Twenty-four days, step by step

First message of the project invoked the deep-research workflow on libultraship and 3DS targeting. Verdict: New 3DS only; the Old 3DS 64 MB region cannot hold it. "new 3ds as target makes total sense".

Contract headers, stream directories, a stub build. The libultraship carve spike proved the runtime compiles on devkitARM with a 13-line newlib patch. Phase 1 streams ran in parallel worktrees and merged in order E→B→A→D→C on 08-12 21:45.

M1 integration: the full game compiled and linked on 08-13 01:33. Then the black screens: bzero, 64DD, sdmc: path mangling, double swap, the SETTIMG low-address guard, the DMA misroute. User at 05:37, 05:42, 12:44, 13:00, 13:16: "still black screen".

"oh! i saw some random shapes briefly for the menu or something. just a white screen with a rainbow rectangle at the bottom".

Race freeze (DMA misroute) at 19:26, then a race-time death exposed as an uncaught bad_alloc (EnqueueList's worst-case reserve), fixed by capping the resource cache. "crashed again when race tried to load", "says fatal error encountered".

"lets do 1, 2, and 3. for CI though, no github actions. this will all be local for now." SMDH metadata, makerom CIA packaging, the real-hardware runbook, a local ci-3ds.sh, the user guide. Two deep-research runs: 60 fps on New 3DS, and stereoscopic 3D. The stereo foundation and GPU profiling telemetry landed the same night.

A deep-research pass on Fast3D vs RT64 confirmed the renderer choice; an asset-cost shift (ETC1, decimation) was added to the plan and marked "done with opus, reconsider with fable".

Fleet on HUD garbage, shadow, boost, sky wedge, tunnel roof, km/h, building pop-in, CI. The measurement pass returned the CPU-bound verdict. Night session 1 while the user slept.

"sure, promote it. i dont see any large regressions. probably better".

User demanded new methods for the remaining visual bugs; deep-research run on the fixing methods. Night session 2: S7 memos, async frame mirror, vblank skip, the env-colour flush that fixed rival-coloured machine bodies.

Race-start hitch fixed (blob preload off the audio thread), gputrace made to reach hardware, the yellow-decal ship fixed by clamping shader-clamped tiles. Hardware: median 48–50 fps, p95 58.

"im showing this to my friend, and he is unimpressed by this port to 3DS. why is it impressive in laymans terms?" A short separate session answered it; the Frontier page is the long answer.

"performance is much better! 3D works great, doesnt seem to affect performance much at all. last race did have 30fps AND had a hard crash (first time ever!)". The crash was the filelog growing without bound; fixed with the retained ring, plus the upstream combiner-key leak found the same afternoon.

Touch menu v1 with STAT/DISP/3D/AUD/INP/LOG/DBG/ABT tabs, config write-back, live setters. 19:21: "playing perfectly, full bleed should be default display mode i think". Full-bleed made the default at 15:21 local.

Then Time Trial froze on its first menu: course-select state never reset in a static non-EK build (retail relied on overlay re-DMA). Fixed 16:18 local.

Rival detail option, TRILOOP packed VBO, APT power/HOME exit fix. 04:22: "minimal appears to be a 8 to 9 fps gain ! sometimes a little more. def worth it."

Suspend-path audit, audio park gate. Fixed 20:56 local. Combined test build staged with A/B inis and the first TEST-PLAN.txt.

"i think brfast helped with framerate … home menu came up! i was also able to return to game. but when i tried to 'close' the software it hung." Close-from-HOME root cause (sticky park event, hot spinner) fixed 10:18 local. Median 57.2 vs 56.0, p10 47.2 vs 42.0 with brfast on vs off.

The combined-lever HUD break (trifast without the atlas UV offset). Fixed at 14:30 local; the rule "combined-lever screenshot before staging" written. 18:59: "ok seems fine right now".

All four levers on with stereo on: median 56, p10 48, 32% at cap. Same-course A/B with trace on: steady 16.4 → 14.9 ms, crowd 20.2 → 18.7 ms. Merged to mainline.

Research 2: core 2 usable, core 3 not, no precedent for cross-core render pipelining. Agents F (dispatch census), G (boot audio 2), H (render thread on core 2). 22:11: "so curious if the multi core will be profitable!"

Sync mode, then pipe mode forked at osSpTaskStartGo, then ahead mode (DP-done acked when the game parks). The console filter fixed the bottom-screen tab bar the user reported at 00:05 ("debug doesnt have tabs anymore").

Hardware round 3 (pipe mode): median 49 → 56, p10 42 → 48 with the trace on; stable, HOME clean, but not yet overlapping. 01:46: "got it. can you work on 2, 3, and 4 right now in parallel?" → Tasks I (bridge on main), J (auto rival detail), E (machine atlas).

Round 4 card: ahead mode + bridge on main + auto LOD, full Grand Prix, trace off.

User captured side-by-side screenshots against the real game: machine-select ships 1.6× too large. Viewport fold fix, verified against the references. "sorry didnt mean to interrupt but its looking good".

"one thing thats really jarring is the 1P/2P/3P markers during a race. they sit on the most foreground in 3D but the ships are somewhere in the distance". Anchored at the labeled ship's depth via the game's own prim-depth register; verified in side-by-side stereo; staged as round 5 with the crash and viewport fixes.