Falling Leaves: corrected shadow alpha on software and browser WebGL

Run 20261103-falling-leaves-paired-shadow-alpha closes the missing paired shadow-alpha evidence for registered scr_fallingl on committed input e1dbee81690cd7e4f424601b07110458517182c6. Original FALLINGL.SCR /s runs the real guest d3drm.dll. No production code or scene asset was changed. This validates the October 10 correction in the screensaver notes, not the superseded interpretation that opaque black shadows were intended.

The canonical module actually served to all four browser launches is 0a58fd7b9ac82c5303048776f442b15e72528b55ddf85564e0df92fcb1aa384d. source-seal.json.gz, source-supplement.json, source-input-SHA.json and source-input.tar.gz retain the committed closure and original input bytes; verification.json checks successful served responses against those pins. Foreign working-tree changes were excluded with git ls-tree/git cat-file on the exact input commit.

What the alpha evidence establishes

The focused node test/test-d3dim-lit-alpha.js passes: an exact material diffuse alpha of 0.5 produces vertex alpha 128, while opaque, transparent and missing-material cases retain their expected alpha.

The original scene has a useful quantization detail. Its black shadow material contains [0, 0, 0, 0.49803921580314636], approximately 127/255. The existing two-decimal trace would print this as diffuseA=0.50. At actual shadow triangle execution, the diagnostic captures material handle 28, its 80-byte material, post-lighting color 0x7f000000, texture handle 0, alpha blending enabled, and blend factors 5/6 (SRCALPHA/INVSRCALPHA). Twelve samples on each route check this chain. This is preserved half opacity within the original DLL's byte quantization; it is not the old forced 0xff000000 result.

On actual browser WebGL, the diagnostic also observes lowered shadow payloads whose vertices all contain black with alpha 127. The fixed-function alpha stage selects diffuse alpha. Sixteen sampled actual drawArrays calls during those shadow submissions have GL blending enabled with SRC_ALPHA (770) and ONE_MINUS_SRC_ALPHA (771). analysis.json and the diagnostic frame JSONs retain the material, vertex, state and GL records.

These probes only read guest memory and observe existing host calls; they do not change guest state, materials, render state, shaders or module bytes. They use the generated region map and existing guest-address translation. The software route has no D3DIM GPU executor. Its actual execute-triangle trace and bound material/blend snapshots establish the software-side chain; the WebGL executor records establish the browser submission chain.

Paired scenes and limits

Both ordinary arms and both diagnostic arms capture three scene frames at 3,000, 4,500 and 6,000 frozen browser slices: 6, 9 and 12 guest seconds. The explicit frozen tick is 2 ms, matching the registered cliTickMsPerBatch: 2 value; browser slices and CLI batches are different execution units. The browser records a 100,000-block slice setting, with its normal cooperative deadline. Guest threads are disabled on both routes. Each arm is capped at 300 seconds and completes before the next starts.

Arm Final ExecTri trace records / black alpha-127 records WebGL draws at the three captures
Ordinary software 12,864 / 6,206 N/A: no GPU executor
Ordinary WebGL 12,930 / 6,420 7,581 → 10,034 → 13,539
Diagnostic software 8,517 / 4,697 N/A: no GPU executor
Diagnostic WebGL 15,009 / 7,455 4,550 → 9,714 → 15,654

All 552 successful served-response hashes match the sealed inputs, including all nine executable/scene/DLL paths listed below on each launch, with no unsealed successful response. This does not erase the separately logged HTTP failures. Within the ordinary arms, successive scene-only PNG crops change by 37,268/43,531 pixels on software and 43,395/57,941 on WebGL. The crop excludes UI and clock; these counts establish motion, not backend parity.

Personally inspected PNGs show gold/green leaves changing position and pose, with offset dark shadows over the textured leaf backdrop, on both software and WebGL. The successful browser responses pin identical module and scene inputs across the arms. Actual local HTTP endpoints, Chrome version, renderer counters and cleanup are retained per arm. WebGL is Chrome's WebGL 2 through ANGLE Vulkan SwiftShader, not Node GL or a physical-GPU test. The owning host enables D3DIM GPU only on the WebGL route; its executor records zero errors and zero fallbacks.

Scene rendering alone would not prove visual correctness. Together with the draw evidence, this closes the narrower missing corrected shadow-alpha validation. It does not establish same-frame pixel parity, native Windows reference parity, full texture/coverage correctness, a complete saver cycle, FPS or completion of the historical 27-app corpus. The captures retain visible edge/texture artifacts and overlapping dark silhouettes; these are not silently certified. Screensaver input is N/A. Missing optional DLL, type-library, UI-resource and local account-endpoint messages remain in the logs; this is not a warning-free-run claim.

Original inputs actually fetched

Paths below are relative to test/binaries/; the browser serves them through the usual binaries alias. The seven executable/scene hashes also match the retained October 30 20261030T2222Z-d3drm-scenes/assets.json. Both archived October 30 logs were decompressed and checked against ARCHIVE-RECOVERY.json.

Input SHA-256
screensavers/FALLINGL.SCR 470e28bcbc3cf92c514cd8e8bbf9eae399c59f79ed2adbf7bf2b91d25cfb477c
screensavers/FALLINGL.SCN 3a12aa6547c1e4c1dab6a6c3023c0f22f3d0e2845a3983e69f63456748228896
screensavers/LEAF.X 323fa9090f26ab88e0e71b4fe35af06aa7ba53825817f954f7e1a129438bd3f3
screensavers/LEAF1.GIF 8f4b1cc27dfa2c7af7ac81ddf757be12bf569cdfae54f0f8a58ca0bf034e1c27
screensavers/LEAF2.GIF 59e114c726b661d14c1c8e60fec899e9f192c8f6a1fe581d7a1bde822c54dc33
screensavers/LEAF2.X a87002dea22968dc093c50789fea29b24e1c72d611841ff0127a12306f2c6df2
screensavers/LEAVES.GIF 18bac057ff316699d5ea6502033adda9e5c8825984a9c7c8d32831950893fbd6
dlls/d3drm.dll fd80f3839a035b6b52362735b22eb8d2523d3434bf18afb3e0f1b5ace84357b0
dlls/d3dxof.dll 1f2d0683668b3039d451d88ed449fd9d49371fab3fadb991636cd6bfb0aa6eee

Two additional CRT DLLs were staged and pinned in assets.json but were not fetched by these browser launches; they are not claimed as executed inputs.

Reproduction and first failure

The retained original-driver.cjs is the recovered October 30 browser harness. browser.cjs narrows it to this registered app, the current sealed module, explicit frozen clock and bounded captures. diagnostic.cjs adds the memory and submission observations described above. In the recovered sealed tree:

node tools/pe-imports.js test/binaries/screensavers/FALLINGL.SCR --all
node tools/pe-imports.js test/binaries/dlls/d3drm.dll --all
node tools/pe-imports.js test/binaries/dlls/d3dxof.dll --all
node tools/check-wat-manifest.js
flock /tmp/wine-build.lock bash tools/build.sh
node test/test-d3dim-lit-alpha.js
# Run each command to completion before starting the next.
timeout --signal=TERM --kill-after=5s 300s node scratch/runs/20261103-falling-leaves-paired-shadow-alpha/browser.cjs software
timeout --signal=TERM --kill-after=5s 300s node scratch/runs/20261103-falling-leaves-paired-shadow-alpha/browser.cjs webgl
timeout --signal=TERM --kill-after=5s 300s node scratch/runs/20261103-falling-leaves-paired-shadow-alpha/diagnostic.cjs software
timeout --signal=TERM --kill-after=5s 300s node scratch/runs/20261103-falling-leaves-paired-shadow-alpha/diagnostic.cjs webgl

All heavy build/test and browser execution occurred on one finite-TTL boat, bx_xb4h5hmk, original expiry 2026-11-03T17:39:19.176Z. Its identity, expiry and original-task delegated normal shutdown were recorded before use.

The first full build failed because this worker's source staging filter omitted the two committed docs/dos-corpus/live/toyvm-*-bundle.js files. first-build.log, first-build.exit and build-command.json preserve that failure and exact invocation. Adding those pinned committed files fixed the staging omission; the same locked build passed without a budget increase. No production defect was established. The first observational selector also expected exact alpha 128 and therefore counted no original-scene shadows; its records are retained. The diagnostic runs account for actual alpha 127 without changing render inputs or execution budgets.

RECOVERY.json, recovery-download.json, the recovery archive and full SHA manifests retain replay commands, source, assets, tested modules, logs, frame state and PNGs. Recovery is verified before the delegated normal STOP. Final lifecycle receipts and publication details are retained in the run. The 56,690,344-byte recovery archive has SHA-256 9b0c0cf3256f04858702350047dad1319532c2d139f59b04deee186b2d7c0a8b; all 122 payload files and 5,842 source/input members were retrieved and hash verified before STOP. The boat reached STOPPED at 2026-11-03T16:25:18.415Z, stop operation stop_2aa02dfb5503861b completed with error: null and forced: false. Snapshot cf344642-29b8-4b43-a6c6-c6aa73d4fc87 is completed, available and verified at 2026-11-03T16:25:17.988Z. All four browser PIDs and profiles are absent, all four local HTTP ports are closed, and owned runtime helpers are absent. Snapshot custody returns to the coordinator; this worker has no successor task.