Registered Globe: original draw/pixel band oracle, November 7

The selected black-band boundaries are expected texture content in one naturally completed registered dx_globe draw. An independent scalar calculation and a separate 60-digit Decimal calculation reproduce 12 selected target pixels exactly, on both sides of six black/nonblack transitions. No causal emulator mismatch was demonstrated and no production repair was made. This resolves the selected band pixels; it does not certify every pixel, seam, or rendering route.

Evidence is in scratch/runs/20261107-globe-original-band-oracle/. Startup/menu/wireframe coverage was not repeated as the deliverable. The registered executable, texture, ordinary scene configuration and 56-vertex / 108-face sphere3.x were preserved. The 296-vertex archive mesh was not used. Mesh SHA-256 is bb8bd85a4f13ef27777690295264f7f2475b3b4bb4e936e3b3394ee08c05360f; executable SHA-256 is ef0f198d8781de117100cbb77dd7f5b3e16990448d5a1a081c570359ceaf377b; tex1.ppm SHA-256 is 24ead200deb44edec992a8614cea9707bb1e2cd5035a8b7b9bd16c6aeff2182a.

One draw and its pixels

recovered/original/passive-band-7E96B3B492C477714506B8CD9BC0EE80.json contains a bounded owning-render-Worker observation. Its complete event order is natural clear → natural upload → submission 3 → successful completion 4 → natural readback completion 5. There is exactly one draw in that interval, 108 triangle-list primitives / 324 submitted TL vertices, with no excluded draw. The observer stops copying after this first readback. No diagnostic readPixels, fence, flush, guest call, guest-memory edit, camera/clock freeze, or input replay was introduced. Existing diagnostic metadata attaches the descriptor to submitted vertices; its before/after callbacks make no rendering requests.

The observed target is rt=131805376, DIB 471576576, 292×254, 32 bpp, format 6, pitch 1168. Submitted viewport is (0,0,292,254), depth range [0,1], no scissor. The natural readback rectangle is [58,41,233,216); its raw bottom-up RGBA bytes and corresponding native DIB bytes are retained. Thus the oracle compares the completed draw's target pixels, rather than an asynchronous desktop screenshot. natural-target.png visualizes these exact readback bytes; the later inspected recovered/original/startup.png is a different camera phase.

Effective state is Gouraud prelit diffuse, texture×diffuse RGB, texture alpha, no additional lighting/fog/specular/alpha test/color key, depth enabled and written with LESSEQUAL, no culling, source ONE / destination ZERO blending. The descriptor records WRAPU=1, repeat U/V, point min/mag and no mip filter. The captured submitted coordinates already include the per-triangle U fixup; this is not a second raw-before-fixup UV acceptance test. Full descriptor, viewport, fixed-function state, vertex colors/RHW/UVs and exact texture bytes remain available in the capture.

The current texture is 256×256 RGB565, pitch 512. Independently parsing the original P6 and quantizing it gives 0/65,536 native-texel mismatches; bit-replicated expansion gives 0/65,536 decoded-texel mismatches. This checks this draw's texture, rather than inheriting the October result as pixel proof.

Independent calculation

oracle.py uses only Python's standard library, with no emulator imports or production sampler calls. It computes triangle barycentric coordinates, selects the nearest depth plane, interpolates UV and diffuse with reciprocal-W weights, applies point sampling with repeat addressing, expands the exact RGB565 texel, modulates diffuse and rounds the 32-bit target's RGB channels. It accounts for the observed fixed POSITIONT path's 1/16-pixel tie bias. Selected samples are at least 2.5 pixels from triangle edges and 0.18 texels from texel boundaries; removing that bias selects the same texels for every retained sample.

Of 5,347 robust interior pixels, 5,246 match RGB exactly and all 5,347 are within one channel value of the scalar result. The maximum error is 1; these 101 one-value differences are retained, not called exact parity or used to justify a speculative fix. Six transition pairs were selected geometrically from texture changes, without filtering by pixel agreement:

Target pixels, left / right Triangle, zero-based Source texels, left / right Predicted = observed RGB, left / right
(214,138) / (215,138) 45 (122,124) / (123,125) (107,97,38) / (0,0,0)
(174,53) / (176,53) 78 (127,13) / (132,13) (0,0,0) / (42,45,53)
(205,90) / (208,90) 39 (122,62) / (125,63) (73,85,100) / (0,0,0)
(193,150) / (197,150) 42 (110,133) / (112,134) (0,0,0) / (141,122,42)
(205,77) / (209,77) 38 (128,46) / (133,47) (0,0,0) / (69,74,88)
(150,183) / (154,183) 4 (91,173) / (93,172) (101,88,32) / (0,0,0)

check-selected.py separately decodes the raw submitted bytes and re-evaluates all covering triangles with 60-digit Decimal arithmetic. It independently confirms the winning triangles, texels, and all 12 exact RGB results. oracle.json, interior-pixels.json, and selected-decimal-check.json preserve weights, depths, next intersecting depths, UVs, native texels and channel values.

An affine-interpolation counterfactual changes predicted RGB at 9/12 selected pixels. The perspective calculation matches all twelve. Every selected UV is inside [0,1), so repeat and clamp agree there: these samples establish expected texture content and perspective interpolation, not general seam/addressing correctness. No native Windows reference, software-route comparison, full scene parity, FPS, audio, historical27 membership or corpus completion follows.

Qualification, source, and lifecycle

Before allocating, the actual selected deferred observer attachment passed a manifest-restricted positive case and missing-observer negative case. Fresh Chromium qualification exercised actual production guest/render Workers, identity/module/freshness rejection controls, trusted input independent-release controls, an exact quit ACK, and passive submission/natural-readback hooks. All owner/module/freshness, ACK and release helpers were inherited unchanged. The generated fixture's ordinary Lock was allowed to proceed; this was a local control, not original guest input. The real driver accepted the immutable 25-minute deadline and rejected 35-minute, expired and NaN values. Local qualification PIDs, ports, profile, writer and held input were cleared. The locked full build passed.

Tested source is Git 6d465979ab58628a0b61e971a90f3e50f16dee0e, with dirty-source.patch empty and new-source.patch retaining the relevant untracked lib/wininit-boot.js. source.json, frozen pins and owned helper patches preserve reconstruction; no source-tree/media archive was made. Actual guest and renderer module SHA-256: e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467. The live module is retained. Compatibility module was pinned, not executed. All 525 staged pins match after execution; all 215 HTTP 200 response bodies match their pinned files. Optional-resource and DLL-search 404s remain in the audit, including optional TLB probes.

Chrome 151.0.7922.108 used the production threaded legacy D3DIM WebGL endpoint, with zero observed fallbacks/errors at the bound capture. The owning WebGL2 context reports provider/vendor Google Inc. (Intel) and renderer ANGLE (Intel, Mesa Intel(R) UHD Graphics 620 (KBL GT2), OpenGL 4.6 (Core Profile) Mesa 23.2.1). These strings do not establish physical-GPU execution.

One UUID-idempotent 40-minute box, bx_rv23s37q, was created at 12:55:41Z, original expiry 13:35:41.302Z. The immutable driver deadline was 13:20:41.089Z. Only one heavy runtime ran on the box; no old box was resumed, held-risk lane retried, worker delegated, or public deployment made. The original received zero native input events. Its exact quit ACK is retained.

The first terminal audit ran immediately after quit and failed its cleanup assertion without reporting the failing field. A subsequent read-only evidence inventory established both own PIDs absent, both ports closed, no profile/writer, no held input, zero pin mismatches and Chrome exit 0. No original runtime was restarted or input replayed. 33 unique artifacts / 14,745,817 bytes, including the live module, raw draw/pixels, screenshot and logs, were recovered and SHA-256 verified before the explicitly delegated normal nonforced stop. The box is archived, expiry null, stop completed at 12:57:47.630Z, forced=false. Local disk remained above 2 GB. No runtime custody remains.

The remaining rendering work is a separately scoped seam/edge or native reference comparison if needed; the selected original black transitions have an exact texture-content explanation and do not justify a sampler repair.