Block 5: can native become the default? (lane b5-finish, 2026-10-28)

Recommendation: yes for the hardware switches in the player builds (play.exe / game.exe), after two small fixes listed under "Blockers"; keep exact as the default of acceptance.exe (it is the oracle). Budget-free timing (MKDS_TIMING=native) is ready to become the build default right after, with one listening check. Implemented in the defaults follow-up: player hardware and builder timing default to native. Acceptance follows the build's generated timing; build the exact oracle with --timing exact --with-acceptance. Native RTC uses host local time, with MKDS_RTC=fixed for deterministic runs.

What "native" means now

ARM7 CPU removed (MKDS_HW_SND=native)

• Arm7::plain_boot runs the ARM7 from its entry (crt0, autoloads, arm7_main: OS, PXI, SND, FS, RTC, WVR, SPI init, and each service thread up to its first blocking receive — OS thread switches happen inside the run) until the main loop's first halt. 861,345 plain instructions, once, before the ARM9's first frame.
• The boot's four waits on the ARM9 are answered in place (arm7::boot): the PXI_Init sync-nibble countdown (the ARM9's half is answered by Bus::sync_responder, 8 → 0 on echo), the CTRDG word wait (skipped; the router answers the word later), os_init_lock's wait for the ARM9's lock owner 0x7F (tables written at once, the final extension 0xBF deferred until the ARM9 has finished its own lock init — writing it early deadlocks the ARM9), and card_init's wait for the ARM9's CARD callback.
• After the boot every tag is claimed by the router. OS (12), CARD (14), MIC (9) and any command without a native handler reach the ARM7's own lifted callbacks as plain calls; the SPI service thread's queue is drained as plain calls (MIC, and PM/touch/NVRAM commands that reach it); an OS reset request (command 0x10) is acknowledged as os_reset_system would (sound shut down, 0x10 sent back). MKDS_SND_BOOT=cpu keeps the old take-over.
• Result on every route: ARM7 CPU cycles 0 (ARM7 lifted cycles 0 interpreted instructions 0).

Budget-free timing

• native::enter charges nothing when the machine is budget-free (native::BUDGET_FREE).
• Frame-event time (Machine::event_time, on when budget-free and the ARM7 has no CPU): Bus::advance (scanlines, V-blank, timers, the ARM7's ticks) runs only in waits (wait_for_interrupt, 1000-cycle steps to the V-blank or the next OS event) and at listed busy-waits (os_hle::SAFE_LOOPS, one SPIN_QUANTUM per spin). At other slice ends tick(0) only exchanges GX commands and PXI words. A guard advances time after 65,536 slices without a wait (an unlisted spin); it fired 15 times per run, all in load-time decompression / fill loops (MI_ReadUncompLZ8, MI_CpuFill32, MI_UncompressBackward), which are compute, not waits.
• Plan 5.4 condition 1 (no time from instruction counts) therefore holds, except for those 15 guard steps.

Overlay residency (plan 5.4 condition 4)

The session forces a residency re-check at the next slice start after FS_AfterLoadOverlay (0x020051F4) is entered, and the flat loop ends the slice there. Measured with MKDS_GATE_WATCH: this game never calls FS_LoadOverlay/FS_AfterLoadOverlay; it loads its overlay through its own archive loader (fs_request_overlay_archive_load 0x02059EA4 → FS_LoadOverlayInfo 0x020055F0 → …), so the re-check is also armed at that loader's code-ready step (0x0200583C: data-cache flush, instruction-cache invalidate, BSS clear), entered once per route. The epoch-based check (watched code writes) stays as the safety net.

Evidence

Exact build and budget-free build of branch b5-finish; races run one at a time.

Proof races (closed loop, acceptance rom.nds route.csv OUT 15000)

Branch b5-finish at b061f40 (before merging block4's 5.5 work). "ARM7 CPU" = cycles the ARM7 CPU ran (lifted + interpreted); "plain" = ARM7 code run as plain calls (boot, sound driver, ARM7 interrupt work).

ms per frame (whole acceptance run incl. its captures, back-to-back pairs on a machine shared with other lanes): exact fiber 27.26 s / 12,362 = 2.21 ms (31.24 s = 2.53 ms in the other pair); budget-free all native fiber 19.41 s / 12,236 = 1.59 ms (19.63 s = 1.60 ms). About 28% less CPU per frame than exact (b5-native's budget-free build with the ARM7 CPU: 1.84 ms). The CPU-less ARM7 still runs the lifted sound driver (663 M plain instructions per race, about 54 K per frame); that is the next cost to remove.

After merging block4 (with b5-spill's 5.5 emitter work), same races on the merged tree (machine busy, so only the relative numbers matter): exact fiber and flat IDENTICAL (1,707,950,080 instructions, 581,521 IRQs, finish 11,462 1st); all four switches finish 11,930 1st with ARM7 CPU 0; budget-free all native finishes 11,260 (fiber) / 11,533 (flat), 1st, ARM7 CPU 0, 96 / 94 busy-wait time steps and 15 guard steps each.

Open-loop equality with the exact model (seed-pinned, MKDS_TICK_PIN)

Masks in masks/ (native_all.txt plus per-mode files; kart entries are now kart-relative, kart +0xC-+0x18, so one file serves every course). n4 = exact-timing build with all four MKDS_HW_*=native; bf = budget-free build, all native.

• Battle is not comparable per frame, by design (plan 5.3, "Lag frames"): the exact model overruns frames on the "Nintendo DS" stage with 7 CPU karts (111 lag frames in 6,500, about one in 40 during play); native timing never lags. The runs are equal until the first in-play lag frame (logical 2341) and differ afterwards. Native battle runs finish their 6,500 frames without errors, with PM commands served through the SPI-queue path.
• Menus: comparing by game frame (inputs re-keyed with inputs/rekey.py), the differences are the frames around screen changes (native loads and saves faster; the exact model has 25 lag frames in menus) and a menu sprite animation phase that counts V-blanks (masked in native_menus.txt).
• Without the seed pin the race RNG seeds come from OS_GetTick and differ (expected, see masks/native_seed_equalized.md); every closed-loop native race still finishes 1st.

Census (BLOCK5_CENSUS.md, new section)

Battle, mission and menus show no VCOUNT/TM1–3 reads, no DISPSTAT reads after boot, no H-blank/V-count IRQs and no new busy-waits: plan 5.4 condition 2 holds on all four routes.

Blockers before flipping the player default

1. RTC is a fixed clock in hle7 (replays must be deterministic). A player build should read the host clock (keep the fixed clock under acceptance/replays). Small change in hle7::Router::rtc.
2. Acceptance must stay exact. The race_diff ... IDENTICAL proof and the oracle need the exact model; flip the default only in play.exe/game.exe (or have acceptance set MKDS_HW_*=exact unless told otherwise).

Not blockers, but known

• Audio of the CPU-less ARM7 was not re-listened in this lane (the driver code and tick model are b5-sound's, which passed its message/SPU oracle and RMS check); do one listening pass before shipping.
• Wireless (Nintendo WFC, multiplayer) is not available natively; the exact model stalls at the WFC menu too.
• The microphone: MIC commands go to the lifted ARM7 code (plain calls, SPI queue); no route sends one, so the native MIC path is exercised only by unit-level reasoning, not by a run.
• CPU per frame: see the table above. The ARM7 sound driver still runs as lifted code (plain calls) and is now the largest cost after the ARM9 itself; a native sequencer (plan 5.2b, "later option") is the next big win.
• spill!/reload! (5.5) are untouched.

Reproduce

Build the acceptance oracle with the builder's --timing exact --with-acceptance options (or set MKDS_TIMING=exact). Keep MKDS_RTC=fixed for proofs, recordings and replays. Set all four hardware switches explicitly for native comparisons.

# exact, logging the seed ticks; then native with the log; inputs re-keyed by the load lead
acceptance rom.nds - OUT_EX N --record OUT_EX/rec --replay inputs/mission.csv --mask masks/native_all.txt   # env MKDS_TICK_PIN=log
grep MKDS_TICK_PIN OUT_EX.log > ticks.txt
python inputs/lead.py OUT_EX/rec OUT_N/rec 1700 4700          # lead after a first native run
python inputs/shift.py inputs/mission.csv mission_m4.csv 1650 4
acceptance rom.nds - OUT_N N --record OUT_N/rec --replay mission_m4.csv --mask masks/native_all.txt   # env MKDS_TICK_PIN=ticks.txt MKDS_HW_SND=native ...
python masks/region_cmp.py OUT_EX/rec OUT_N/rec --mask masks/native_mission.txt --from 1700 --offset 4
# menus / battle: python inputs/rekey.py OUT_EX/rec OUT_N/rec inputs/menus.csv menus_n.csv; compare with --by-index
# screenshots for writing scripts: MKDS_SHOT_FRAMES=a,b,.. (PNG files in the --record directory)