out.pos` set, varyings zero: `gfx/src/shader/maxwell_translate.c:1956`	LAS sun shadows were wrong without it; unwritten w=0 divides	S / low	NX	no	smo-gfx
G8	(diagnostics are listed as T4)
G9	Lower-left origin by negative viewport + reversed winding + flip at present `NX:dm/gfx/nvn/nvn_mtl.m:3529,7349,7366,8391`	rows stored bottom-up, flipped at present `gfx/src/internal.h:40`, `mtl_replay.m:370,483,723-759`, `mtl_present.m:363`; LAS title text reads correctly in `ios-sim-las-title.png` (session notes)	none now	n/a		equivalent, keep	none
G10	Present hardening: residency committed before present `NX:docs/reference/rendering.md:279-286`; `QueueFinish` waits GPU only `NX:docs/reference/nvn-objects.md:165-172`; present rescue `NX:docs/reference/frame-rate.md:50-57`; present snapshots `:70-82`; dynamic interval holds `:59-69`	`gfx/src/mtl_present.m` (not read in full: verify each)	windowed Mac, iPhone smoothness	M / low	NX / D2R	no	`gfx-present`
G11	Pipelines: Metal 4 base + specialization + archive files `NX:docs/development.md:296-308`; seed manifest + user manifest `:281-293`; look-ahead `NX:docs/reference/shaders.md` "precompile look-ahead"; key normalization `NX:dm/gfx/nvn/nvn_pso_skey.h:22`	persisted keys, background prewarm, binary archive, wait-or-build `gfx/src/mtl_shader.m:1-60,610`; archive off on iOS Simulator `:52`	first-visit stutter (`mtl_pipe_get`), cold-start	M-L / medium (Metal 4 path is Mac-only in theirs; iOS stays off)	NX	no	smo-gfx (same file)
G12	Inline uniform updates replayed as ordered GPU copies, barriers inside encoders, alias visibility: `NX:docs/reference/rendering.md:459-482`	CPU `memcpy` at replay `gfx/src/mtl_replay.m:1577`, ring copy per draw `:830-` (ordered on the CPU timeline; wrong only if the buffer is also GPU-written earlier)	AC final pass reads UBO+SSBO; verify with a pass dump	M / low	NX/MNG	no	`gfx-cond`
G13	`fp` math: `precise::` min/max/clamp, iOS relaxed math: `NX:docs/reference/shaders.md:338-350`	our FMNMX uses `fmin/fmax` `maxwell_translate.c:1057-1059`, `fastMathEnabled = NO` `mtl_shader.m:122`, no iOS math-mode flag verify	NaN lighting (white half on Mario in Galaxy 2)	S / low	NX	no	smo-gfx / ios-sim
G14	`SR_Y_DIRECTION`: open on their side (`NX:docs/reference/shaders.md:385`, Ryujinx returns +1)	we return -1 always `maxwell_translate.c:1423`	LAS-like shaders under upper-left origin	S, verify		no	smo-gfx
G15	Geometry shaders as compute + capture: `NX:docs/geometry-metal-research.md`, `NX:dm/gfx/tools/dmsh_geometry.py`, `NX:dm/gfx/tests/t_geometry.m` (498 checks); tessellation not drawn (`NX:docs/reference/shaders.md:464`); ACNH keyboard has 8 GS, SMO 3 (thunder), MK8D 1: none translated there either	none; backlog `gfx-features` (L)	low for the 4 titles	L / medium; their path rests on Ryujinx's GS lowering, ours has none: design only	RYU (design)	backlog	later
G16	Image atomics contract `NX:docs/image-atomic-binding-contract.md`, bounded MSL 2.3 emulation `NX:docs/reference/shaders.md:615-676`	the brief says we have `docs/image-atomic-binding-contract.md` and `docs/geometry-metal-research.md`: neither file exists in this checkout or on any branch (`git log --all --name-only`); backlog `shader-surface`	SULD/SUST titles (not our 4)	L	NX	backlog	later
G17	Texture details: new textures zero `NX:docs/reference/rendering.md:15`, cube array = layer-faces `:65` (overlap: ACR commit `22752de`), block-height truncation `:108`, sparse `:117`, 1D as 2D `:51`, buffer-texture row wrap `:93`, stencil carried in depth-stencil copies `:542`, `copy_t2t_draw` for RG11B10F to RGBA16F `:529`	`gfx/src/tex/napple_layout.c` (same packaged-layout byte as theirs `:4`), sparse getters are stubs `gfx/src/api_texture.c:72,147`	per-title bugs not yet seen	S each	NX/DEKO (layout origin: both cite deko3d zlib `NX:dm/gfx/nvn/nvn_tex.c:1-8`; check our file header)	partly	`gfx-tex` after ac-render
G18	Depth16 clear rounding `NX:dm/gfx/nvn/nvn_depth_clear.h:8`, partial colour/depth clears as triangles `NX:dm/gfx/nvn/nvn_mtl.m:8105`	clear quads `gfx/src/mtl_replay.m:512-606`; index != 0 falls back to a pass restart `:547`	partial clears on MRT	S-M	NX/MNG	no	smo-gfx
G19	Same-pass feedback via snapshot for listed shaders `NX:docs/reference/rendering.md:260-270`, `fb_prepare` `NX:dm/gfx/nvn/nvn_mtl.m:4939`	none	magenta 16x16 blocks on A-series GPUs (iPhone)	M / perf cost, per-profile list	NX	no	ios later

Kernel and threads¶

Id	nxnative	Ours	Benefit	Eff / risk	Lic	In progress	Owner
K1	`SignalToAddress` type 2, count <= 0 with waiters: value-2 (`NX:docs/reference/threads.md:61-66`, mesosphere `KAddressArbiter`; Ryujinx and yuzu agree)	writes value-1: `runtime/src/threads.c:566-570` (`cnt <= 0 ... val - 1`); count>0 branch matches	condition variables / mutexes in every title	S / low	ATM (reference only; implement from the documented semantics)	no	after ac-render (owns kernel waits)
K2	`WaitForAddress` type 3 (64-bit wait-if-equal, SDK 19+) `NX:docs/reference/threads.md:67-73`	types 0-2 only, type 3 returns invalid enum `threads.c:546-552`	none of our 4	S	ATM	no	later
K3	Same-core handoff when a higher-priority thread wakes/starts on the caller's core: `NX:docs/reference/threads.md:7-14,43-51`, `dm_yield_to_core_peers` in `NX:dm/host/hle_thread.c`	we model three virtual cores explicitly (`runtime/src/threads.c:16,253`, `NAPPLE_FREE_CORES`); may already give this: verify with an equivalent test (`NX:dm/host/tests/t_preempt*`)	removes racy guest crashes (LAS physics, ACNH loaders)	M / medium	ATM semantics	no	after ac-render
K4	Yields that make progress return at once (`thr_yield` `NX:dm/host/hle_thread.c:660`, `NX:docs/reference/threads.md:101-109`): ACNH shows a black plane-icon screen for 77 s without it	our yield hands the core to an equal-priority waiter, else `sched_yield` `runtime/src/threads.c:367-373`; no back-off to remove	ACNH loading time: probably not affected (no back-off), verify once it renders	S	NX	no	none
K5	`sleep_precise` (3/4 of remaining per step) `NX:docs/reference/threads.md:86-100`	our SleepThread implementation not read	30 Hz title pacing (27 Hz measured on their side)	S, verify	NX	no	later
K6	`svcGetInfo` FreeThreadCount = 96 minus live `NX:docs/reference/threads.md:28-35`	constant 0x100 `runtime/src/kernel.c:208`	Unity only	S	ATM	no	none
K7	Thread pause + context for GCs `NX:docs/reference/threads.md:36-42`	implemented `runtime/src/threads.c:414,449`; `docs/runtime-services.md` still says "existence check only" (doc stale)	none	docs only
K8	`SetThreadCoreMask`: ideal core moved into the new mask `NX:docs/reference/threads.md:15-20`	returns InvalidCombination if ideal not in mask, ideal -3 checks current `threads.c:390-405`	cannot tell which matches Horizon; Atmosphere reference needed	S, open question	ATM	no	see (e)

Services, input, storage, audio, movies (findings, not code)¶

Id	nxnative	Ours	Benefit	Eff / risk	Lic	In progress	Owner
S1	(their `time::` hooks, `NX:docs/reference/services.md:508-535`: steady clock, snapshots, user clocks)	`time:u` table ends at cmds 0-5,20,100-102,200,400,401,500,501 `runtime/src/services/time.c:185-199`: no 300 (CalculateMonotonicSystemClockBaseTimePoint per switchbrew; confirm)	MK8D update boots until here (ios-sim log)	S	none	no (found in our logs, not from nxnative)	ac-render's area (services): after merge, or tiny `mk8d-boot` session
S2	MK8D Mii: six default Miis with valid CharInfo, random for `BuildRandom`, Mii editor cancelled; MiiModel system data `0100000000000802` mounted as `MiiResource:` with a stand-in archive (layout in `NX:dm/games/0100152000022000/README.md` "Miis"; `NX:dm/host/hle_mii.c:137`)	`runtime/src/services/mii_core.c:21-42` has default Miis; system-data mount: only `OpenDataStorageByDataId` for DLC `fs.c:553-572` (verify MiiModel)	MK8D crashes at boot without both	M / low	NX	no	MK8D follow-up
S3	Device and account save as two saves when a game mounts both `NX:docs/reference/services.md:163-169`, `NX:dm/host/hle_fs.c:81`; MK8D deletes `CommonSave:/userdata.dat` at boot	save types are separate directories by type `docs/runtime-services.md` "Save data"; verify MK8D	MK8D re-asks Mii every boot if shared	S	NX	no	MK8D follow-up
S4	AOC mount cache size = RomFS working memory `NX:docs/reference/services.md:186-192`, `romfs_work_size` `NX:dm/host/hle_fs.c:347`	arises from SDK-level hooks; at IPC level likely N/A	MK8D DLC	S, verify	NX	no	MK8D follow-up
S5	Game's own HIPC clients (`olsc:u`, `npns`) answered `NX:docs/reference/services.md:50-56`, `NX:dm/host/hle_sf.c:33` (ACNH aborts without)	`olsc` exists (`services/olsc.c`), `npns` not in `runtime/src/services/`; manager enforces the NPDM access list	ACNH after first frames	S, verify with `NAPPLE_IPC_LENIENT=1`	NX	no	ac-render
S6	History reads fill all 16 entries with consecutive numbers (SMO sticks dead otherwise) `NX:docs/reference/services.md:200-209`; six-axis sensors per controller and shake `:217-224`	`input/src/hid_shm.c:148` count capped to 16; outer=2n, inner=n fix (overlap: ACR); six-axis rings exist `:44,87`; shake verify	SMO sticks/motion, cap throw	S	NX	partly	ac-render / input
S7	Audio renderer details (SDK 4/6/8 DeviceSink ids, update lock) `NX:docs/reference/audio.md:38-100`	N/A at IPC level (`audio/`, `services/audio.c`)	none	none
S8	Movies: 16 video output slots `NX:docs/reference/movies.md:109-115`; VP9 via libvpx, Vorbis via stb_vorbis, H.264 via VideoToolbox; vendored in `NX:dm/third_party/` (BSD-3 / MIT-or-PD)	ours uses `runtime/src/movie.c`, `vt_h264.c`, Homebrew libvpx (`THIRD_PARTY_NOTICES.md`); ios-sim added libvpx for the Simulator (commit `5fb4673`)	ACNH opening movie (VP9 1080p + Vorbis)	M	NX/third-party	ios-sim (libvpx)	ios-sim
S9	Frame pacing / present interval handling `NX:docs/reference/frame-rate.md`	`runtime/src/timescale.c` etc. not compared	smoothness	M			later

Loader, rewriter, ingest, memory¶

Id	nxnative	Ours	Benefit	Eff / risk	Lic	In progress	Owner
R1	Jump-table shapes (7+): `NX:docs/reference/rewriter.md:23-100`; halfword tables padded to the largest entry, patch area 42 MB to 7.9 MB for ACNH `:134-143`, ACNH note `NX:dm/games/01006F8002326000/README.md` "Rewriter"	`crates/napple-rewrite/src/jumptab.rs:18-24` (3 shapes + function-span fallback), `rewrite.rs:198` "landing pads too far"; ACNH data builds today	ACNH-size modules (63 MB text), game updates	M / low	NX	no	`rewrite-xcheck` (disjoint from mk8d-recompiler's A32)
R2	Exclusive windows: exit order, out-of-line retry blocks, adrp in window `NX:docs/reference/rewriter.md:103-132`	`crates/napple-rewrite/src/windows.rs`, `check.rs`	atomics correctness (job systems hang)	M, mostly tests	NX	no	`rewrite-xcheck`
R3	Spin-wait rewrite for power `NX:docs/development.md:310-320`; `NX:dm/tools/patch_nso.py`	none	battery/thermals (iOS)	L / medium	NX	no	ios later
R4	Per-build workarounds keyed by build id, `MATCH_CUTS` function fingerprints `NX:docs/reference/rewriter.md:7-22`, `NX:dm/tools/nso_match.py`	per-build signatures `knownSignatures/`	updates keep a workaround	M	NX	no	later
R5	Extraction: LibHac `nx_extract` `NX:dm/tools/nx_extract/Program.cs` (231 lines)	own Rust: `crates/napple-core`, `napple-build/src/bktr.rs`, NCZ support; nothing to take (ours is ahead: NSZ, no .NET)	none	none	LH
R6	Survey + provisional profile for unknown titles (`NX:dm/games/profile_guess.py`, `NX:scripts/build_ipa_from_nsp.py`, `NX:docs/agents/02-survey.md`)	Studio/CLI `napple build`, no survey step read-only check	generality, coverage	M-L	NX	no	later
R7	Guest memory layout (4 GB layout, `sys_resource`, heap from pool minus used) `NX:docs/reference/memory.md:7-50`, ACNH values in its game.tsv (`mem_reserve 0x100000000`, `app_pool 0xCD500000`, `mem_used 0x7a00000`)	`runtime/src/memory.c` layout; iOS budget is backlog `ios-memory`	iOS budget	M	NX	backlog	ios-memory

Tooling, tests, docs¶

Id	nxnative	Ours	Benefit	Eff / risk	Lic	In progress	Owner
T1	Symptom index + entries "Symptom, Cause, Fix, Found on" per subsystem; `NX:docs/agents/symptoms.md`, `NX:docs/reference/*.md`, `NX:dm/tools/docs_symptoms.py`, `NX:dm/tools/tests/test_docs.py` (links, anchors, length limits)	`docs/changes/*` dated notes; no symptom index	faster bring-up of the next title	S-M	NX	no	docs session
T2	CI `NX:.github/workflows/tests.yml` (host + CPU graphics tests, TSAN targets for audio/memory, iOS compile smoke `NX:dm/ios/tools/build_smoke.py`)	`.github/workflows/` has pages.yml and release.yml only	regression safety	M	NX	no	`ci` session
T3	~40 GPU regression tests `NX:dm/gfx/tests/t_*.m` (clear_ds, clear_tex, depth16, feedback, present_snap, pso_archive, geometry, copy_alias ...)	`make -C gfx test`	test ideas for G1-G4, G18	S each	NX	no	with each gfx row
T4	Diagnostics: `DM_NVN_PASSDUMP`, `TEXDUMP`, `CBDUMP`, `HILITE`, `DEPTHOVR`, `KEYDUMP`, `MAGSCAN`, `GPUPROF` `NX:docs/agents/04-graphics.md` section 7	`NAPPLE_DRAWLOG*`, `NAPPLE_SHADER_PROBE`, `NAPPLE_DEBUG_SOLID_FS`, `NAPPLE_DEBUG_SKIP_VS` (`gfx/src/mtl_shader.m:470-500`, PERF note)	pass-level dump with clears and per-pass render-condition state would have shown the AC final pass	M	NX (design)	no	smo-gfx
T5	iOS build and launcher settings table `NX:dm/ios/tools/*`, `NX:dm/games/settings.tsv` model, `ios_metal.py` relaxed math	`tools/build_ios*.sh`, `docs/ios.md`		M	NX	ios-sim	ios-sim
T6	Per-title notes with "what it needs / open" (`NX:dm/games/<id>/README.md`, 65 titles; ours under `compat/`, `knownSignatures/`)			S

(c) Discrepancies: our behaviour contradicts theirs (candidate bugs on our side)¶

#	Ours	Theirs	Evidence	Confidence
c1	Conditional render always drawn, warning "treated as always enabled"	Evaluated; AC and MK8D require it, SMO wants it off	`gfx/src/mtl_replay.m:1557` vs `NX:dm/gfx/nvn/nvn_mtl.m:539-575`, AC/MK8D `game.tsv`	high for AC/MK8D semantics; whether it explains our culled quad: unknown
c2	`SignalToAddress` type 2, count <= 0, waiters: value-1	value-2	`runtime/src/threads.c:568` vs `NX:docs/reference/threads.md:61-66`	high (documented kernel behaviour)
c3	Graphics `c0` words 0x610 and 0x644 left 0	2047 and base instance	`gfx/src/mtl_replay.m:790-796` vs `NX:dm/gfx/nvn/nvn_mtl.m:5234-5238`	high
c4	Varyings start at 0	(0,0,0,1)	`gfx/src/shader/maxwell_translate.c:1956` vs `NX:dm/gfx/tools/dmsh_build.py:580`	high
c5	`time:u` has no cmd 300	their host answers the SDK `time::` calls	`runtime/src/services/time.c:185-199`, ios-sim update log	high
c6	`docs/runtime-services.md` says SetThreadActivity is an "existence check only"	implemented	doc vs `runtime/src/threads.c:414`	docs stale
c7	Window origin lower-left: our targets stored bottom-up, flipped at present	viewport flipped, rows top-down	`gfx/src/internal.h:40` vs `NX:dm/gfx/nvn/nvn_mtl.m:3529`	both work for LAS; keep, watch MIRROR_Y copies
c8	`SR_Y_DIRECTION` = -1	their open question (Ryujinx +1)	`maxwell_translate.c:1423` vs `NX:docs/reference/shaders.md:385`	unknown
c9	No unattach of duplicate colour slot	unattached	none vs `NX:dm/gfx/nvn/nvn_mtl.m:2673`	medium

(d) Do NOT take, and where we are ahead¶

Do not take: - Any Atmosphere / yuzu code: GPLv2, reference only (K1-K3). Implement from documented semantics. - Ryujinx translator and dotnet toolchain as a runtime dependency: wrong architecture (offline SPIR-V chain vs our run-time translator), .NET at build, large per-game library sets (up to 344 MB on iOS for MH Rise per their note). Use it only as an offline oracle (E1). - SDK-symbol HLE (HLE_FORCE, in-process audio renderer, movie decoder hooks, hle_sf.c): we answer IPC; copying hooks would fork our model. - Per-game profile TSV system and per-build rewriter cuts as a first move: our rule is data-only per-game knowledge via knownSignatures; adopt only the findings. - LibHac nx_extract: ours is native Rust and ahead (NSZ, no .NET). - iOS-only items now (power QoS, spin-wait rewrite, present snapshots): out of scope until iOS device work (backlog says iOS out of scope). - Geometry/tessellation implementation: depends on the Ryujinx GS lowering; none of the 4 titles needs it to reach gameplay (SMO thunder only). - Their vendored copies: keep our own provenance. Anything adapted needs a THIRD_PARTY_NOTICES.md entry.

Where we are ahead (nobody copy backwards): - Vertex integer inputs keyed per runtime attribute format, no offline guess and no shadow libraries (gfx/src/shader/maxwell_translate.c:245,1900); stride-not-multiple-of-4 repack, fans/quads/8-bit indices exist (mtl_replay.m:721,1050,1108,1179). - Depth mode and origin resolved per draw in the pipeline key, no second entry point per vertex library (mtl_replay.m:724, mtl_shader.m:463). - Compute c0 dispatch size and UBO/SSBO tables already match their layout (mtl_replay.m:801-806), as does preserveInvariance (mtl_shader.m:121). - Own AArch32 route in flight (A32 interpreter + lifter; they have none). Own NSZ/BKTR ingest in Rust. Transactional save data with crash-safe commit (docs/runtime-services.md "Save data"). SetThreadActivity/GetThreadContext3 implemented. - IPC-level services with an NPDM access list: SDK-version differences do not matter, which is why their SDK 4/6/8/20 audio-command fixes do not apply.

(e) Open questions¶

Does evaluating render_cond fix the ACNH black image, or only the culled-quad bug remains? Run ac-render's build with NAPPLE_DRAWLOG and a pass dump of the last pass (conditional state, clears) before changing code. AC (their note) needs it by frame 900; ours draws ~1200 frames in 45 s.
Why is our final-pass quad culled when upper-left winding code is equal? Candidates: vertex data from the SSBO (c0 table, gpu_va), UBO projection, front-face sense after SR_Y_DIRECTION, depth_mode clip. Needs a CPU evaluation of one vertex.
SMO skinned geometry: is our translation of fa3fa316 numerically equal to Ryujinx's? Proposed oracle: build NX:dm/third_party/ryujinx offline (MIT) on the same shader bytes, compare outputs for random inputs. Is that within what the user wants (their repo, Do not run builds of nxnative applied to this review only)?
MK8D: confirm with a run that update v1376256 + base + DLC reaches the title through the 64-bit path after time:u:300; if so, deprioritise the A32 lifter for the user's goal. (All MK8D evidence here comes from logs, not from a run of mine.)
Brief says we have docs/image-atomic-binding-contract.md and docs/geometry-metal-research.md; they are not in this checkout or any branch. Are they on an unmerged local branch?
SetThreadCoreMask when ideal core -3 and the current ideal is outside the new mask: Horizon result vs ours (K8)? Needs the Atmosphere source, not read here.
LibHac / SPIRV-Cross / glslang licences are not in nxnative's tree; confirm upstream before any reuse.

Method and limits¶

Read in full: NX:dm/games/{01006F8002326000,0100000000010000,01006BB00C6F0000,0100152000022000}/README.md, NX:AGENTS.md, NX:docs/agents/04-graphics.md, NX:docs/reference/{threads,rendering(137-290,329-560),shaders(1-20,301-520,697-741),rewriter(1-135)}.md parts, NX:docs/development.md (52-73, 281-334), NX:dm/third_party/README.md, game.tsv of MK8D. Headings only: remaining reference docs. Spot-read code: NX:dm/gfx/nvn/nvn_mtl.m at the cited lines. Ours: cited files at cited lines only; large files (mtl_present.m, tex/, a32jit) not audited. No builds, no game runs. Sibling notes used: ac-render-progress.md, ac-303-stall-findings.md, smo-title-screen-visual-bugs.md, mk8d-is-aarch32.md, ios-sim logs (all in the shared notes folder), ACR/MK8/PERF change notes via git show.