PTCT texture and animated-scene investigation — 2026-11-06

The original issues texture operations, but the observed beam sequence renders without texturing. The new failure boundary is earlier: its sole observed texture allocation/upload attempt reaches our GL bridge before any context exists. This establishes neither an intentional wholly procedural intro nor an emulator texture defect. No emulator repair is justified by this evidence. The next investigation should establish why initialization reaches GL before context creation, including original initialization/callback ordering. Complete intro correctness and Windows-reference parity remain unverified.

Original and retained history

Fixture test/binaries/demoscene/ptct/PTCT.exe is 74,752 bytes, verified SHA-256 89028685eb2968dcc9a9dd6b7941e4fad47a70e505820dc1be350a0d30526cc4. The registered ptct launch uses resolution chooser command 1; tools/site-app-shots.json retains --input=1:wait-dlg-control:1:5000,2:dlg-cmd:1. The original launch was used without guest patches, forced GL state or replacement media.

The October 30 paired-browser section in docs/gl-d3d-corpus-status.md establishes changing gray/white beams, not complete intro correctness or same-frame parity. Its 20261030T2351Z-ptct-paired-browser run remains absent locally after searching retained evidence and recipes. I personally inspected both retained historical PNGs under history-ptct-767b07cfd830cd64 and history-ptct-7a53441d5e9f8326: gray polygon beams on black, respectively cropped and inside a desktop window. Their command, module identity, execution date and reference session are unknown. They are historical visual context only.

objdump -p exposes only glEnd under OPENGL32, plus LoadLibraryA and GetProcAddress. Runtime API tracing records dynamic resolutions including glGenTextures, glBindTexture, glTexImage2D and glTexCoord2f. Sparse imports are not the API profile.

Actual texture and state evidence

Both fresh original runs issue one glGenTextures, one glTexImage2D, and 827 glBindTexture bridge calls over the retained interval. The trace hook executes before context lookup, so issued calls must not be confused with accepted uploads. The WebGL run additionally records current-context state and original return addresses at entry:

Guest time Call Context at entry Original return address
208 ms glGenTextures(1, 0x074ffeac) current 0, contexts 0 0x412310
208 ms glBindTexture(GL_TEXTURE_2D, 0) current 0, contexts 0 0x412322
208 ms glTexImage2D current 0, contexts 0 0x41236d
256 ms wglCreateContext current 0, contexts 0 0x411c08
256 ms wglMakeCurrent current 0, contexts 1 0x411c1b

The upload arguments are target GL_TEXTURE_2D, level 0, internal format GL_RGBA8, 128×128, border 0, external GL_BGRA, GL_UNSIGNED_BYTE, pointer 0x5f7324. Its 65,536 captured pointer bytes include zeros and configuration strings; they do not establish a successfully generated texture image. Small read-only excerpts of original unpacked guest instructions retain the call sites (guest-texture-init.asm and three 384-byte snippets). No executable was patched or copied into evidence.

In lib/gl-compat.js, OpenGLHostBridge.call traces before _call; _call returns when _current() is absent, before glGenTextures or upload dispatch. The explicit WebGL context observation therefore explains the unaccepted initialization. Software observes the same calls at 208 ms and its first context creation/current at 336 ms; its cumulative accepted software uploads and unsupported uploads both remain zero. No later upload is issued on either route. Zero accepted uploads alone would not have established any of this.

All 165 observed frontend geometry batches on each route have both 1D and 2D texture units disabled. All 113 retained bounded geometry samples match across routes in raw vertices, modelview, projection, frame ordinal, enables and bound names. Their vertex alpha values are zero; issued blend factors are GL_ONE, GL_ONE, consistent with the earlier tracing note. Neither run issues glViewport; the default drawable is 634×453 inside a 640×480 guest window. The WAT mirror trust latch stays zero. Both retain matrix error 0x504 (stack underflow); that observation alone does not identify an emulator defect. The frontend histogram covers observed geometry dispatch, not an assertion about unobserved future scenes or every possible native fast path.

Animated scene review and timing

One software run and one WebGL run reached respectively 180,480 and 180,320 ms of stepped guest time. Each recorded 36 presents and 165 geometry batches (5,940 vertices); WAT counters match: 27,878 calls, 165 spans, 5,940 vertices, 23,744 flushes. These are animation observations, not gameplay or performance measurements. Only 36 presented frames over this stepped interval were observed; this is not evidence of the complete intro or normal-speed playback.

I personally inspected the following original screenshot pairs. Absolute guest times differ; corresponding presents and all retained draw samples have a constant software-minus-WebGL offset of 80 ms. Screenshot checkpoints occur in 160 ms stepping batches and differ by 160 ms. The table names checkpoint time; scene-call-comparison.json records each last-presented frame and its time.

Phase Software checkpoint WebGL checkpoint Inspected content
1 5,600 ms 5,440 ms Single gray beam on black
5 26,080 ms 25,920 ms Two overlapping gray beams
15 77,280 ms 77,120 ms Layered gray fan on the right
20 102,880 ms 102,720 ms Bright white overlap stripe and left gray beam
35 179,680 ms 179,520 ms Dense white fan in the lower right

The sequence changes on both routes. Small visible raster/edge differences are not established as defects by these phase comparisons. There is no claim of pixel parity at equal absolute guest time. Agreement between emulator backends does not establish Windows-reference parity. No verified Windows capture is available here.

Reproduction, identity and limitations

Evidence: scratch/runs/20261106-ptct-texture-correctness/. Run-only capture.cjs uses the existing dev server, trusted browser Launch, registered chooser startup, WineFrozen.step(10,16), built-in detailed API tracing, OpenGLCompat.hooks.trace, WAT census/trust exports, and read-only geometry and module observation. Commands were node <run>/capture.cjs software, then node <run>/capture.cjs webgl. Chrome 151.0.7922.108 used ANGLE SwiftShader; URLs include debug&no-threads&perf-stream&frozen=16&gl-renderer=<route>. The software capture is the browser's presented front image, not a depth-buffer screenshot. gl_sw_front is 131805184; the other entry is 131805216.

Tested source base is Git 99967820f4fe619cc1022be991e5c20f757b9f2b plus runtime-source.patch; tested-inputs.json pins relevant files and both built modules. Both actual instantiated modules match standard build/wine-assembly.wasm, 2,178,624 bytes, SHA-256 3b7511d12f3e8d364eb6bf29cdeed8b53af07abc53644d4f186e46f8620da149. Compilation/instance identity was observed, not inferred from a URL or file on disk. capture-verification.json verifies 118 actual served files per route against pinned/current source and media hashes with no unmatched identities. media-references.json also pins the served MSVCRT DLL. No source/media archive was made.

An early harness serialization error left software startup.json and phase-0.json through phase-24.json containing literal undefined. These files are preserved as failures. PNGs, progress records, console trace and the continuous in-page census were unaffected. At 123,824 ms a second read-only connection corrected BigInt serialization in the same live session, without restarting or changing guest/render state. read-correction.json, late valid checkpoints and final-observations.json retain the full cumulative trace and bounded earlier geometry samples. WebGL used the corrected serializer before its first launch. Screenshot checkpoint times before correction are reconstructed from the retained progress/stepping recipe, not invented snapshot metadata.

Both runs report no page exception. The server records 404s for optional build-info.js and favicon.ico; these are retained rather than concealed. A locked full tools/build.sh passed and generated standard and compatibility modules before capture. No emulator source changed, so no repair regression is claimed or needed. This result is a documented initialization boundary, not acceptance of complete texture/scene correctness. Follow-up requires evidence about original pre-context initialization and a verified reference before any semantic repair.

Cleanup completed inside the 09:03:01–09:48:01 UTC execution window: final-cleanup.json independently reopens 20 absent process IDs, two refused server ports, two absent browser profiles and four free build/runtime locks; free /home exceeds 5.4 GB. No test box was used (ID/state/expiry not applicable). No guest/runtime remains owned. The requested terminal result.json is an investigation record using app ID ptct; no matching candidate-corpus manifest entry was found, and it does not certify a registered candidate or gameplay.