Seed-equalized proof of the native switches (lane b5-native, 2026-10-27)

Acceptance evidence for flipping MKDS_HW_FS, MKDS_HW_ARM7 and MKDS_HW_TIME to native, and the first 5.4 (budget-free) result. Mask: native_seed_equalized.txt (beside this file).

Method

• Seed pin (MKDS_TICK_PIN, test only, os_hle::TickPin): the race seeds its RNGs from OS_GetTick (course_init_checkpoint_progress 0x020401AC stores tick mixes into the race manager's RNG states +0x47C and +0x498; race setup and the menus' random picks do the same, os_hle::TICK_SEED_SITES). The exact run logs those results (MKDS_TICK_PIN=log, 20 calls in the reference route), the native run returns the logged values at the same call sites (MKDS_TICK_PIN=<log file>). Nothing else is changed.
• Open loop: frame_compare replay of the reference inputs (b5-harness recording, 250 rows), 12,100 frames, with --mask native_seed_equalized.txt on both runs.
• Load lead: native loads finish earlier, so the race starts earlier in V-blank count: FS native and all three switches by 4 logical frames, the budget-free build by 18. Inputs from the race load on are re-keyed by that lead (inputs.csv rows >= 1976 minus 4, >= 1963 minus 18), and frame f of the exact run is compared with frame f - lead (region_cmp.py --offset). ARM7 and TIME native alone have no race lead.
• Regions: the 5.0 recorder's (karts, race manager, item slots, object RNG, race config, records, controllers, DTCM/ITCM, main RAM per 64 KB, 2D/3D I/O, palette, OAM, VRAM, polygons, PXI, SPU).

Results (race: logical frames 1980..12063, all unmasked regions)

The lead comparisons also mask the V-blank counter (0x027FFC3C), which differs by the lead by definition. Hardware frames of the finish differ by the lead (11,462 - 4 = 11,458; - 18 = 11,444).

Before the race (menus and loads, frames 0..1979) the runs are equal outside the loading transitions, which move with load speed: TIME 1,683/1,703 frames equal up to frame 1719, ARM7 1,387/1,695, FS 1,471/1,697, all three 1,258/1,703, budget-free 1,075/1,704; the race load itself (1720..1979) is shifted.

Remaining divergences and why (all masked, see the mask file)

• TIME: tick-derived values only (the 5.3 mask), plus the cartridge lock word owner at the sample.
• ARM7: the native RTC is a fixed clock (RTC area 0x027FFDE8 and an ID derived from RTC + tick at 0x02180488, copies at 0x02174F78/0x02175078); lock words 0x027FFFB8..; DISPCAPCNT at the sample (ARM7 alone only).
• FS: the card driver's work area (0x02180040..0x021803C8, 512-byte page buffer) is not used.
• Sound timing (FS and ARM7): each kart starts with a sound emitter; its voice-handle list (+0xC/+0x10) and active count (+0x14) depend on when the ARM7 sound driver reports a voice done, which both switches move by a few ARM7 bursts (the SND command queue 0x021732B0.. and SoundState's handle pool are in the already-masked main_02170000). Kart positions, physics, items and the RNGs are unaffected.
• IPC delivery (task 2): with MKDS_HW_ARM7=native + MKDS_HW_TIME=native the PXI/SPU streams move between adjacent frames (masked pxi/spu), but no game state differs, so 5.3's model does not need synchronous replies for correctness. The budget-free mode answers native-service PXI words synchronously (Machine::pxi_sync), which removes the polled deliveries in the PXI wait loops (47 -> 1 per race).

5.4: budget-free build (MKDS_TIMING=native when building)

Generator: mkds_gen --timing native emits ARM9 budget! as nothing; loop headers and the census busy-waits keep it as poll! (4,164 poll sites; 122,736 budget! sites are empty). The ARM7 table keeps its cycle budget. Runtime: waits end the slice; native PXI replies are synchronous; Session::step_frame runs to the next V-blank count.

Closed-loop race (acceptance rom.nds route.csv OUT 15000, each run alone on the machine):

About 20% less CPU per frame than the exact build; the rest is the ARM7 CPU (sound), the hardware model ticking in waits, and the acceptance driver's captures.

Plan 5.4 conditions, checked against the census

1. No hardware unit derives time from instruction counts — partly. TM0/TM1 are virtual (5.3), FS and the small ARM7 services are native. Still cycle-driven: the ARM7 CPU running the sound driver (lane b5-sound) and Bus::advance (scanlines, V-blank). Without a budget nearly all hardware time passes in the frame-end wait, so this is benign, but it is not yet "advance only at frame events".
2. No code observes cycle timing — holds for the reference route: VCOUNT 0 reads, DISPSTAT 2 (boot), TM1-3 0 reads, TM0 only through OS_GetTick (virtual), no H-blank / V-count IRQs, GXSTAT reads see the immediate geometry engine. DISPCAPCNT's end-of-capture bit is hardware-timed but only shows in the mask. Menus beyond the route, battle and missions are not in the census yet.
3. IRQs and switches only at safe points — holds under MKDS_HW_TIME=native; the watchdog still fires once per race at MI_CpuFill32 (0x0214D01C, a 1.1 M-instruction fill while loading; harmless).
4. Residency re-checked at FS_AfterLoadOverlay — not implemented as stated; residency is still re-checked at every dispatch after code was written (code_epoch) and at slice starts, which also works without a budget (the race loads its overlays and finishes in both runtimes).

Reproduce

# exact, logging the seed ticks; then each native switch with the log
frame_compare replay EXE rom.nds inputs.csv OUT_EX 12100 --mask native_seed_equalized.txt --env MKDS_TICK_PIN=log
grep MKDS_TICK_PIN OUT_EX/replay.log > ticks.txt
frame_compare replay EXE rom.nds inputs_m4.csv OUT_N 12100 --mask native_seed_equalized.txt \
    --env MKDS_TICK_PIN=ticks.txt --env MKDS_HW_FS=native --env MKDS_HW_ARM7=native --env MKDS_HW_TIME=native
python region_cmp.py OUT_EX/rec OUT_N/rec --mask native_seed_equalized.txt --from 1980 --offset 4