MDK2 measured native boundary observation

Input is cd788045394d877eb28cdb9f01c3acc62cc13bb4 and mdk2-installed-original-gameplay-20261104.md. Evidence is scratch/runs/20261104-mdk2-measured-native-boundary-gameplay/ (RUN). The previous Medium press exceeded the harness's 100000-boundary ceiling. Its final 13000 boundaries and terminal report remain unknown. That observation does not establish a game or emulator defect.

The replacement contract is one continuous history, at most 50 million native boundaries, 512 records and 360 seconds, including personal review pauses. The caller cannot restart or renew that history. Existing original-byte, module/instance/thread ownership, stack/history, native input, poll, ACK and frozen-register checks remain. Record admission is capped at 512 even when a caller supplies its own bound. No guest instruction, configuration or pacing change is involved.

Both retained modules execute the original decoder at 0x47cea0, containing the previous fault PC 0x47cee7, through the real serialized binding/observer and shipped Worker queue. Explicit test data supplies 65536 decoded output words and a real original decoding table/bitstream path. The old caller throws at 100000 boundaries. The repaired caller returns after 156000 boundaries, including the final slice's bounded caller padding, over nine continuous checkpoints. Full 512MB memory, registers, lazy flags, decoded output, commands and subsequent genuine native trap match the corresponding control. This is measured original native work with initialized fixture data, not reconstruction of the lost game's decoder state. The generous finite ceiling accommodates multiple such operations; the independent fixed wall and record limits still apply. The actual game's needed work is established only by its fresh run.

Observation exhaustion now unwinds the native run call with an identity-checked private exception and retains a queryable terminal Worker report. Only boundary, record and deadline exhaustion receive this treatment. Genuine native traps, original import failures and unsupported native history keep their original failure behavior. The next slice cannot execute native work after observation exhaustion, including expiry while idle. The original caller persists the terminal report before its first-failure closure. Both-module real queued tests exercise boundary/record/deadline exhaustion and unchanged stopped state; old record/deadline callers reject where the repaired Worker retains its report.

Validation before Chrome: 143 integration checks, 38 local checks, both-module observer parity, required full flock build, and the complete actual fresh-directory entry through filesystem/HTTP pins, local installation and recorder to an independently spawn-guarded mocked Puppeteer boundary. That entry spawned zero Chrome and closed both servers. Accepted unchanged prior 135-integration/module proof remains referenced by path/hash; no historical screenshot rehearsal ran.

The local assignment and exact predecessor parser now identify this task and the Stronghold fresh-complete release. Browser/bootstrap/command/cleanup bounds remain 720/60/40/180 seconds with absolute 23:50Z/23:53Z cutoffs. The original attempt uses a long-lived external /tmp/wine-local-runtime.lock owner after GLtron's 21:22:51Z and Stronghold's explicit 21:23:22Z releases. Box ID, state and expiry are NA; coordinator retains shutdown ownership.

One original Chrome launched. Fresh startup, title and opening-menu captures were personally viewed. Ordinary Escape passed its original native press/rest readbacks and was balanced. The existing 16ms clock was retained by the same small run-only selector adaptation that omits the published tick20 operation and its acknowledgement; every other selector check remains. Original helper bytes and guest code/configuration were unchanged.

At native-selector-75, 3200 frozen steps, the fresh image already shows the highlighted controller and readable New Game label. The worker submitted the preceding opening-state classification (selection: none, highlight/label false) instead of reassessing that new scene. The original selector correctly rejected wrong native selection at 21:31:53.734Z, before another step or input. The inspection approval had already been published by the preceding controller. This is an operator review error, not a selector, game or emulator defect. The erroneous receipt is preserved alongside its explicit correction and the actual image. A separate earlier legal-overlay boolean error at capture 69 is also explicitly corrected without rewriting its receipt.

The attempt stopped immediately. Quit command 78 requested ordinary closure; no New Game/Medium input, loading, gameplay or actual construction trace ran. The offline boundary repair therefore has no fresh Medium/gameplay validation. No second Chrome, diagnostic game or postrepair original run followed.

The causal offline controller repair moves the original selector's menu review classification check ahead of inspection-approval publication. It rejects contradictory fields without guessing pixel meaning or correcting human input. The original selector repeats its own check later. A regression executes the actual generated Python control and local transport: old ordering publishes the bad approval and then rejects the same selection; new ordering rejects before any approval is written. Honest New Game review fields pass the original rule. Final validation is 39 local tests including that old-fail/new-pass regression and a fresh complete-entry zero-Chrome test, the focused generated-controller check, the prior passing 143 integration checks, both-module observer parity, and the required full flock build. Failed offline test setups and their logs are retained. The review repair was tested only offline after the original.

Chrome PID 1192996/start ticks 187494444 exited 0 with null signal through browser.close at 21:32:01.026Z. Node PID 1192857 and external flock PID 1192856 (both start ticks 187494231) exited 0 and are absent. Both HTTP servers closed; ports 41947/33819 and owned DevTools 34095 refused connections. The exact profile was removed after process absence and the runtime flock was verified free. Explicit heavy release was posted at 21:32:51Z. No foreign process was signalled and no box lifecycle operation ran.

Original runtime Git is 1329b1ba6eb00144ff77d4cd92a8513d4fc953ad; shared launch Git is ae34cf1305b9057930b7a6c685a9062a37a70bb5 plus RUN's launch-time relevant binary diff. Separate final diff/new-note patches preserve the later offline repair. The original module hashes remain cffca869b541e6eff12ae3a6b74c4476977e0a7e23cd43d8ec5ba8faecd4bcee and 580f49f3c78d26e6855efa0beb37046d03ea24690ff98cfcb3e94ce5e6cb1ee3. The prior run's 576 sealed files / 34295096 bytes were verified unchanged. Existing media/modules are referenced by path and hash; no source tree, media/module snapshot, tar, npm, boat or subworker was used.

The next original attempt requires separate authorization and honest fresh scene review through the repaired preapproval check. Visible gameplay remains unproved. This single assignment stops after publication and release.