MDK2 Medium through loading: composed producer repairs, original withheld
Input is 86dc176a3b25a1eda252d1f4b83a1e30a8bb417b and
mdk2-demand-entry-continuity-20261104.md. Evidence is
scratch/runs/20261104-mdk2-medium-through-loading/ (RUN). No original, Chrome,
game HTTP server, ordinary game input or runtime flock was started. No fresh
gameplay, Lua execution, or genuine Level creation is claimed.
The actual producer sequence was inspected through the Python operator, author, next and control entry points, measured selector, pending controller, construction controller, loading controller, original browser command function, and owning observer. Medium runs repeated 16-slice press polls, followed by a 32-slice release and possible further 16-slice release polls before its capture. Pending observations cost 16 slices. Construction selects 16 with Level absent and 64 with the original Level1 table identities published. Loading sends eight 64-slice commands under one reviewed capture, then requests another review. Movie cancel uses 16/32; post-cancel observation uses 64. Every original controller's native, visual, command, ACK, ledger and failure gates remain.
There were four composed-path defects. First, observation began only after Medium's execution. The driver now binds the trusted keyboard readback saved by the original command path and attaches before the first difficulty execution. It requires the reviewed Medium scene, original NewGame/Medium ledger, trusted screen edge, balanced prior history and unchanged event prefix. Pending input must be an exact suffix of that acknowledged renderer queue; foreign active or queued input rejects before execution. Original nativeReady still runs before keyboard dispatch. Subsequent polls compare against the last observed frozen state, rather than the older screenshot's step count.
Second, the observer command adapter asserted exactly 16 while real controllers send 32 and 64. It now validates the producer's exact 16/32/64 request and completion without dividing, skipping or replaying commands. The same observed history continues through the original scene validators for Medium fade, pending, construction, loading, comic and cinematic reviews. The Worker retains the same observer, deadline, records and boundary count throughout. An idle-loop health check now closes the original normally on a terminal/expired observer; the command path checks it before further mutations and closes on its first command failure. Key release and quit retain the cleanup path.
Third, next.py already routed loading, cancel, post and cinematic continuations, but both operator.py and author.py rejected those modes. Their generated run-specific candidate sources now admit the existing routes. Cancel uses the input-review prefix, matching next.py's existing native input preflight. The sealed helpers are unchanged. Local assignment identity and absolute cutoffs are renewed to this task and 23:15/23:18Z; 720-second browser, 60-second bootstrap, 40-second command and 180-second cleanup caps remain.
Fourth, an observer failure disabled tracing but returned to the queued guest slice, permitting an unobserved suffix before the command could reject. The owning adapter now throws at that first rejected boundary, including a paused boundary or a count limit. It preserves the original import's own exception. Both modules execute the original finalizer POP ESI/RET with a deliberately missing finalizer prefix: the old adapter reaches EIP 0x1800000 / ESP 0x74ff008; the repair stops at EIP 0x44f515 / ESP 0x74ff000. The same single queued command is delivered. Guest bytes and unaffected registers/flags match; complete memory hashes are retained, with different decoder-cache endpoints explicitly expected. Normal-path full-memory/register/flags/command/trap parity remains separate.
The actual selector/controllers/driver/inspection/ACK protocol test uses explicit synthetic native-state and transport responses. It exercises repeated Medium press and release polls, fade, pending, absent Level, published Level and the eight-command loading branch. It passes 16 quanta with one start, 15 resumes and 16 checkpoints. The input producer leaves Medium unobserved and rejects the actual 64-slice construction command; the repair passes the complete fixture. Trusted-edge, extra queue, foreign active input, changed history and expired review negatives reject before the next execution. Python tests execute the actual operator -> author -> next code for all 11 modes and verify the selected controller/action and cancel preflight. The old author/operator fail that test. These fixtures contain no original pixels and cannot establish human latency.
The retained normal and compat modules separately execute the genuine original queued PUSH/CALL, finalizer, completion consume, poll, demand return and buffer entry with the current preload. Full-memory endpoint hashes, registers, lazy flags, command sequence and original trap are compared with control. The original UNKNOWN poll-prefix, missing-demand-history, wrong ESI, module, owner, thread, frozen state, changed registers, boundary/record overflow and deadline negatives remain. Synthetic repositioning between native fragments remains explicit fixture setup. It supplies no original gameplay history.
The complete fresh-original capability gate has not passed. The sealed original commands establish a concrete timing failure: first Medium step 91 occurred at 14:53:28.000Z; first pending step 98 at 14:53:48.007Z, a 20007ms interval. Both owning-module observer tests replay that interval with an explicit test clock. They preserve the production 10000ms wall limit and reject at the first checkpoint continuation. A faster fake transport is not evidence that personal VIEW/review and genuine native execution fit this deadline. The full real controller/selector/owning-observer path with those human reviews still needs a measured passing rehearsal; native fragment parity and fast protocol tests are separate evidence. No fresh original is spent on this known failing historical review schedule. The caps remain 10000ms WALL / 100000 boundaries / 512 records, including review time, with no restart or deadline reset.
Stronghold's explicit 20:20:46Z release was observed. This worker released its unused local slot at 20:29:22Z for the GLTron successor. All activity was offline; box ID/state/expiry are NA and the coordinator owns future boxes and shutdown. RUN records exact Git SHA, relevant binary diff and this new-file patch, plus existing module/media paths and hashes. No source snapshot, tar, module/media copy, boat, remote box operation, npm install or subworker was used. Prior seals are preserved. The remaining work is a measured complete personal-review/native rehearsal within the unchanged limits before a separately assigned original.
Validation passed: 132 integration checks, 35 local lifecycle/caller checks,
final producer/containment/input-scene checks, standalone normal/compat native
parity, and flock /tmp/wine-build.lock bash tools/build.sh. The 111 helper
closure, four Git dependencies, 432 filesystem/HTTP entries and both retained
module hashes verified. Current build modules also match the standalone parity
hashes. The input run's 888 sealed files / 1650054 bytes remain unchanged. No
owned Node fixture, Chrome, HTTP server, port or runtime flock remains active;
the short-lived board watcher was already absent and received no signal.