NFS III CLI consumer contract: original launch refused, 2026-11-13
The required finite input contract is missing from the ordinary CLI queued
consumer. This run ends at that concrete prelaunch blocker. No new NFS III
original was launched, so renderer selection, race, native Pause/Continue and
reviewed driving remain unproved. Parent b8ac865a5 still supplies the exact
native catalogue → setup registration → chooser registry-transfer route;
nothing here replaces those native exports with fabricated installation state.
Evidence is scratch/runs/20261113-nfs3-local-consumer-native-race-0845/.
The actual production CLI, using Solitaire as an explicitly labelled consumer
fixture, demonstrated these results:
- Unknown action parsing returned
ok:false, without a DOWN call. - Frozen input returned
queued:truebefore any renderer input call. The fixture replaced the actualrenderer.inputQueuebeforestep:1. The scheduled realhandleKeyDown(65)then ran against the replacement queue and set the physical state to32768. Suppliedowner, expireddeadline:1,nonceandreviewedfields were ignored. - A second fixture request with that same known nonce executed another real DOWN against the replacement queue. This was a deliberate duplicate-input test with known replies, not a replay of an UNKNOWN transaction.
- The run-scoped read-only identity check refused an actual queue replacement and succeeded after restoration. That separate check does not guard production's later scheduled consumer.
- An independently armed Python subprocess issued a later ordinary UP after
a full 640×480 composed-canvas capture. The real DOWN returned at
1794533983335ms; capture brackets were1794533984851and1794533984867ms (1516–1532ms held). Both brackets had physical state32768, unchanged bound owners and an empty input queue. Real UP entered at1794533985285ms and returned with state zero; actualtakeInputobserved DOWN and UP. This proves this fixture's held-frame/release ordering only. The personally reviewed image is Solitaire, not NFS Pause or driving.
The negative qualification is about the requested envelope, not an advertised
CLI feature regression: these extra JSON fields are not supported by the
current interface. test/run.js parses cmd, enqueues plain actions and
later calls renderer input methods without carrying owner/deadline/intent
predicates into that consumption. lib/control-server.js times out replies
without cancelling pending actions. Consequently a read-only check followed
by a separate ordinary input request cannot establish the requested freshness,
one-use and UNKNOWN-terminal guarantees. Independent UP alone does not stop a
still-pending DOWN from applying later. This run did not manufacture a timeout
or claim to prove that last ordering dynamically.
The actual actor was the Node main-thread CLI interpreter, PID 972045,
kernel start ticks 258119213, HWND 0x10001. The observer retained actual
instance/exports, memory and shared buffer, renderer, window, canvas, input
queue, D3DIM encoder/module/adapter/control and command-buffer references.
No guest-main Worker or browser trusted-DOM identity was invented. The
production D3DIM adapter owns a real Node Worker but does not expose its
underlying thread ID; that limitation remains explicit. The fixture did not
exercise a game D3D device or claim encoder readiness as race rendering.
freeze.json rereads all 138 parent path/SHA pins unchanged, including the
133 media files and original EXE 0defab3e…548ad. The on-disk production
module remains 04c53419…f0a17; it was checked after this fixture, not
independently hashed inside the live instantiated module. Source identity is
base-commit.txt plus relevant-source.patch; new helpers and this note have
individual patches. Media are referenced by existing paths and hashes, with
no copy or source archive. Earlier browser 1500ms HELD proof is retained by
reference and hash; browser full22/1024×712 assumptions and old 200ms UNKNOWN
results were not transplanted or replayed.
The capacity dependency's actual recovery-result.json was observed after
controls were prepared. Its worker remained live and its recovery outcome is
recorded by path/SHA in dependency-observation.json; no independent cleanup
or repeated floor decision occurred here. Root0115/0120 recovery data and
Pawn/LMarbles work were untouched. Capacity recovery cannot satisfy the
missing consumer contract, so no original-effect admission was attempted.
The independent 120-second watchdog was observably armed before fixture
launch. The fixture acknowledged ordinary quit; interpreter, watchdog and
UP subprocess all have actual parent waits, exit0 and separate stdout/stderr
EOFs, with all three PIDs absent. Box ID NONE, state NOT_CREATED,
expiry null, stop receipt N/A. No production source changed, so a
locked build was not required. No agents, other workers or public deployment
were launched.
Next work must supply and qualify a CLI consumer adapter or production action that carries the actual owner references, deadline and consumed reviewed intent through the final scheduled action and queue delivery, cancels stale pending DOWNs on release/error/UNKNOWN, and gives independent UP its own bound receipt. Use focused real-consumer tests for changed ownership, expiry, duplicate intent, error and UP-before-pending-DOWN ordering. Only after that passes may a fresh capacity decision admit the unchanged native registry route, personally reviewed renderer selection/race, two settled native Pause/Continue frames with stable world timer, and one finite reviewed drive with held frame before UP/Escape.