Production desktop idle CPU audit

Goal: all non-realtime games on the production desktop should use little CPU while idle. Local acceptance complete; merged, not deployed.

2026-09-10 final status: user authorized the shared merge and explicitly accepted synthesis CPU around15%. Main fast-forwarded to4a341973. Existing tracked edits were restored with conflict reconciliation; recovery autostash 6eb880b3 is retained, and unrelated work remains unstaged. Post-merge paired build passes (canonical1073704/compat1074165, layout e5ea7699f58899e4), as do browser park/cooperative deadlines, clock43/context, Win16 WaitMessage, dialog timer/posted/host input, touch and single-app regressions. No unresolved merge conflicts. Historical open checklists below describe earlier stages; the shared merge is no longer blocked. Quiet-box precision and realtime Marbles investigations are optional follow-ups, not outstanding turn-based idle fixes. Production remains untouched as requested.

User clarification: do not deploy to production. The production desktop inventory defines which games to fix; implementation and acceptance stay local. No credentials are needed, and publishing is not a completion requirement.

Inventory fetched from production lib/apps.js on 2026-09-10. Production host.js reports source version 298. Work is isolated in /private/tmp/wa-rodent-phone; shared main has unrelated concurrent changes.

Measurement contract

Fresh Chrome per app, visible page (document.hidden === false), 1280x900 desktop viewport, 5s startup settle and 10s cumulative process CPU deltas via CDP SystemInfo. Percentages below are renderer CPU as percent of one core, not total-machine utilization. Raw files include browser/GPU/utility costs, load averages, guest state and screenshots. These first-pass measurements are headless Chrome on Apple M1 and the host is loaded: use them to identify core spins, not to claim sub-percent performance precision or real iPhone battery. Final acceptance still needs headful local verification across that inventory.

The initial screenshot is part of the evidence: an empty launch/splash screen does not establish that a dealt game board sleeps. Start-game commands come from the guest's actual menu labels/IDs, not guessed command numbers.

Temporary harness: /private/tmp/audit-idle-games.js; raw production baseline: /private/tmp/wa-idle-prod-baseline/*.json and matching PNGs.

First production sweep

Non-realtime scope also includes Heroes II's player-turn adventure map, not just its title screen. Bricks is a Klotski sliding-block puzzle and belongs in the non-realtime board-game inventory (15 card/board titles, not 14). EmPipe, Funtris and CWordZap still need explicit state classification: inspect static menus/paused states but do not mistake active realtime simulation for an idle wait. Rodent is included because it initiated the investigation. SkiFree, Pinball, DX-Ball, Blobby Volley, Rattler and RollerCoaster Tycoon have realtime active play, not a turn-based idle board.

Local fixes and evidence

1. Win16 USER.112 WaitMessage completes its Pascal far return, then parks with queue yield 15. Fresh queue/clock parks override the host's 120ms recent-input busy window. Rodent idle 101.1% -> 0.66%, paused 101.9% -> 0.60%, playing ~102% -> 2.22–2.56%. See the Rodent reverse-engineering notes.
2. Four Stones' main is in message wait 7; its only helper is in event wait 1 at 0x00405320. Across 10s the main slice counter stayed 523, while helper wait polls grew 2,557,559 -> 5,269,727. hasActiveThreads() means alive, not runnable. After a no-work cooperative pass, parkedThreadDelay() lets the host sleep to the earliest observed wait/sleep/timer deadline. It does not call waitSingle again (which would consume auto-reset events), and remains conservative for pending spawns, productive passes, unknown yields and true Workers. Local CPU 0.80%, from 101.24% on the preceding local run. Regression checks pending messages, timeouts and runnable work.
3. Both native dialog idle pumps now request queue park 15 only after their pending work is drained. The DialogBox continuation keeps EIP/ESP intact; compiled tests verify an idle park, subsequent WM_TIMER, and posted WM_COMMAND. Common-dialog tests verify cross-instance completion after park. Actual Blackjack opening-dialog CPU 0.74%, table awaiting bets 1.44% after its Close button; the rendered table was inspected.
4. Heroes II alternates four stack contexts inside its clock polling wrapper; the single-context detector kept resetting before K=8. A bounded four-context MRU retains each context's independent same-time/no-activity proof. Existing ABI and one-park-per-ms safeguards remain. Menu headless CPU108% ->13.68%, actual player-turn map22.51% ->7.36%. Headful menu 16.15% with same-browser blank renderer0.005% at host load3.42. Residual is still too high to close the audit; about235 clock parks/s plus palette presentation remain. Production headful confirmation is105.59% with zero new parks (host load7.73); local repeat14.99%. Palette census rules out identical updates:15 updates in3s all change RGB, alongside11 surface updates. No timing-default or animation suppression change was made.

After-input acceptance, local304

Four Stones, headful Chrome: click Start! at desktop(90,50), then the first column's entry area at(126,84), holding each click220ms. Screenshot shows the human red stone at the bottom left and the computer's blue reply in column4; the red selection marker remains above column1. After both moves, renderer CPU is 0.94% and the main is back in message wait7. Evidence: /private/tmp/wa-idle-fourstones-move304/fourstones.{json,png}.

Do not substitute Game/New (40001) for Start! (40005): New resets the board. Clicking a hole in the board rather than the column entry also does not prove a move. Earlier no-stone screenshots were rejected as gameplay acceptance.

Headful FreeCell game16371 and Taipei game10112 were screenshot-verified as actual populated boards, not launch/splash screens. Renderer samples0.58% and 0.71%, respectively, but host load9.68/19.42 means these establish absence of a core spin, not precise final performance. Raw wa-idle-dealt-headful304.

Marbles actual one-player board reached using existing playability readiness predicates: wait for menu pixel guest(300,105), held Play click at page(500,650), wait for skill-panel pixel guest(100,300), then Enter held500ms. Fixed5s startup delays were insufficient; earlier clicks landed in intro transitions and their results were rejected. Actual board has two competing marble grids and continues simulating without input (32.07% renderer, host loaded12+). This is active realtime gameplay, not static turn-based idle acceptance. Practice/paused state still needs checking. Raw wa-idle-marbles-board304, screenshot inspected.

Implementation checkpoint: isolated commit 1ce2532e, after phone UI/runtime commit f298b7f9. No main merge or production deployment. The subsequent user clarification prohibits deployment; the earlier credential request is withdrawn.

Three-app headful check (wa-idle-three-apps304): populated FreeCell, Four Stones and Blackjack's opening dialog run together. Across20s, main slice counters FreeCell5 ->5 and Four Stones6 ->6 remain parked; Blackjack319 ->712 is about20 queue-park wakes/s, not an immediate loop. All three remain running, document visible, no page errors; combined renderer1.45%. Host load10.56 means repeat on a quiet box for exact CPU acceptance, but no shared core spin appears.

Heroes settled headful run:60s warmup followed by30s measurement at host load2.43 still15.59% renderer,7037 clock parks. Startup compilation is ruled out as the residual explanation. Profile: WASM3.6% wall time, native/program7.7%, remaining host scheduling/presentation costs. Evidence wa-idle-heroes-settled304.

Marbles p key did not produce a verified paused screen; its resulting 19.53% sample must not be recorded as paused acceptance. Actual pause binding remains to be established (guest pause routine0x41a975, thunk0x401370). The physical Pause key also fails to pause: wa-idle-marbles-pause-key304 screenshots show the opponent counter changing1 ->2 and both grids advancing. Its42.52% loaded-host sample is active gameplay, not paused acceptance.

Extended production sweep: Bricks menu0.46%, EmPipe8.34%, Funtris102.18%, CWordZap2.16%, Heroes II menu107.04%. Funtris has a sleeping helper (not a runnable one); the same local scheduler fix reduces it to 0.76% and EmPipe to 0.55%. Heroes II's menu is explicitly in scope per the user's follow-up; see its RE notes for the clock-context fix and remaining work.

Local source/cache version 304, canonical and compat build pass, layout hash 7dc7fccf292ba722. Focused Win16 stack, dialog timer/command, common modal, park-sleep, cooperative-deadline and Worker scheduler tests pass.

Completion checklist (still open)

Visible local304 inventory sweep

/private/tmp/wa-idle-cards-headful304,10s per game, separate visible Chrome processes; all ten screenshots inspected in wa-idle-cards-all.png, no page errors and all guests remain running in message wait7. Host load ranged6..23, so these rule out core spins but are not precise quiet-box performance claims.

Golf's higher result is not an unparked main loop:52 guest runs over10s, with message wait7 and timer deadlines between runs. Profile timer/presentation cost and repeat under low host load before accepting the5.32% as persistent.

Follow-up wa-idle-golf-settled304:30s warmup,20s sample gives 1.00% renderer, main slices125 ->205, no errors. Host phases total57.5ms/20s, of which guest work7.9ms, present30.8ms and other18.8ms. Thus the5.32% outlier is not persistent. Load5.44;3s CPU profile is96.7% idle and0.1% WASM.

After-input checks (both headful,15s samples):

• Minesweeper0.94%, wa-idle-mines-revealed304: click first tile at(100,145) held180ms. A1 is revealed; counter advances004 ->019, screenshots inspected. The earlier(100,130) edge click did not reveal a tile and was rejected.
• Blackjack1.75%, wa-idle-blackjack-dealt304: Close notice(760,360), Play menu(168,30), Minimum Bet(222,52), then OK(286,855),180ms holds. The5-unit bet is on the table, player7+8 totals15, dealer shows3; helper strategy window is visible. Main returns to yield7, slices134 ->164 over15s, no errors. This verifies real deal input after the modal idle fix.

Heroes actual player-turn map, headful (wa-idle-heroes-map-headful304): NEW GAME / STANDARD / OKAY through real mouse input, screenshot inspected. Renderer10.37%, GPU2.59%, audio utility0.74%,4441 clock parks over20s; no page errors, map remains running. Music and palette animation remain enabled. This completes the map-state verification, not a claim of sub-percent CPU for the audible animated game. The separate menu audio-isolation evidence above explains why real-time synthesis must be distinguished from idle guest polling.

Remaining work

Additional visible local304 checks (details in each game's RE notes):

• Funtris Game/New starts a visibly falling tetromino,1.68% renderer over10s. Its sleeping helper fix does not freeze gameplay. Active play is realtime.
• CWordZap Start + ready click + Pause produces a real paused board with a Resume button; pixel-identical across15s,1.40% renderer and302 queue parks.
• EmPipe Next + Pause produces a static actual pipe board. Normal audible repeat9.77% renderer; AudioContext-only diagnostic0.70%. Both delay censuses return50ms on every call (about20 wakes/s), so music is the main residual, not a busy guest loop. One earlier host-phase-instrumented excessive-wakeup result remains unexplained; do not call it a reproduced runtime leak.

Bricks headful local304 follow-up: actual started Klotski board 0.93% renderer, then 0.84% after a real drag,15s samples. Main slice counters remain4 ->4 and9 ->9 respectively, message wait7, no page errors. The drag changes1838 board pixels and the rendered counter reads MOVES: 1; this is not an untouched launch screen. Raw wa-idle-bricks-board304 and wa-idle-bricks-drag304, screenshots inspected. Host loads18.15/15.65 limit numerical precision, but zero guest runs during each sample establishes that the board is not spinning. No additional runtime change was needed.

EmPipe combined phase+delay census (wa-idle-empipe-combined304) confirms296 turns/15s, all50ms, zero guest work and11.3ms host phases. Renderer7.97% with music. The phase hook alone does not reproduce the earlier excessive wakeups. No additional runtime fix follows from that outlier without a reproducible cause.

Outstanding acceptance decisions / work

• All15 traditional card/board titles now have visible actual-board evidence; Four Stones, Bricks, Blackjack and Minesweeper additionally have explicit move/deal/reveal follow-ups. Funtris gameplay and EmPipe/CWordZap paused boards are verified, so those are no longer generic untested checklist items.
• Heroes' audible menu/map still use about10–16% renderer CPU. Its core spin is fixed. User explicitly accepts roughly15% for synthesis: "synth is not free." This residual is no longer an acceptance blocker; preserve the music and animation rather than making audio optimization a prerequisite.
• Marbles' actual active play is realtime; Pause/P have not paused it. Keep any practice/paused investigation distinct from the15 turn-based titles.
• A same-browser blank-page floor was measured for Heroes. The three-app shared-scheduler sample passed at1.45% on a loaded host; a quiet repeat remains useful for precise CPU acceptance, not for discovering a core spin.
• Implementation and evidence are committed in the isolated worktree; shared main integration remains separate because of overlapping concurrent edits. Paired source304 artifacts remain unchanged. Do not deploy.

Local integration, source305

Prepared in /private/tmp/wa-idle-integrate-20260910 from committed main 453a00c2 plus the complete isolated phone/idle series through51c8b2a0. Preserve main's symbolic region owners, larger DLL tables and newer tests; retain the Win16 builtin-name extent needed by the phone fix. Regenerate the map rather than accepting either parent's generated hash.

Main's guest/profiling clock split requires one semantic integration repair: helper park deadlines and the host main-Sleep remainder now read _waitNow, the clock that created those deadlines. Input grace still uses host time. Both new split-clock assertions failed before this repair and pass afterward.

Full build gates pass (933 test files accounted for, not933 executed tests), canonical1067238/compat1067695 bytes, matching layout e5ea7699f58899e4. Focused tests pass: browser park/scheduler, cooperative deadlines, clock43+ contexts/batch, Win16 WaitMessage, dialog timer/command, cross-instance modal, Worker scheduler48, ThreadManager, touch controls and single-app layout.

Visible three-app smoke (wa-idle-integrated-multi305),15s: FreeCell game1467, Four Stones and Blackjack notice all alive/error-free, renderer1.51%. FreeCell and Four Stones main slices stay9 ->9; Blackjack186 ->480 queue parks. Rodent visible smoke (wa-idle-integrated-main305)1.19% renderer, real board rendered. These verify the integrated local build, not a production change. Heroes menu smoke in the same output directory:13.12% renderer with music, no errors; screenshot inspected. This is within the user's accepted synth cost, with the clock-context park fix retained.

Shared-main merge still awaits a safe window: active dirty host/index/touch/ renderer edits belong to other work and were not stashed or overwritten.

Follow-up integration includes committed main2cadff13, notably the newer retained modal host-input callback path. It merges without conflict with the idle pump. Source306 canonical1067255/compat1067712 build passes with the same layout hash. Dialog timer/posted/host-input, renderer modal routing, cross-instance common modal, browser park and cooperative deadline tests pass. The shared checkout still has overlapping uncommitted changes; it is not safe to replace or stash them while their owners are working.