[Framerr]

Goal

Diagnose the reproducible first-capture bug in #14 without undoing the approved DMA flush fix.

Reporter observed 3/3 first photos corrupt across independent restarts, with one second photo corrupt and all third photos correct. A separate rapid-photo test reported one vertically wrapped image.

Experimental change

• In MaxwellDMA::CopyBlockLinearToPitch, match the expected 1920×1080 RGBA8 image-to-buffer transfer (7680-byte pitch and row length, 1080 rows, supported block-linear source layout).
• For the first two matching copies per emulator process only, bypass AccessAccelerateDMA().ImageToBuffer() and use the existing CPU unswizzle/readback fallback.
• Every other transfer uses the original acceleration + approved flush path unchanged.
• Log the first twelve matching copies (source/destination, geometry, index, fallback) to confirm that the probe actually reaches the photo path.
• Control: FRAMEROPIA_FIRST_PHOTO_DMA_CPU=0 disables the test and restores normal behavior.

Caveats

This is an A/B diagnostic hypothesis, not a proven fix. The exact shape match may fail if Pokopia uses another pitch/copy geometry; inspect logs before interpreting a negative test. CPU fallback may introduce a temporary slowdown for the initial photographs. No generalized buffer-cache change and no old PR #16 patch.

Validation

• Windows CI build succeeds, artifact-only, no release
• Log reports Pokopia photo-shaped DMA copy ... fallback=true for first photo
• Across 5 fresh emulator starts, record outcomes of first three photos per start
• Repeat with FRAMEROPIA_FIRST_PHOTO_DMA_CPU=0 as control
• Check vertical wrap, geometry stability and FPS

Do not merge without reporter confirmation. Baseline nx2 remains unaffected.

[Framerr]

Reporter test: severe photo regression — reject this experiment (2026-11-08)

The reporter tested the first-photo CPU DMA fallback experiment and reports all saved Pokopia photographs are corrupt. This is much worse than the previous accepted DMA flush behavior (initially 48/50 good, though first-capture corruption remains reproducible).

Result: FAIL. Bypassing accelerated ImageToBuffer for the first photo-shaped transfers does not fix first-capture corruption in the tested setup, and is associated with severe regression. Do not merge. No root cause has been established. Restore the accepted nx2 baseline (PR #17) for further tests. Keep branch for reference. No new release.