MDK2: observe demand before file-request publication
The first early-attachment dependencies are repaired offline. The actual command
producer attaches at the first reviewed, input-free construction quantum, including
a checkpoint where no file request exists. It no longer executes that quantum
unobserved while waiting for an already active level1 request. The observer reads
the genuine demand argument from ESI at the original PUSH/CALL boundary, before
the file-request global is necessarily published. No original was launched.
Input is 1b1eada7d4885008e51a379142f22391c1fd0e7b and
mdk2-continuous-construction-onward-20261104.md. Evidence is
scratch/runs/20261104-mdk2-demand-entry-continuity/ (RUN). Old seals are unchanged.
The command fixture executes the actual composed driver functions. With an empty request sidecar, the input producer executes one quantum with zero attachments; the repair attaches once and checkpoints once. These are explicit protocol fixtures, not original pixels. Historical producer comparisons change only the observer-hash admission constant so both producers use the same current observer.
Both retained modules execute the original 0x41fce5 PUSH ESI and CALL to
0x44fc40 through the pinned shipped Worker queue. The paired fixtures begin with
empty request/active globals and a valid type10 level1 argument. The previous
observer rejects wrong demand request; the repair records that argument and
the independently observed empty global. A wrong ESI still rejects. Existing
nonempty global/active identities must match the argument. Type, name and size
are checked by readback; no history is supplied by a snapshot.
The same observer then spans checkpoints and the original finalizer, completion consumption, poll, demand-return and buffer-entry fragments. There is exactly one observed demand-call record, before finalization. Full-memory endpoint hashes, registers, lazy flags, returned command sequences and the original trap match the control on both modules. Fragment repositioning and request publication are explicit synthetic fixture setup shared with the control; they are not a claim that a fresh game completed this path continuously. Late attachment to the same fixture still rejects the demand return with no demand-call record invented.
Request publication across a construction quantum is now retained as two request identities. It supplies neither a call event nor an output-progress delta across different requests. An unchanged request still requires unchanged buffer, size and name. File callback decoding remains evidence interpretation, rather than a prerequisite that would delay attachment to the main Worker.
The original observation caps are unchanged: 10000 ms wall time including reviews, 100000 boundaries and 512 records. Checkpoints retain the observer, deadline, counts and native call history. Existing owner, first-script, frozen queue, review/ACK, changed-register, scene, expiry, unknown-prefix and overflow guards remain. No guest code, state, configuration or pacing changed.
The first remaining original dependency is earlier than this reviewed construction
checkpoint. The retained selector executes Medium press and release quanta while
the approved scene is still difficulty, before it returns the next screenshot
for personal review. The actual composed medium-unobserved fixture executes a
16-slice command in that scene with zero attachments. Its earlier demand history
therefore cannot be promised by the repaired construction attachment. The code
evidence establishes a coverage gap; it does not establish where a historical
original's demand call occurred. Also, the construction controller can choose 64
slices while this observer command contract accepts 16. Neither dependency is
solved by extending the observer deadline or ignoring commands.
The task stops offline as authorized. The actual end-to-end capability gate has not passed, so there is no known-failing original, repeat diagnostic, title input, loading capture, Lua/Level claim or gameplay claim. The next repair must cover Medium's earliest execution under its existing trusted-input checks, then prove the entire command/review path fits the same wall/count/record caps. It must also reconcile the actual controller's loading quantum with the trace contract before spending a fresh original. The prior runtime assignment remains prior admission; no renewed runtime-release document was issued for this task.
Stronghold's explicit 20:00:07Z release was observed. This worker acquired no runtime flock and started no Chrome or game server. Its offline local release was posted at 20:10:41Z for GLTron. Box ID/state/expiry are NA; the coordinator owns future boxes and shutdown. RUN retains exact Git SHA, relevant binary diffs and this new-file patch, helper/Git/runtime pin checks, test logs and existing module paths/hashes. No source archive, media/module copy, box operation, npm install or subworker was used.
Final validation passed: 115 integration tests, 35 local lifecycle/caller tests, standalone both-module full-state/command/trap parity and the locked full build. All 111 retained helpers, four Git dependencies, 432 filesystem/HTTP entries and both retained modules verified. The predecessor seal still covers 1140 unchanged files / 2096728 bytes. The local suite's owned Node cleanup fixture passed; it does not launch diagnostic Chrome.