Boids natural display completion

The closed original from 7c84a0ae0 has no complete capture. Retained logs identify neither the first blocked publication join nor page/controller latency. Its readiness counters do not establish current pixels or motion, and its endpoint remains closed.

The shared native/original capture driver now supports an opt-in, bounded scalar trace. It records page entry, public capture entry/return, ticket transitions and encoding completion, together with host damage flags, executing slice depth, paint depth, repaint flags, publication identity/time, weak object identities for display and producer canvases, and up to eight source-writer epoch snapshots. Console delivery times and controller request/completion times distinguish page queueing from publication and reply delivery. Trace failures cannot alter capture returns or refusals. It performs no guest call, readback, flush, repaint, stepping, input or scheduling change; the 3000 ms bound is unchanged.

Authored native Device2 fixtures cover ordinary rendering, a single first Flip followed by native Sleep, and an 1800 ms native delay before the first Flip. Their selected complete source/output captures match independent predefined literal blue/white frames over all 307,200 source and 1,185,280 output pixels. Both already-displayed and later natural publication paths complete within the unchanged bound. The delayed-first fixture's selected capture occurs after that delay; it does not identify the original's cause. The first old Device2 oracle assumed only half-integer raster centres; the existing independent ownership oracle also checks the predefined integer-centre frame, which matches these captures exactly.

The driver tests run with tracing on and off, including absent, wrong and replaced ownership, failed/UNKNOWN/pending tickets, corrupt or late data, bounded trace storage, observer failure and sticky refusal. Native controls additionally replace the source device and instance. Real renderWorker WebGL context loss causes the next natural record to fail with no later accepted frame. A separate native run retains an actual 640x480 GPU ImageBitmap callback, awaits the real bridge close, then delivers that callback after epoch 1 changes to 2: the bitmap closes and the presentation count does not advance. All native/browser/watchdog waits are zero with both output streams at EOF and no remaining token-owned processes. The trace cost measured in the context run is 1.97 ms total over six samples, maximum 0.545 ms, including console emission. This is a measured run, not a universal overhead bound. Trace-on and trace-off native captures take 549 and 503 ms end to end respectively; those individual runs do not establish a causal slowdown.

Evidence is in scratch/runs/20261117-boids-natural-display-completion/. Source recovery uses exact Git commits and binary patches; existing media paths/hashes are referenced without copying media or archiving source trees. The mandatory locked canonical build passes after the concurrent NE worker updates its ownership review. Trace commit 4f0a33f3682cebc517616660f47cdd5dcd6615d5 was strictly pulled and pushed. A Git-only published native run passes full literal pixels and all eleven native ownership/failure controls. Its trace costs 5.12 ms total over seven samples, maximum 2.085 ms.

Exactly one authorized diagnostic original then runs on that current published commit, without overlays or an older core. Its 1,504 Git source pins, all 38 unchanged registered media paths, and actual page/guest/render Worker debugger module hashes are verified. The module is a105f6ce02932b3edebe9376a8b5eb4658f2159c52bb3cc76ae5ac4eadb05b0c. Preparation remains 240/260 seconds and runtime/watchdog 300/320 seconds. Defaults, clocks, EXE and media bytes remain unchanged; no input, unpause, replay or second original occurs.

All three complete current-display captures pass the unchanged 3000 ms controller limit, at publications 58, 122 and 244. Page entry through encoding takes 1263, 1075 and 321 ms; readback takes 804, 496 and 148 ms. The first and third requests read an exactly validated already-displayed publication. The second request begins at publication 121 with an executing slice, dirty host and scheduled repaint; it remains pending until ordinary publication 122, 321 ms after the request. It never accepts the older pending frame. The bounded trace retains 31 rows, with stable source/output/producer identities and no source-writer join or context error. These successful observations do not retrospectively identify the old 7c84 timeout's cause or establish a semantic timeout repair.

All three outputs were personally viewed. The first contains cyan terrain edges; the later two show small then larger colored birds at different positions and orientations. Those latter complete displayed frames are 16,413 ms apart and differ in 16,065 of 307,200 source pixels. This establishes actual passive original flock motion through guest Worker and WebGL render Worker. Counters are supplementary diagnostics, not its pixel proof. It does not establish smooth FPS, native Windows parity, physical GPU use or full 27-title acceptance. Only named selected distant/near flock source/output screenshots and one authored source/output pair remain; there is no frame archive or capture bundle.

The browser's actual close is zero and the terminal record says observed. The outer execution returns 143 before native/watchdog join and EOF receipts are written. All four recorded original supervisor/runtime/watchdog/browser PIDs are absent, but their absence is not a substitute for missing waits. The cause of that outer termination is unknown and remains a separate concrete shutdown boundary. No native guest Quit is claimed. The closed original is not accessed again. No box is created: ID and expiry null, state NOT_CREATED, stop receipt N/A.