MDK2 finite menu command allowance, 2026-11-04

The run-only menu controller now supports an explicit 113-command maximum, with 64 still the default. Offline regression reproduces the demonstrated 57+9 >64 rejection and completes the same Medium command envelope under a fresh supported contract. No browser, game runtime, allocation or lifecycle operation occurred. No production defect or production change was needed.

This follows corrected staging and the native menu boundary, commit 2488b55b1cfe108fde7b53355e7386bf374aa865. New owned helpers and evidence are in scratch/runs/20261104-mdk2-finite-menu-budget/ (RUN). Predecessor files remain unchanged.

Finite contract

The supported upper bound is conservative and derived from the existing menu envelopes:

Component Bounded command cost
Shared execution envelope 512 slices /16 =32 minimum execution quanta
Observation/poll commands At most three commands per quantum: state, step, capture; native step/state pairs cost only two
Three one-shot input envelopes At most five additional commands each: before-state, tick, keydown, keyup, final capture
Inspected comic acceptance Two commands: state and capture
Explicit command ceiling 32×3 +3×5 +2 =113

The three input envelopes are title Escape, New Game Enter and Medium Enter. Their original minimum reservation remains nine commands /48 slices; observations reserve three /16 and acceptance reserves two /zero. A 32-slice release costs one step/state pair, so this formula deliberately overcounts. Additional native polling consumes the same aggregate slice and command budgets. Repeated observations, fade observations or repeated acceptance calls receive no new allowance; even misuse that consumes no slices still hits the finite command cap. This is execution credit, never permission to infer readiness or replay an input.

select(dir, review, phase, {maxCommands:113}) explicitly chooses the allowance on the first menu call in a fresh namespace. The API accepts positive safe integers through113. It writes commandLimit into the same persistent menu ledger before submitting the first command. Later calls inherit the stored limit; an explicitly repeated value must match. A legacy ledger without this field retains64 and cannot be upgraded in place. Invalid, fractional, nonfinite, overflowing or above113 values fail before submission and latch the first error.

The shared 240,000ms elapsed cap, 512 slices, individual step bound64, first-error latch, reservation-before-input and per-command write-before-submit accounting remain. The spool still takes the minimum of its existing deadline, menu start+240seconds and recovery reserve. No ledger, elapsed clock, phase flag, ACK or error latch is reset. Native consumption, callback effects and release checks, original selection/readiness checks, personally inspected capture hashes, input balance/history, exclusive lock, contiguous command/ACK validation and no-retry behavior are unchanged.

Offline verification

menu-tests.tap: 90/90 tests against the actual candidate select/Spool implementation and independent command peer. Coverage includes:

existing-suites.tap: 177/177 predecessor native/controller tests pass unchanged. candidate-existing-suites.tap: the same 177/177 tests pass with the small offline candidate-suite.cjs preload routing existing controller imports to the actual new helper. This exercises the existing construction, comic/post-cancel, opening, inspection, native readiness/observer, original PE/script, scene and recovery suites without editing predecessor consumers. Tests are sequential to respect their existing short ACK deadlines. These are offline protocol/native-memory tests, not fresh game execution.

Reproduce from repository root:

node --test --test-concurrency=1 scratch/runs/20261104-mdk2-finite-menu-budget/test-native-menu.cjs
node --require ./scratch/runs/20261104-mdk2-finite-menu-budget/candidate-suite.cjs --test --test-concurrency=1 scratch/runs/20261104-mdk2-corrected-staging-native-scene/test-*.cjs

Only six necessary small predecessor helpers/tests were copied: native-menu, its test, native-readiness, native-opening, opening-boundary and memory-fixture. Four support helpers retain identical bytes; menu implementation/test contain the owned changes. A small candidate-suite preload is new. No source tree, module, media or dependency tree was copied. No tracked src/, lib/, tools/ or ops/ change was made, so the production build gate does not apply. provenance.json, relevant.diff, helper delta and owned new-file patches record exact Git SHA and SHA-256 identities; artifact-hashes.json seals the small run inputs, logs and patches. Git remains the source archive.

Exact recipe for a separately authorized fresh runtime

  1. Create a new run and browser namespace; retain the predecessor's original execution source/module/media pins and corrected small-file staging recipe. Do not reuse its closed browser, commands, ledger or latch. Reconstruct these seven small owned files from RUN's new-file patches if needed and verify their recorded SHA-256 values. Stage the candidate native-menu.cjs beside its unchanged native-readiness/opening-boundary dependencies. Ensure the driver and every controller resolve that same candidate file. Keep all other driver, inspection, native, construction, comic, scene and deadline helpers unchanged.

  2. Re-run the above candidate tests and the corrected prebrowser checks before Chrome: pinned bytes, CJS/ESM dependency imports, registered MDK2/WebGL selectors, isolated HTTP and corrected absolute staging paths. Generate fresh staged/served pins including the new helper. Existing missing retained test-only source/capture prerequisites must be resolved or explicitly accounted for as in the predecessor; do not silently omit a failing check. Launch only under separately granted runtime authorization and its close/recovery deadlines.

  3. Use the original bounded input-free startup route. Personally view and approve the latest fresh title capture. At the first menu call, approve the capture through the unchanged inspection guard, then choose113 explicitly. For example, with R the fresh staged run, B its fresh browser directory and REVIEW the actual latest personally inspected title review:

    node "$R/inspection-guard.cjs" "$B" "$REVIEW"
    node "$R/native-menu.cjs" open "$B" "$REVIEW" 113
    

    open is valid only if the existing original-title guards authorize it; this command is not a generic startup action. If the first authorized menu phase is instead an input-free observe, pass113 there. Never pre-create or edit the ledger to enable the allowance.

  4. Retrieve and personally inspect each resulting PNG and its matching state/hash. For each permitted subsequent phase, approve its own latest capture through the inspection guard and call node "$R/native-menu.cjs" PHASE "$B" "$REVIEW" 113. Use observe only for the original identified opening/fade, newgame only for native-ready highlighted New Game, and medium only for native-ready highlighted Medium. The ledger must preserve its original startedAt/counters/113 limit across calls. Observation caps never establish readiness. Native Enter polling/release still uses the unchanged callback witnesses. Do not repeat the prior erroneous opening classification after New Game is already visible and native-ready.

  5. Continue to construction/comic/post-cancel only when their unchanged reviews/native prerequisites authorize it. This allowance changes no downstream consumption, cancellation witness, scene, readiness or ACK contract. Stop immediately on the first failure, deadline or cap; preserve the latch, close/recover under the authorized driver protocol, and hand runtime custody back to the coordinator. No restart, reset, second browser or increased cap is implied.

This task claims a supported, tested finite offline contract only. Medium/comic/playable-scene behavior remains unobserved in a fresh runtime. No box was assigned or accessed here; the predecessor's box was bx_wm47edh3, original expiry 2026-11-04T08:33:42.970Z. The coordinator subsequently reported STOPPED, snapshot verified 2026-11-04T08:22:50.376Z; that state was not independently queried by this offline worker. Lifecycle custody remains with the coordinator. No successor, subworker, HeroesII or public deployment was started.