MDK2 inspected native menu attempt — 2026-11-04

The run-only input observer integration passed, but this attempt stopped at an operator observation-sequencing error. New Game was not activated, Medium was not inspected or selected, and neither comic nor tutorial was reached. There is no demonstrated production defect and no production change.

Evidence and executable run-only helpers are retained in scratch/runs/20261104-mdk2-inspected-native-menu/ (ignored). Read the cause note published at 632639d35, the per-command route and the grenade notes before this attempt. The predecessor's original failure remains Enter on the unready 3,136-step copyright/title capture; an ACK did not establish New Game. Old step counts were not used as selection readiness in this attempt.

browser-driver.cjs now imports the tested selector's exported inputSnapshot(renderer), installs its standalone function with page.evaluateOnNewDocument, and spreads its result for sharedRenderer into the actual main state observation. The original renderer's absent/undefined inactive input becomes explicit null; the public getMouseButtons() supplies the zero mask. There is no renderer write or modification of old evidence. The actual driver state callback was tested in a VM with the installed adapter: absent, undefined and null inactive state normalize correctly; missing renderer, getter or valid mask fails.

Before Chrome, all 25 selector controls passed, including freshness, readiness, wrong selection, ACK, trusted event, aggregate 512-step / 64-command / 240-second bounds and stable post-close recovery negatives. The selector itself was copied unchanged and never invoked live. ACK mocks passed delayed startup, mismatched/failed/absent receipt rejection; transport mocks passed independent per-command deadlines and stopping after the first failure. Syntax, imports, all 432 remote runtime pins and an actual small-file upload/download recovery probe passed. Initial boat-dev lookup failed because it is not installed; the installed boat CLI was used before runtime, as in the predecessor.

One browser launched at 04:02:58.999Z. Initial startup observation had its own explicit bound: at most 4,096 steps and deadline 04:05:58.999Z. Each advance command requested 64 steps and waited for its matching ACK. Eight-step- command observation chunks were interleaved with captures. The startup bound was not the selector budget and did not authorize bypassing its protocol.

Every listed image was personally viewed by the worker:

Actual steps Capture Visible result
0 startup Registered MDK2 runtime log; no native scene.
512 / 1,024 startup-512, startup-1024 Gray application surface.
1,536 / 2,048 / 2,560 Corresponding startup-* Black native surface.
3,072 startup-3072 MDK2 copyright/title, legal overlay, closed icon wells; no New Game label.
3,584 startup-3584 Controller with red highlight and readable New Game, legal overlay absent. Viewed after the sequencing error; no activation followed.

The error was submitting another eight acknowledged 64-step startup advances in the same tool invocation displaying the 3,072 capture, before personally viewing the first native title and switching to the tested selector's inspected 16-step observation protocol. The extra 512 steps had no selector ledger and cannot be represented as compliance with its aggregate containment bounds. first-failure.json records this operator error. The already-submitted sequence finished; the only subsequent browser command was orderly quit. No source- backed Escape attempt, game click, Enter, movement, firing, reload, retry or second launch occurred. The final visible New Game label does not excuse the sequencing error or establish New Game selection. No held game key was created.

Validation establishes 65 matching successful ACKs, contiguous commands, every next command timestamp at or after the previous receipt, and 56 step commands / 3,584 completed steps / maximum 64. All eight captures match their recorded full PNG SHA-256. Every observed input queue is empty, active input null and mouse mask zero. The only two DOM input events are the initial trusted, balanced Launch-toolbar mousedown/up, with an empty target id; no game-control event occurred. An initial validation assertion incorrectly expected zero DOM events; its failure is preserved and the assertion was corrected to account for these toolbar events, without changing evidence. No browser first-failure artifact appeared; all 200 recorded HTTP pin checks matched.

Original execution source remains 1329b1ba6eb00144ff77d4cd92a8513d4fc953ad.

Retained module Full SHA-256
Canonical cffca869b541e6eff12ae3a6b74c4476977e0a7e23cd43d8ec5ba8faecd4bcee
Companion 580f49f3c78d26e6855efa0beb37046d03ea24690ff98cfcb3e94ce5e6cb1ee3

Both modules were referenced from the per-command run's retained paths, verified and staged once with the pinned Git blobs and original media. Media path/size/full SHA references are in staged-path-references.json; no media is stored in this run's evidence. Canonical compilation and instantiation with 2,261 exports were observed. Companion execution was not attempted. Exact checkout Git SHA, relevant binary diff, retained accepted-to-execution patch and owned new-file patches identify source; no source archive or rebuild was made. No guest write, cheat, console command or configuration edit occurred.

Chrome 8410 exited normally at 04:04:49.135Z; driver 8388 is absent and server port 39361 is closed, verified before stable inventory. The browser segment was 110.136 seconds, below 15 minutes. Two identical post-close directory hash passes passed. All 220 recovered unique files / 6,655,875 bytes, including both actual modules, match size and full SHA-256; all 432 runtime pins remain unchanged. Recovery finished before 04:49Z. No production build was required for a documentation-only commit and ignored run-only integration.

Hand back bx_wnjq4738 / RUNNING, API idle, health ok, hydrated / expiry 2026-11-04T04:59:39.962Z, with owned browser/helpers closed, to shutdownOwner=coordinator. No lifecycle action or public deployment ran. Next unresolved work is a separately authorized attempt that personally views each startup capture before submitting any dependent advance, switches to the tested selector on the first inspected title, and separately inspects selected New Game and highlighted Medium. This single task stops at the preserved error; no successor is launched.