2026-11-01 sched-applets (batch 2026-11-01-r1, Mac)

Branch claude/sched-applets-2026-11-01-r1. Scheduler semantics fixed for the ACNH start flake; swkbd and the offline web applet verified on real titles.

1. Scheduler: two Horizon divergences fixed; the ACNH race is NOT fixed by them

What was fixed (generic, Horizon semantics)

• Equal-priority rotation only at priority 59. Napple's timer preemption (kernel-threads, 2026-10-27) rotated equal-priority threads every 10 ms at every priority. Horizon's 10 ms timer rotates one priority per core (KScheduler::RotateScheduledQueue with the preemption priorities {59, 59, 59, 63}, mesosphere / yuzu PreemptThreads); threads of other priorities keep their core against an equal-priority thread until they block or yield. NAPPLE_SLICE_ALL=1 restores the old rotation for comparisons.
• A preempted thread keeps the head of its priority. Horizon leaves a thread preempted by a better one at the front of its run queue; Napple re-queued it behind every equal-priority waiter (core_put + core_take with a fresh sequence). Preemption points (timer signal, svc-exit checkpoint, SynchronizePreemptionState) now use core_switch, which re-queues at the head when a better thread takes the core.
• Main thread pinned to its ideal core (KThread::Initialize: affinity = the ideal core only; it was 0x7). ACNH pins it itself at 0.96 s.
• Unit tests (test_threads_preempt, 12 checks): P2 rotation at 59; P5 no rotation at 47 (fails with NAPPLE_SLICE_ALL=1); P6 the preempted thread resumes before an equal-priority waiter (fails without core_switch, checked with a patched copy). Debug aids: NAPPLE_THREAD_LOG (+ core mask / priority changes), NAPPLE_THREAD_STATS=1 (per-thread timer / svc-exit switches, GetCurrentProcessorNumber), NAPPLE_PREEMPT_SIGNAL=0.
• Measured effect (NAPPLE_THREAD_STATS, ACNH base, 25 s): thread 0x54 (prio 44, core 0) lost its core to equal-priority rotation 21 times with the old rule, 0 with the new one; other threads' counts are unchanged (better-priority preemptions).

The ACNH base race (open)

All counts are ACNH base 7fc1baff976aeca4, 30 s headless starts, NAPPLE_GFX_CBTRACE=1, run lock held one start at a time; "overlap" = a run with recorder anomaly dumps (two threads on one command buffer).

Runtime	Starts	Guest crashes	Runs with overlap
base-commit scheduler (threads.c of df2f318, same build otherwise)	20	0	8
rotation fix only	10	2 (main+0x1b61d3c null+0x28, wild at main+0x19c3594)	3
final (rotation + head-of-queue)	40	2 (main+0x1b61d3c null+0x28)	11
final, NAPPLE_SLICE_MS=0 (no timer preemption at all)	10	1 (different: null pc main+0xabb7c at 6 s)	2
final, NAPPLE_PREEMPT_SIGNAL=0	10	0	5

So the overlap happens about as often with or without the scheduler changes and also with timer preemption switched off: the b8 observation "0 of 60 with NAPPLE_SLICE_MS=0" was a small 5 s sample. Neither preemption nor equal-priority rotation is the cause. Crash counts (0-2 per series) are too small to compare. ACNH update+DLC: 10 of 10 clean with the final runtime (and 10 of 10 with the rotation fix).

What the traces show (call chains converted to main+ offsets from the dumps): - The colliding threads are always the guest main thread 0x51 (core 0) and a worker, 0x57 (core 1) or 0x58 (core 2). Both run the same job path job dispatcher (ret 0x19b7598) -> job item method 0x19b654c (ret 0x19b672c) -> 0x1af9960 -> 0x1b319e4 -> view vtable+0x298 = G 0x1ad9018 -> F 0x1ae75c0 (F binds a stack command buffer context into the shared slot sub->array[k][bufidx] and records). Workers enter from their loop (0x19a39c4 <- 0x19a2cf8), the main thread from the job manager (0x19b6ea0 <- 0x1b4d164 <- 0x197a248), which runs worker 0's share itself. - The job is a parallel-for: the dispatcher pops a job under the worker's spinlock (worker+0x130), the item claim takes [job+0xc4].. under the job's spinlock (job+0x1c); both are ordinary ldxr/stxr locks, executed natively (host exclusive monitor), so the same item is not claimed twice. - The job manager's "wait for the workers" after it kicks them (0x19b6ed0) does not wait: when a worker's state word [worker+0x108] is non-zero it only performs discarded loads (ldr wzr, [xN, #0x108]) and returns, the shape a compiler leaves of while (state != 0) {} on a non-atomic field. The update's copy (0x390d2b0) has the same shape, yet the update never overlaps, so this is a lead, not a proven cause: the real join of the main thread with the workers is elsewhere and was not found in this session. - Ruled out: rewriter exclusive windows (ldxr/stxr stay native unless they contain x18), ArbitrateLock/Unlock, the condvar pair, WaitForAddress / SignalToAddress including modify-by-waiter-count (match mesosphere on reading), per-core data (GetCurrentProcessorNumber is called once per thread), timer preemption (table above). Next step for whoever continues: find what makes the main thread wait for the job group before it re-records the same views (follow 0x1b4d164 / 0x197a248 after the manager returns, and the per-job done counter [job+0x70] incremented by the dispatcher at 0x19b75b0), then check the wait primitive it uses against Horizon (and the gfx sync objects if it waits on GPU progress).

2. swkbd on a real title (verified)

Link's Awakening, fresh save, new file: the title needs L+R (A does not skip the attract movie; the title appears at about 97 s headless), then A on an empty file opens the name keyboard at about 109 s:

applet: CreateLibraryApplet id 0x11 mode 0
applet: swkbd: header "" sub "" guide "" initial "Link" length 1..8 password 0 invalid-flags 0x12c (config ok, work buffer ok)
applet: swkbd: entered "Napple"
applet: swkbd: header "" sub "" guide "" initial "Napple" length 1..8 ...   <- the game's next call carries the answer it got back

NAPPLE_VERBOSE=1 dumps the config words: header (+0x24), sub (+0xA6) and guide (+0x1A8) are all zero in LAS's KeyboardConfig, so the empty texts are the title's own (the non-empty parse is covered by the unit tests). ACNH's name entry (reached at about 55 s on a fresh save) is the game's own keyboard, not swkbd; SMO and MK8D open no keyboard offline. The inline keyboard therefore stays tested only (no title here uses it).

3. Offline web: a real title's manual (verified)

The build extracts HtmlDocument romfs to <data-dir>/html/ as it is: html-document/<name>.htdocs/html/<RegionLang>/index.html (SMO: 12 languages, MK8D: 13). An OfflineHtmlPage document path is relative to html-document/ (yuzu's offline applet joins it the same way), so web_offline_resolve now looks under <root>/html-document/ first, then the root as before. New: web_manual_path (the manual for a RegionLang, fallback USen, EUen, any), web_manual_lang (Switch language code -> RegionLang), applet_web_open_manual (the offline applet run with the WebArgs storages the HOME menu pushes: shim offline, kind HTML page, document path) and the trigger NAPPLE_OPEN_MANUAL=<secs>[:<RegionLang>]. SMO, windowed, NAPPLE_OPEN_MANUAL=25 NAPPLE_SHEET_AUTO=confirm:4000 NAPPLE_SHEET_SHOT=...:

applet: manual: opening htmlcontents.htdocs/html/USen/index.html (language USen) from .../3ca12dfaaf9c82da/html/html-document
applet: manual: shown, exit reason 0, last page file:///.../html/html-document/htmlcontents.htdocs/html/USen/index.html

None of the four titles opens the offline applet on its own in the first minutes; on a console the manual is opened from the HOME menu, which is what the new path reproduces.

4. Soak tooling (tools/soak/soak.py)

--lockf FILE runs each start under lockf -k FILE (this batch's run lock) instead of gamelock.py; every run record counts NAPPLE_GFX_CBTRACE anomaly dumps and recorder warnings (cb_anomalies, rec_warnings; an ok run with any keeps its log); the summary lists runs with anomalies; list tolerates a missing version.

Verified here (Mac)

Check	Result
make -C runtime GFX=napple test, make -C runtime test-host	pass (threads_preempt 12/12, kernel_sched 24/24)
applet service tests (test_svc_applet, _ae, _album, _host, _libs, _web)	pass (86 / 52 / 63 / 539 / 263 / 179 checks)
ACNH base 40 x 30 s, final runtime	38 ok, 2 guest crashes main+0x1b61d3c, 11 runs with overlap dumps (acceptance not met; see section 1)
ACNH update+DLC 10 x 30 s, final runtime	10 ok, 0 overlap dumps
LAS / SMO / MK8D 60 s, final runtime	ok / ok / ok (3532 / 3526 / 3290 frames), 0 overlap dumps
swkbd on LAS (fresh save, L+R at the title)	swkbd: entered "Napple", read back by the game
SMO manual (windowed, NAPPLE_OPEN_MANUAL=25)	shown, exit reason 0 (Done), screenshot above
All game runs were headless (except the SMO manual run), audio off, under lockf /Users/neto/NappleWork/.run.lock, one start per lock hold.

Files

runtime/src/threads.c, runtime/tests/host/test_threads_preempt.c, runtime/src/services/applet.c, runtime/src/services/applet_web_core.{c,h}, runtime/tests/test_svc_applet_web.c, tools/soak/soak.py, docs/architecture/runtime-services.md (appended), docs/todo.md (ACNH start flake), this note and its image.