Step by step: twenty-four days, then forty-three more, then one week

Burst 1 · Research to first race (08-11 → 08-14)

1. 08-11 03:16
   Deep research: is a 3DS port feasible?²⁴ Verdict: New 3DS only; the Old 3DS 64 MB region cannot hold it.Later (09-25 → 10-19): it fit under simulation on 09-25, booted on an Old 3DS on 09-29, and shipped for the Old 3DS in v1.0.0. — this page, chapter 3; RELEASE-NOTES.md v1.0.0
2. 08-11 03:44
   Plan for a team of agents "can we segment plan so we can have multiple work trees and multiple contributing agents ?" A memory budget, a five-stream plan and a reviewer agent.
3. 08-11 → 08-12
   Phase 0 and the carve spike libultraship compiles on devkitARM with a 13-line newlib patch; Phase 1 streams merged 08-12 21:45.
4. 08-13
   "start the crew", then black screens Full game linked 01:33; then bzero, 64DD, path mangling, double swap, the SETTIMG guard, a DMA misroute. 13:17: "i saw some random shapes briefly".
5. 08-14 01:01
   First race "omg it works ! i can race ! it looks messed up and runs at 12 to 20fps but it works!" (after a race-time bad_alloc fix).
6. 08-14 04:26
   Hardware prep, CI, docs CIA packaging, a hardware runbook, local CI; research on 60 fps and on stereo 3D.
7. 08-14 14:06
   Pause 1 "we ran out of fable credits until next week". Fog (pink road) still open.

Burst 2 · Playable, then hardware (08-20 → 08-22)

1. 08-20 02:55
   Resume; the road appears Fog solved (per-vertex blend, not the depth LUT); fog, GPU profiling, stereo, perf and culling merged.
2. 08-20 05:38
   "run as many subagents on plausible tasks"²³ A fleet on HUD, shadow, sky, tunnel, km/h, pop-in. The measurement pass: CPU-bound, not GPU-bound.Held (09-01 → 10-24): an emulator measurement at the time (GPU 0.4 ms vs CPU 26–58 ms); hardware agreed (GPU 5.6 vs build 16.3 ms), and the Old 3DS probe agreed on 10-24. — locked60-campaign.md; NATIVE-RENDERER.md ceiling_probe
3. 08-20 22:45
   The HUD bug After eight wrong fixes: the bridge's ILP32 high-bits backstop dropped 13 SETTIMG a frame. "omg finally!!"
4. 08-21 13:03
   "menus … still crazy slow at around 10fps" A per-opcode profiler named G_LOADBLOCK at 22.34 ms a frame; fixed within the hour. "blazing fast on the UI".
5. 08-21 15:54
   First hardware run "wow it runs amazing ! … sound doesn't work btw". Audio fixed by 21:22 after eleven SD trips (the ROM-side driver needed host porting).
6. 08-21 22:49
   "framerate at a crispy 60. then... maybe 3D?" Race-start hitch and decal fixes. Hardware median 48–50 fps, p95 58.
7. 08-22 01:31
   Pause 2 "ran out of tokens again". Checkpoint 854b9d6.

Burst 3 · Product grade (08-27 → 08-28)

1. 08-27 06:21
   Fleet drop Traffic grind (bridge 19.7 → 10.4 ms), sky wedge, CCMUX 11, off-axis stereo, shadow audit.
2. 08-27 16:00
   Hardware with stereo²² "3D works great, doesnt seem to affect performance much"; first hard crash: the filelog grew without bound, fixed with a retained ring.Later (09-02 → 10-02): the 256 KB cap truncated a session (raised to 1 MB, 09-02); the ring then caused 26–270 ms and 1.1–2.8 s SD-write hitches (09-30), fixed by an async, dropping writer. — 2344e88; difficulties.html
3. 08-27 16:53
   Touch menu and full bleed "the bottom screen should have a menu for these features"; full-bleed made default. 19:30: "i did the first GP!"
4. 08-28 00:18
   Rival detail, packed VBO, HOME exit²¹ "minimal appears to be a 8 to 9 fps gain !"Later (10-19): the New 3DS preset runs full rival detail; automatic scaling is used on the Old 3DS. — gdx3ds_preset_n60.c; RELEASE-NOTES.md v1.0.0
5. 08-28 04:27
   "moonshot! lets go for it"²⁰ GPU vertex transform in one session: works, imperceptible on hardware, breaks the venue floor. Shelved.Later (09-29): the broken floor was traced to float24 cancellation on world-space vertices (up to ±32,000); a machines-only transform was rebuilt for the Old 3DS and ships in both presets. — o3dsgpuxf-progress.md; gdx3ds_preset_n60.c
6. 08-28 13:17
   README and pause 3 README-3DS and the release package. Then the third pause.

Burst 4 · The last big lever (09-01 → 09-02)

1. 09-01 21:27
   "the goal to 60fps"¹⁹ Bridge output cache NO-GO (87% of commands are host-built each frame); brfast memos instead, bridge 11.5 → 4.9 ms. Research: nobody has shipped beyond the per-frame walk.Later (10-02): the native race renderer (10-01/02) skips building the game's display-list sections on rendered ticks and draws from GPU-resident meshes. — native.html
2. 09-01 23:53
   "the home button crashes the game!" Suspend-path audit and audio park gate; the first combined A/B test card.
3. 09-02 14:02
   Hardware round 2 HOME works; close-from-HOME hang fixed. brfast: median 57.2 vs 56.0, p10 47.2 vs 42.0.
4. 09-02 14:41
   LOCKED-60 campaign HUD atlas, TMEM bookkeeping, per-triangle cost, boot audio: four agents, landed by 13:40.
5. 09-02 18:35
   "VERY slow. massive step back" A combined-lever HUD break; fixed, and "screenshot the combined levers before staging" became a rule.
6. 09-02 21:56
   Rounds 3 and 4 Median 56, p10 48, 32% at the cap with stereo on; trace-on A/B steady 16.4 → 14.9 ms.
7. 09-02 21:59
   Research 2¹⁸ Core 2 usable, core 3 not, no precedent for cross-core render pipelining. Agents on a dispatch census, boot audio, a render thread.Later (10-19): the render thread shipped in v0.9.0 and was left out of the v1.0.0 presets (above). — gdx3ds_preset_n60.c

Burst 5 · Multi-core (09-02 night → 09-03)

1. 09-02 → 09-03
   Render thread, sync → pipe → ahead Forked at osSpTaskStartGo; ahead mode acks DP-done when the game parks. Pipe mode on hardware: median 49 → 56.
2. 09-03 13:20
   Round 5: median 59.6, p10 51.4, 53% at the cap¹⁷ Ahead mode, bridge on main, auto rival detail. Then a hard crash (texture-cache LRU on core 2), fixed by render-thread ownership.Later (10-19 → 10-26): v1.0.0's tuned presets set renderthread = 0: the native race path fits 16.7 ms on one core. Card 6: New 3DS 3D races at 60 in every window without it. — gdx3ds_preset_n60.c; o3ds-card6-analysis.md §4
3. 09-03 13:46 → 16:32
   Menu scale and stereo markers Machine-select 1.6× too large (viewport fold); 1P/2P markers anchored at the ship's depth.
4. 09-03 18:08
   This post-mortem "lets write up a whole post mortem on this project."
5. 09-04
   State of play (v0.9.0)¹⁶ Ahead mode, bridge-on-main, auto rival detail and the crash fixes all merged. The GPU vertex transform stayed shelved; the tunnel-roof bug (#27) was still open at v1.0.0.Later (09-29 → 10-19): a machines-only gpuxf = 1 (the floor stays on the CPU) is on in both consoles' auto presets in v1.0.0 and v1.1.0. — gdx3ds_preset_n60.c; o3dsgpuxf-progress.md

A feasibility pass on 09-07 found the fix should start with the development loop, not the renderer: no network path to the console, no frame oracle, no desktop reference build. Over two weeks that loop was rebuilt (a network control and log link, input-recorder determinism, a Linux reference render container), and the first renderer tier landed: an owned F3DEX2 interpreter replacing libultraship's, a native race path (track, machines, sky, HUD, effects), and the first 64DD Expansion Kit pass. In full on The road to 60, all over again.

1. 09-24
   The milestone reopens A stale note said the user owned no Old 3DS; they did. Under a simulated 64 MB grant: RDRAM arena halved to 8 MiB, audio_table streamed.
2. 09-25
   It fits, in simulation¹⁵ Replay 49/49 under memtype=0 with 2.65 MiB to spare.Later (10-16 → 10-26): a 10-16 hardware capture peaked at heapUsed ≈ 24.2 MB; card 6 logged a race watchdog at heapFree=3264. — STATUS.md; o3ds-card6-analysis.md §5.1
3. 09-28
   The console says memtype=0 Matches the simulation byte for byte; the shipped RDRAM-16 build dies in its boot preload.
4. 09-29 00:10
   "it ran ! at like 9fps" Wall p50 96 ms, GPU busy 3.6 ms. "is there anything we can do to avoid the display lists"
5. 09-29 day
   The parallel wave Audio-lock wait 25.8 → 0.2 ms, the HOME freeze (backlight session), bit-exact ARMv6 SIMD audio, the slow-motion discovery.
6. 09-29 13:09
   "we want this running at 60fps" A design sub-orchestrator and five research and probe subagents.
7. 09-29 15:16
   "playable but nausea inducing"¹⁴ Frame skipping: logic 57–60 ticks/s, drawn 19–28 fps. Draw cost measured at 10.9 µs per draw.Held (10-26): DrawProbe measured 10.9 µs flat over 30–300 draws (2,926 cycles). A uniform-free draw is 8.0 µs; a trimmed citro3d reached ~6.8 µs in the emulator, unmeasured on hardware; race draws now bypass it. — cpu-profile-answer.md §A.2
8. 09-29 19:39
   First release candidate does not boot¹³ Cold size-optimised code crashes, stock boots; code generation reverted. The latent bug was never found.Held (10-26): still unproven; code generation stays stock and hot-first ordering was never re-tested on hardware. — rel1crash-progress.md; agent-D-cpu-cuts.md
9. 09-30 01:10
   Second candidate boots, hitching¹² "like 15fps ... a lot of awkward hitching": synchronous SD log flushes, 26–270 ms on main.Later (10-02 → 10-19): a full ring later forced 1.1–2.8 s writes; fixed by an async writer, then a log that drops lines rather than block. v1.0.0's own figure still shows the gap (57.3 vs 51.9 fps). — difficulties.html; RELEASE-NOTES.md v1.0.0
10. 09-30 02:09
    "start as many agents as necessary" Seven agents hit the session limit together and were resumed; then the disk filled.
11. 09-30 → 10-01
    A hardware profiler¹¹ CPI 4.34, 43.6% fetch stalls; a probe found the cliff from CPI 1.06 at 16 KB of hot code to 4.1 at 24 KB.Held (10-07 → 10-08): native-renderer builds cut instructions ~3.6× and I-misses ~4×, but CPI rose to 5.4–6.4 with I-fetch stalls at 28–32%: less code, not faster code. — cpu-profile-answer.md §A.3
12. 10-01 08:57
    "we should go for a real 60!" Pose replay (40 images/s, artifacts) set aside against a research verdict of "marginal"; an I-cache-first native race renderer begins.
13. 10-01 22:41
    First native-renderer build freezes A verify mode rolled back the live GPU command queue; hardware only.
14. 10-02 09:28
    "time trial was silk 60fps" 199,472 draws verified, 0 bad; Grand Prix 45–50 with log hitches. Saves after the first never persisted (fixed).
15. 10-02 23:19
    "omg it ran amazing!" Card 3: race median 16.5 ms, 0 hitches. "lets focus on 3D at 60fps".
16. 10-03
    3D, first pass¹⁰ Black sky on an interim build, then "60fps for 2d for both", 3D in the 40s and 50s. The right eye is interleaved per chunk.Later (10-26): each chunk's two eyes became one queue entry (CMDBUF_JUMP): Old 3DS race entries 9 → 7, 3D menus 40.2 → 44.6 fps; v1.1.0 default. — ba0692e; a2905ed
17. 10-03 → 10-07
    Rate-limit pause One branch's uncommitted edits recovered on 10-07.
18. 10-07 22:51
    3D, second pass Mute City ~60 in 3D with end-mode pacing; heavier venues 16–18 ms.
19. 10-08 02:08
    "all 3 crashed" Bottom-screen framebuffer in VRAM, which a CIA maps read-only for the CPU; the emulator does not enforce it.
20. 10-08 02:56
    "did a long play on main, was good!"⁹ VRAM change reverted: 3D Grand Prix race windows at 59–60 fps; 0 bad draws.Later (10-24 → 10-26): a development card, not the shipped build. A 16-machine race later ran 53.7 fps; one submission brought the 1P race to 59.7–59.8. — cpu-profile-answer.md §A.4; o3ds-card6-analysis.md

1. 10-08
   Card 4 closes the 3D push⁸ 59–60 fps in 3D Grand Prix race windows on a development card; the rest of v1.0.0 followed. Commit- and document-sourced from here.Later (10-19 → 10-26): v1.0.0 shipped 3D as “not yet a locked 60 fps”; a heavier 16-machine 1P race measured 53.7 fps (10-24); card 6's baseline ran 56.4 / 59.3 and one submission 59.7–59.8 (10-26). — RELEASE-NOTES.md; cpu-profile-answer.md §A.4
2. 10-07 → 10-08
   Expansion Kit: own title and save The last blocker to a second, side-by-side install.
3. 10-08 → 10-15
   Rate-limit pause The longest gap of the project.
4. 10-15 → 10-17
   The bottom screen becomes DASH⁷ Lap, position, energy, splits, composited into the framebuffer at well under a millisecond of CPU time.Later (10-23): on Old 3DS hardware the DASH cost 4.2–5.6 ms a frame and was put on a 1 ms budget. — d9deb37; e99777b
5. 10-16
   The Old 3DS memory ceiling, in simulation⁶ One large mid-race asset load can exhaust the budget (the generic archive loader); a hardware capture peaked near 24.2 MB. Left as a known limit.Later (10-26): v1.1.0 moved ~320 KB off the main heap and reserved the track at boot; card 6 showed the failure mode is fragmentation. Tight memory stays a known issue. — d7010a8; a2905ed; RELEASE-NOTES.md v1.1.0
6. 10-16
   An audio-crackle hypothesis, refuted⁵ Vblank vs audio rate measured and ruled out; a smaller resample fix ships.Held (10-20): v1.1.0 turns the clamp and declick fixes on for the New 3DS, not yet confirmed to cure its crackle; Old 3DS residual crackle still open. — RELEASE-NOTES.md v1.1.0 Audio; 9ead784
7. 10-18
   A New 3DS froze at the title screen The attract demo's full-field race overflowed the GPU command buffer; see The road to 60, all over again.
8. 10-18 → 10-19
   Custom HUD mode shelved The three fixed DASH layouts ship; an idle view outside races.
9. 10-19
   Disk streaming and Japanese inputs by default The 64DD image streams from SD on every build; Japan-region inputs accepted.
10. 10-19
    v1.0.0 Four build targets from one tree, three downloads (Old 3DS, New 3DS, New 3DS + Expansion Kit; the Old 3DS Kit build stays from-source), replacing the New-3DS-only v0.9.0.

1. 10-20 → 10-25
   Multiplayer Local wireless, same-Wi-Fi and online rooms, lockstep races with each console showing its own racer, a relay server, version gating. Old 3DS: 2D 60 in a 16-racer online race; 3D locked to 30. See MULTIPLAYER.md.
2. 10-24
   The Old 3DS 3D ceiling probe⁴ Waiting out the previous frame's GPU work first made the frame no cheaper (13.5 → 14.2 ms), so a deferred-wait redesign was dropped: 3D races here are CPU-bound.Later (10-26): card 6 found 3D menus waiting 7.66 ms a frame on the GPU queue (per-entry latency); one submission per stereo chunk took them from 40.2 to 44.6 fps. — o3ds-card6-analysis.md §3
3. 10-26
   HOME close, fixed³ Card 5's Old 3DS hard lock was a 73 KB allocation in the 64DD save flush; the flush now allocates nothing.Later (10-26): while chasing this lock a 617 KB drop in free heap was reported as a regression. The two logs compared were a 2D multiplayer session and a 1P 3D session; the comparison was wrong, not the build. — o3ds-card6-analysis.md; 9d19a5c
4. 10-26
   One submission for both eyes² From a homebrew developer's public question. Card 6 A/B: Old 3DS 3D menu 40.2 → 44.6 fps, race GX entries 9 → 7, 1.5–1.8 ms of headroom; New 3DS neutral. v1.1.0 default.Held (10-26): the menu result (6 vs 9 blocks, ±0.2 fps) is solid; the race figure rests on two blocks per arm. The Old 3DS frame has 9 queue entries, not the 11 the pre-card notes assumed. — onesubmit-progress.md; o3ds-card6-analysis.md §0
5. 10-26
   Track arena and a wider 3D sky¹ The course track moves into one 777,360-byte arena reserved at boot, aimed at the fragmentation failure card 6 hit; sky, floor and clouds widened by the eye shift.Later (10-26): built for card 6's failure (a 96,768-byte allocation refused with 1.6 MB free: fragmentation, not exhaustion); tested in the emulator, not yet on hardware. — RELEASE-NOTES.md v1.1.0 known issues; o3ds-card6-analysis.md §5.1
6. 10-26
   The bottom screen, last pass DASH layout, D-pad navigation in the multiplayer sheets, one SETTINGS button.
7. 10-26
   v1.1.0, prepared for release This page's last entry.