CLI CPU service clock

The CLI defaults to an aggregate virtual CPU clock. This repairs the old publication-dependent charge: the same 4,000 retired blocks used to represent 200 ms in one outer batch and 2,000 ms in ten. Device cursor, waveOut eligibility, GetTickCount and scheduler deadlines now use one service clock.

The model is floor(totalRetiredBlocks * 1000 / blocksPerSecond) + idleMs. A block is the interpreter's existing retired basic-block unit, including callback and loader work. It is not an x86 instruction or a hardware cycle. The default is 5,000 blocks/second, preserving the previous nominal single 1,000-block, 200-ms slice capacity. It is deliberately independent of game startup results. --cpu-blocks-per-second=N selects a different abstract CPU. This change establishes accounting, not a claim that this default reproduces any particular PC's speed or makes every game playable.

All guest threads share one aggregate service budget even when Node runs them on separate host CPUs. Host parallelism does not create extra virtual cores. A shared integer ledger banks wrap-safe deltas from each actual WASM instance's get_blocks_now. Imports publish live progress before externally observable host calls; export facades bank final progress, including nested execution. Native blocks_now includes the active enclosing run budgets during recursive runs; otherwise entering a nested run temporarily hid outer progress and the ledger interpreted that regression as a 32-bit wrap. The outer prefix is saved and restored alongside the existing budget, without changing guest contexts. Workers use the same ledger for their local get_ticks import. Export facades record time when an execution actually retires work. Sleep and bounded-wait origins use that saved time, including Worker replies, so later sibling service or delayed result delivery cannot postpone a deadline. Read-only diagnostics retain the original execution time. Worker timer hints retain their observation time and spin/vblank parks retain absolute native deadlines. Fractional ms survive slice boundaries; clock reads and outer publications add no time. This specifies time at service observations, not deterministic ordering of races between parallel guest threads or deterministic external input delivery.

When all observed frames are blocked, the runner advances to the earliest sleep, bounded wait, timer, spin, vblank or audio deadline. It gives runnable siblings their existing turns first. The idle hint does not poll waitSingle or waitMultiple, consume a semaphore, restore a callback frame, or run a parked thread. Reserved grants, callbacks being offered and in-flight Workers forbid idle advance. A complete pass with no retired work is required before idle advance, so a late signal gets another ordinary scheduler poll. Unknown asynchronous states remain conservative. An empty message poll can advance to the next observable millisecond; an indefinite object wait with no device deadline cannot invent elapsed time. waveOut still requires its own cursor eligibility; idle advance does not force a completion.

Service mode uses elapsed-time wait deadlines without legacy minimum-batch poll floors or early timeout shortcuts. Zero-time origins remain valid wait origins. Existing callback routing, wait completion, semaphore grants and round robin policy remain owned by their existing implementations.

--clock=batch, an explicit --tick-ms-per-batch=N, and an explicitly registered cliTickMsPerBatch retain the old batch contract. --real-ticks retains its wall clock. Scripted startup dilation and gameplay benchmark clock mappings require --clock=batch; they are separate time models. Runtime tick-ms changes require that explicit batch selection.

Regression: node test/test-service-clock.js runs actual built WASM with the same audio header, work and CPU state across 1/2/10/40 partitions, before and after an identical idle wait. It covers main plus actual guest-worker transport, shared clock reads, runnable sibling exclusion, earliest idle deadlines, reserved grants, timed waits and non-consuming clock reads. SERVICE_CLOCK_BASELINE=1 selects the old policy and fails the same partition assertion. The default test also invokes actual-WASM nested-run and actual CLI idle-boundary cases. The nested counter test failed before the native fix (10 -> 1) and passes with 1,007 total blocks. SERVICE_CLOCK_CASE=worker isolates real Worker transport; SERVICE_CLOCK_CASE=idle isolates all-blocked deadline advancement. Existing clock/audio/callback/scheduler/wait/grant tests are retained. Evidence and original GLTron outcome: run 20261110-gltron-partition-invariant-clock-repair-1740 (ignored artifacts).

Original/default GLTron remains unqualified. The repaired native-counter build ran for 120 seconds in cooperative and real Worker modes; the final wait-origin revision ran another bounded 120-second Worker check. All exited normally with separate stdout/stderr EOF, accepted screenshot commands and full black 640x480 captures personally reviewed. They reached audio (48,942 accepted cooperative callbacks; 33,074 in the final Worker run), but recorded zero scene draws/vertices. No gameplay was observed. The accounting regression is repaired; abstract CPU throughput and original startup/audio progress still need separate investigation. Full27 remains open. No rate was selected from menu success.

Real-time CPU/device alternative

--real-ticks is a separate, existing general policy: the guest CPU runs at the interpreter's available host throughput, while guest timers and headless devices use elapsed host time (1x unless an explicit time scale is selected). It makes no emulated Pentium frequency claim and gives no throughput guarantee on a busy host. It is suitable for measuring whether available execution can keep up with a real-time device. It does not calibrate the abstract service CPU. The deterministic service default and its aggregate 5,000-block rate remain unchanged. No GLTron-derived rate is introduced.

An app-independent control exposed a defect in that alternative: the CLI bound headless audio to batchClock.batchTicks() even when guest threads selected wall time. One outer publication could consume 200ms of audio while the guest observed only 100ms; ten publications could consume 2,000ms at the same elapsed time. The repair binds shared audio to guestClock.publish(), the same selected, non-consuming clock already published to guest threads. Batch and service modes retain their existing source; real-time mode now also applies its selected scale/startup mapping to device consumption and completion eligibility.

node test/test-guest-clock-source.js executes the actual CLI binding with production host imports, actual built WASM and a 22,048-byte waveOut buffer. It holds elapsed time and 4,000 retired x86 blocks fixed across 1/2/10/40 work partitions. At 100ms a 22,050Hz stereo 16-bit device has consumed 8,820 bytes and its buffer remains pending. The test also covers 3x time (completed), a half-speed elapsed mapping (pending), and non-consuming batch reads. The pre-repair binding fails the same clock equality assertion. The service-clock partition, actual Worker, idle and nested-counter regression remains separate. Evidence: scratch/runs/20261110-gltron-independent-cpu-cost-cause-1820/.

One original cooperative run with this repaired real-time alternative used a predeclared 180-second CLI limit, 240-second total child deadline including setup/cleanup, and independently prearmed 270-second watchdog. It exited zero after 183.566 seconds, with actual parent waits, separate stream EOFs and no pending command. No guest setting/media change or CPU-rate search occurred. The original music refill returned 467 times; its first return had occupancy 157,312, exceeding the original 155,648-byte exit requirement. All 10,243 observed ring states conserve bytes: 1,944 decodes produced 7,962,624 bytes, 355 reads consumed 7,827,040, leaving 135,584. Decodes still cost 59,252–86,974 retired blocks; no decoder semantic repair was made or inferred.

Full 640x480 captures were personally inspected: the original main menu, GAME submenu, then arena and yellow cycle with GAME IS PAUSED. Ordinary Return down/up navigated and started the game. A late keypress:13 emits WM_CHAR, which did not unpause this SDL game; the final capture remains paused. Moving gameplay/input qualification and Full27 remain open. A requested A character arrived after normal CLI exit and was not submitted. The next input check is ordinary Return down/up to unpause, then the game's turn controls. This run establishes refill return and original scene/menu progress under the independently specified real-time policy; it does not change or qualify the default service CPU.