MDK2: retained construction observation across queued quanta

The first continuity dependency is repaired offline. The actual composed command producer now checkpoints its existing observer after a construction quantum and resumes that same observer before the next quantum. It retains native call state, records, counts and its original deadline. No new original or Chrome diagnostic was launched. Resource completion, Lua execution, Level creation and gameplay remain unproved.

Input is 91d6a3a1583b956c1c0de4e41309f8eaaac3f0e1 and docs/re-notes/mdk2-repaired-launch-original-20261104.md. Evidence is scratch/runs/20261104-mdk2-continuous-construction-onward/ (RUN). The predecessor seal is checked without rewriting any predecessor artifacts.

The repaired Worker binding leaves tracing armed while frozen for review. Its checkpoint samples the existing observer and retains the owning register and lazy-flag values. Resume requires those exact values, the same Worker/session, page compilation/transfer binding, and independently frozen main/file sidecars. The driver repeats its input, queue, lifecycle, scene and review checks before execution. A boundary delivered during a review is an explicit first fault. Indexed start, end and correlation records preserve each quantum's evidence; later commands cannot overwrite the only copy of an earlier checkpoint.

Neither checkpoint nor resume calls observer.finish, attaches another observer, clears call history, resets counts, resets the timer or changes guest state. The limits remain 10000 milliseconds of wall time, 100000 boundaries and 512 records for the entire observation. Reviews consume that same wall-clock budget. Terminal cleanup still disables tracing and finishes the observer once. Unknown history before attachment remains explicitly unknown.

The actual retained command functions reproduce both predecessors: the pre-input driver silently executes a second step after finishing its only trace; the input driver rejects that second step. The repaired producer executes two reviewed 16-slice commands with one start, two checkpoints and one resume, without finishing between them. These are protocol fixtures, not original gameplay.

Both pinned modules also execute the original native fragments through the ordinary shipped Worker queue, with a checkpoint/resume after each returned slice. Finalizer entry/return, completion consumption, Sleep/poll return, demand call/return and script-buffer entry span those checkpoints. Full-memory hashes at the same fragment endpoints, registers, lazy flags, returned command sequences and the original undefined-instruction trap match the control. Fixture setup is explicitly synthetic; no resource or Lua state is manufactured in an original.

Changed ownership, frozen state and scenes reject before a second execution. Both-module negatives cover changed owning registers during review, a delivered boundary during review, boundary exhaustion, terminal history and review expiry. The expiry test successfully checkpoints, waits past its original deadline while paused, and rejects resume without extending it. Existing binding, first-script, ACK, input, deadline and native-history negatives remain.

The next actual prerequisite prevents a safe original. The live driver attaches only after the type10 level1 request is already active. Its accepted mid-poll suffix cannot supply the earlier demand-call entry. A new both-module regression executes the original queued Sleep-return suffix, then delivers the demand-return boundary: the observer rejects demand return history/completion missing, and no demand-call record is invented. Preserving the observer fixes termination after one quantum; it cannot recover that missing prefix. The existing ten-second wall deadline also remains binding across personal reviews. This candidate therefore does not establish the capability required to spend the authorized original run.

The worker stopped offline as instructed. There is no renewed runtime release or changed launch assignment/deadline: the prior local contract retains its prior assignment, so it cannot be presented as this successor's release. The repaired launchOwned caller and passive Browser/page recorder remain unchanged. The next worker needs earlier, genuinely observed demand entry before loading, with a review path that fits the retained observation bounds, before another original.

Stronghold's explicit 19:41:24 release was observed. This task acquired no local runtime flock, started no HTTP server or Chrome, injected no ordinary game input, and created no gameplay images. Its local heavy release was posted at 19:54:54Z for GLTron. Box ID/state/expiry are NA; the coordinator owns future boxes and shutdown. Only owned offline test processes and the board watcher need closure.

RUN records the exact Git SHA, relevant binary diff and new-note patch, retained helper/Git-dependency checks, and existing media/module paths with SHA-256 values. No source snapshot, archive, media/module copy, native install, guest code/state/ configuration/pacing edit, box operation, npm install or subworker was used.

Final validation passed: 109 integration tests, 35 local caller/lifecycle tests, standalone both-module observer full-state/command/trap parity and the locked full build. The local suite includes its existing owned Node termination fixture; it does not launch Chrome. All 111 retained helpers, four Git dependencies and both retained modules verified. The predecessor's 927 files / 39310861 bytes remain unchanged. No filesystem/HTTP release was attempted because the continuity gate for an original did not pass.