TR2 completed-owning-frame publication pacing — 2026-11-08
Current pacing remains unresolved. Browser qualification stops at a concrete
production journal-read failure; no original-game run or box was started.
SurfaceVersionJournal.read() passes a shared Uint8Array.subarray() to
TextDecoder.decode(). Chrome 151 rejects it with TypeError: The provided ArrayBufferView value must not be shared. The observation catches this failure
during ordinary composition and poisons its epoch. A negative/empty Join is not
0 FPS. Logical/simulationFPS, publication pacing and full observer cost remain
unknown.
Evidence is in scratch/runs/20261108-tr2-qualified-publication-pacing/.
The tested source is published
d7c1f187fe4eaaa0fde0879a9684a7c75599526a, including Join
663f1c1399fc270d6c29a05ea077742cdc63e299. source-audit.json checks 1,212
published source/tool files against the working tree; all match. The relevant
binary diff is empty. Peer test and operational edits were excluded. Git plus
the retained helper patches preserves source identity; there is no source/media
archive. No emulator rendering divergence was demonstrated or repaired.
The installed Join note/source, retained final protocol/build/publication results, default OK activation and frame identity contract were reviewed. The prerequisite really is published and reports its locked build and thirteen focused passes. Its Node production association regression was rerun here and passes again: one actual owed publication, five installed-host publications, eight discard rows, palette/ABA/coalescing/fence/delayed/duplicate/retirement/ timeout controls. These Node results do not establish browser compatibility. Accepted native Enter, Venice translation and heading were not repeated.
The bounded API admits 1–4096 records, each with a 64 KiB shared slot. Records are never reused or drained. Reaching the record limit deactivates the producer observation and poisons its writer epoch. Host arming refuses an existing Join; retirement does not create a supported same-process rearm path. A future run must allocate enough capacity for its entire observation and retain raw records, coverage state and the discard ledger. A failed or incomplete read cannot silently become an empty successful window.
worker-join.json exercises actual production guest and render Node Workers,
the production Encoder, native queued flip/fence, broker, canonical decoder and
raster-canvas compositor. Fixture-only exports expose existing WAT routines;
no game or GPU throughput is involved. All four same-entry queue/fence attempts
complete ordinary pixel swaps, but all four source records are rejected as
overlapping-or-poisoned-source-writer. The guest's WorkerLink writer ticket
spans its fence wait, while source sealing requires the render Worker to be the
only writer. This conservatively unsupported execution shape needs separate
integration work; it does not prove every TR2 slice has that shape. The Node
fixture retains four rejected records, four ledger dispositions, zero qualified
rows, zero duplicate completions and capacity 16 without exhaustion.
browser-fixture.json uses real browser guest/render Workers and the installed
WineAssembly.armD3dimPublicationObservation({maxRecords:16}) method in a fixture
host shell. Arming succeeds. The endpoint requests {api:'legacy',backend:'webgl'};
no GPU draw or active-game GL qualification is claimed. Four native queue/fence
completions retain exact owner/request identities and rejection reasons. The
actual canvas alternates red values 17 and 34. Canonical uploads carry the owed
request, but there are no qualified cache/publication rows. The first ordinary
repaint poisons observation; the final direct journal inspection exposes the
shared-view TextDecoder exception at lib/render-worker.js:298. The raw four
rejections survive in the producer receipt, while the page discard ledger is
empty because its journal reads fail. That missing accounting is a failed
coverage guard, not evidence of zero dropped/discarded frames. No events were
deduplicated into another timing window.
An isolated browser control (journal-browser-control.json) reproduces the
failure with one committed 66-byte journal entry, without Workers or a live
writer. Decoding an ordinary copied byte view returns the exact expected record.
This diagnostic control does not patch the production reader or certify a fix.
Its fetched host and render-worker source hashes match the published blobs.
First source-completion clocks and callback-receipt clocks remain separate in
the raw records/events, including their performance time origins. No qualified
page-publication clock exists in this browser run. No cross-domain subtraction,
publication intervals, three ten-second windows, median/p95 game intervals or
rate is reported. The production records' qualified:false/null consumed
version are retained.
Matched import-cost measurements address part of the previously unqualified
default impact. import-cost.json retains 132 raw timed trials: eleven
alternating symmetric groups per import, 200,000 actual WASM calls per trial,
with published pre-Join baseline shape (279ede77e), new-unarmed and new-armed
arms. Each arm calls the same production math_sin or get_ticks implementation
in the same JavaScript realm. Median group-paired added costs are approximately
18.2 ns/call unarmed versus baseline and 177.1 ns/call armed versus unarmed for
math_sin; corresponding get_ticks deltas are 9.1 and 158.0 ns/call. Raw ranges
are wide, including an armed get_ticks outlier above 1 microsecond/call.
These are Node import-wrapper seam costs, not gameplay slowdown. They do not measure the complete default module facade, arming, journal/source proof, canonical/page publication, sampler or serialization overhead. No matched same-living-scene observer/control intervals were reached, so full observer cost and uninstrumented game performance remain unqualified. The initial cross-realm benchmark is explicitly invalid and retained separately; its larger numbers are withdrawn. The first Node fixture failure was a missing Node Encoder dependency; its error is preserved, and the corrected fixture preloads the unchanged production dependency as the browser normally does.
The next production work is a browser-safe immutable journal read, with a real browser regression and measured copy/decode cost, followed by qualification of the guest-fence writer boundary without weakening coverage. Repeat full identity/cost qualification before allocating the one watched original. Ordinary rendering must remain intact and raw rejected/discarded/dropped/coalesced counts must remain reviewable. This report stops at the integration blocker rather than treating Node positives as browser acceptance.
Both local fixture Worker owners retire normally. Both local browsers close
normally with Chrome exit 0/signal null and retained target-destruction and
disconnect events; no unexpected original-browser loss occurred because no
original was launched. Local runtimes are serialized with
flock /tmp/wine-runtime.lock; available disk remains above 5 GB. No console
input, private input handler, guest/game patch, forced scene, clock/seed/binding
change, deployment, other-agent pane operation or subworker was used.
Box ID/state: not allocated. Expiry/stop receipt: not applicable. No old box was resumed, stopped or deleted. The independent deadline/retrieval/normal stop protocol was not entered. Full27, fullFPS, backend parity and audio remain open, as does simulationFPS without a separate simulation association.