Q3 fixture blur matches its intended 8×8 rendering
The retained image matches the generated fixture geometry and default smoothed publication. No causal production defect was demonstrated. Production code, shared providers and review guards remain unchanged. No box was allocated and no fixture input or original arena gameplay was attempted.
Evidence: scratch/runs/20261109-q3-fixture-visual-cause-2348/.
The exact starting Git commit is in git-base.txt; the small owned patch is
owned.patch. Git remains the source archive. Historical sources/media are
referenced by existing paths and authenticated hashes, without source archives.
The separately model-inspected original is
scratch/runs/20261109-q3-readiness-wire-cause-2335/fixture-startup.png:
9,633 bytes, SHA-256
69ebb61ccc6f16cd1c1f44c6d29bbe9b91f99189ca8b1ebba753647a80b275e2.
Its dimensions are 945×889. The white page surrounds a 320×200 black/red canvas
at image coordinates [8,8) through [328,208). The image contains the expected
coarse red triangular silhouette, with blurred boundaries between large steps.
It remains the image previously refused for continuation; this investigation
does not publish a review gate or reverse that stop.
The exact current generated fixture consumer was composed through
q3-current-foreign-closure-v1.cjs, including startup and physical-input
integration. Its fixture-source SHA-256 is
ebaaa700cc0145e7df997600aafdb4c553c660dc90df1777be3c1dbfbd335384.
Executing that fixture function with controlled WASM/Worker services records
its actual public x86 thunk calls: the client/window and GL viewport are 8×8;
GL_TRIANGLES receives (-1,-1,0), (1,-1,0), (0,1,0), in solid red over black.
glFinish, SwapBuffers, bridge drain and a Worker slice follow. The independent
pixel-center coverage calculation produces these top-down source rows:
00000000
00011000
00011000
00111100
00111100
01111110
01111110
11111111
The fixture creates screen with 320×200 backing dimensions, no CSS size or
transform, and draws the published source layer with
drawImage(layer.canvas,0,0,320,200). It does not disable default canvas image
smoothing. The actual host bridge _presentFrame sets the source canvas to the
frame dimensions, replaces its pixels, publishes the layer onto the owning
window, then calls the fixture callback. Controls execute that unchanged host
publication method and verify an 8×8 source, exact target canvas identity,
320×200 destination, and smoothing enabled. Wrong HWND, wrong frame token and
a same-looking replacement window each prevent publication.
Independent bilinear reconstruction enlarges each source pixel by 40 horizontally and 25 vertically. Of all 64,000 canvas pixels, 63,976 match exactly and 24 differ by one red-channel value. Every green/blue value is zero and every alpha is 255. All 776,105 pixels outside the canvas are opaque white. This is quantitative evidence for expected low-resolution steps and smoothing, rather than a blur attributed to failed publication or a different screenshot target. Negative controls reject inverted triangle rows, a corrupted green channel and a shifted canvas boundary. Native computed CSS, devicePixelRatio and intermediate layer pixels were not separately retained: DPR=1 is consistent with these pixels, not an independently observed native value. There is no forced scale/viewport command on this fixture route; it captures the page rather than a canvas clip.
Native provenance was independently reread and SHA/size verified against the
previous settled inventory. native-provenance.json preserves the small selected
records plus original artifact paths/hashes. Readiness producer records keep the
same render target, context, HWND 65537 and endpoint 1 while draws, presented and
writeSeq advance from 1 to 2 and vertices from 3 to 6. Capture request 100 uses
page target 864D14FBC3193C9B07B1102BB14A732C and session
DB4E2E0B64496DB11854531A4BE0D26E; consumer delivery matches that session.
The capture receipt links the PNG hash to unchanged page-canvas, guest and render
identity tokens and the selected CPU/render module SHA
3e491c0840bb7771328738a0d160b4bbaa3d7ab72074cbd38a7b41e5d7e8ec18.
These counters are publication observations, not intermediate source pixels.
tools/q3-fixture-visual.test.cjs also compiles and executes the actual composed
bound capture consumer, with controlled observation/CDP services. It verifies
one page-session PNG screenshot with unchanged options, returned PNG bytes/hash
and canvas identity, then independently rejects canvas replacement during
capture. No input is emitted. These controls do not execute a native GPU or
replace the separate model review. The existing service's seven review and
eight native-wire revocation controls also pass; no guards were relaxed.
The locked build passes (see build.log and build-wait.json).
The first concrete blocker is the earlier fixture presentation refusal combined with the absence of a demonstrated rendering defect supporting a corrected fixture. The pixels are accounted for by intentional 8×8 geometry and smoothing. Increasing its resolution or changing its display filter would change that fixture contract; neither is a demonstrated repair. Therefore the conditional corrected-fixture allocation prerequisite is false. Missing native computed DPR/CSS and intermediate frame pixels remain explicit limits, not reasons to repeat live timeout work. A future task would need an explicitly specified visually reviewable fixture contract and qualification before gameplay. Request 347 was not reinvestigated. No movement, fire or alive-arena claim.
No runtime was created: box ID/state/expiry/watchdog/stop receipt are all N/A. The prior archived box was left untouched. This single task ends after publishing the tested controls and this finding, and releasing all owned paths.