MCM Worker-service prerequisite: current Boids protocol differs
No production latency repair was established, and no original MCM was launched.
The current small production Boids route uses GPU stream v2 d3dim-gps,
whereas the November 7 latency control used the retired legacy D3DIM WebGL
protocol. The old readiness field, guestWorker.d3dStats.renderEndpointId, is
absent even while the current guest and renderer make visible progress.
Changing that diagnostic readiness predicate is not a runtime fix or backend
credit. MCM remains a provisional historical27 D3DRM candidate; throttle,
combined turn at speed and ordinaryQuit remain unproved.
Evidence: scratch/runs/20261120-mcm-worker-service-latency-quarry-control/.
analysis.json correlates requester bounds with bounded read-only spans;
current-gps/{capture,spans,evaluations,workers,roles}.json (evaluations is
.jsonl) retain the raw observations. current-gps/boids.png was personally
reviewed: dark scene, thin blue wire outlines and one magenta triangular shape.
The earlier startup.png shows multiple colored shapes in the same Boids scene.
Neither is an MCM capture or an input consequence.
Actual correlation
One current-protocol control completed full unchanged capture-chain validation.
Before-capture live validation took 278.37 ms; post-screenshot live validation
took 31.92 ms. Its first renderer identity request lasted 184 ms by requester
wall bounds. Two natural exclusive renderer drains overlap about 182.6 ms of
that request: one finishing at 04:39:09.582Z, followed immediately by a 155.92-ms
drain ending at 04:39:09.739Z. In the guest, slice 25's ex.run lasted 184.83 ms,
including a 166.44-ms synchronous GPU doorbell wait for ring 134006592. The
renderer drain completed immediately before that wait returned. This directly
locates a real service interval inside ordinary renderer work and a guest
completion wait, but does not explain the historical 3.5–5.4-second samples.
The guest observer retained 198 rows and the renderer 124, with zero drops. There were 16 completed guest runs, 31 completed renderer drains, and no local sleep/message-wait span during this short window. Publication fences had a maximum observed duration of 0.255 ms. The observation covers only a short startup window; it does not rule out later latency. There was no admission callback, no input, no artificial busy task, no promise-flag comparison and no freshness/deadline adjustment.
Instrumentation was confined to run-only HTTP transforms. It records bounded
start/end timestamps around the original calls, retaining their arguments,
receiver, return values, exceptions and existing async structure. It adds no
yield, guest call, GL synchronization or fence operation. The synchronous
Atomics.wait wrapper calls the original once. Instrumented scripts were syntax
checked before launch. No production source was edited or claimed for a causal
fix. Observer overhead was not separately measured, so these are attribution
receipts, not performance results.
There are two measurement limits: raw span times use browser
performance.timeOrigin + performance.now, while requester bounds use
millisecond Date.now, without a clock calibration claim. The optional drain
end read/write metadata is empty because those properties are unavailable on
the Ring object; exact done_seq and read/write positions were not captured.
Actual span completion and guest wait return are retained, not invented fence
sequence values. The old latency cause and any safe yield repair remain open.
Identity and scope closure
The two actual Workers each exposed one executing module, recovered once per
incarnation, SHA-256
fab7c60e39c079c2972458d30c3a36143c6afd68c80f76d97e9e4a0c2f789d92
(2,420,759 bytes). It is not the old e415 module. The full chain rechecked
live page/guest/renderer membership and identity. Only immutable recovered CDP
module bytes were reused. The renderer endpoint's actual options identify
d3dim-gps with requested backend webgl; actual GL renderer/executor identity
is unknown, so no exercised WebGL/D3D backend pair is credited.
The first scope ended after its obsolete readiness predicate timed out. A
bounded read-only diagnostic within that live scope had already captured the
working scene and advancing guest doorbell counters. That scope was closed
permanently. The second scope used the actual manager port's d3dim-gps option
for readiness and completed the observations above. Its actual identity and
module inventory are retained. Neither closed scope was resumed.
The requested parent retained-timing-analysis.json and
comparison2/comparison.json are missing from this checkout and were not found
by recursive filename search in scratch. The retained notes, probe/probe.json,
comparison2.log and focused4/controls.json were consulted unchanged. Their
availability/hashes are recorded in retained-controls.json; no missing JSON
was reconstructed or treated as present. The negative await comparison was not
repeated.
All three finite native supervisors report no leaks; direct browser, driver,
supervisor, watchdog and board-watcher waits include honest pipe EOF receipts.
The extra read-only page diagnostic had its own finite supervisor and direct
wait. Tracked PIDs are nonlive, and the short /dev/shm/mcms20 TMPDIR/profile was
removed. No enclosing tool-pipe EOF is inferred. During the two runtime scopes,
independent 100-ms disk samples stayed above 1.5 GB; later sealing observed
1,498,259,456 bytes free, after both scopes closed. No further normal runtime or
build was launched. Only the task's identical duplicate recovered module was
removed after byte verification; its recovery reference points to the retained
unique module. No foreign cleanup occurred.
No production code changed, so no build or unchanged test campaign was run. Run-local syntax, span pairing/correlation, module identity, disk, deduplication and custody checks pass. Native Windows parity is unknown; no Wine was used. No test box was created: box ID null, state NOT_CREATED, expiry null, stop receipt null. No deployment or backend promotion occurred.