Write PSD interleaved pixels directly into the final buffer. Direct RLE output is already on main. The unrelated export-side sample-encoder inlining hint is removed locally.
Whole-photo imports improve by about 4–7%, and the authored designs by about 20–26%. An added export-side inlining hint accompanied a separately confirmed 2.3% float PSD export slowdown (2440.168 → 2497.423 ms). Removing that one line leaves export source equal to main; follow-up float export is +0.4% and 16-bit export +0.9%, with both paired intervals crossing zero. These results supersede the much larger previous interleave-only percentages.
Measured on 2026-10-19 against main `d95fdf3181c01ad41e95456091a394b105a9afb2`, independently of the other splits. Apple M4 Max, 16 CPUs, 128 GiB, macOS arm64, rustc 1.99.1; release O3, LTO off, 16 codegen units, debug off. Builds used nice 10 and warm APFS-cloned caches; measurements used default priority without concurrent builds or profiling. Five alternating paired samples follow a warmup; confirmation and fix runs use nine pairs. Time change compares the two medians; intervals resample paired ratios. Negative time change means faster. Small differences are not automatically established improvements or regressions.
The full-resolution CC0 Pixls photographs are P1000475.RW2 (24 MP foliage centre crop, no resampling) and IMG_4059.CR2 (12 MP city), each converted to U8/U16/F32 with the ProPhoto RGB profile. F32 derives from 16-bit development, not HDR capture. MIT psd-tools masks.psd supplies a 640 x 1394 web layout with 22 imported layers and three masks; mask-index.psd supplies a 1000 x 867 illustrated card with a mask. The sources and hashes are pinned by the existing corpus. These are the available authored layered files, not a large layered-photo corpus. No generated noise, repeating ramps, enlarged charts or duplicated photo layers are used.
The shared harness times in-memory operations. Input loading, image conversion, thumbnail rendering, hashes and equivalence checks are outside timers. Every save uses a fresh PcraftWriter and a real rendered thumbnail. Common reads use main-encoded files; own-output reads are separate. PSD exports use ZIP merged images at U8/U16, Raw merged images at F32, and normal RLE layer channels at U8/U16 / ZIP prediction at F32. Results do not measure filesystem throughput or browser performance.
Real PSD exports remain byte-identical and imported layer/mask float samples match main. The committed literal previous-interleave oracle covers depths, bands, fills, truncation and inversion.
Current local checks: affected `pixels::` unit tests (five passed, one ignored manual benchmark); strict all-target IO Clippy; affected IO wasm32 check. All passed. Follow-up exports at float and 16-bit depth remain byte-identical to main; no synthetic timing test was run.
Local raw evidence and the shared harness: `.local/perf-real-261019/` in the repository checkout (ignored). This comparison supersedes the earlier synthetic performance table. Small synthetic unit fixtures remain correctness checks; ignored synthetic timing tests were not run.
## Earlier split validation
These checks passed when the split was prepared; they were not all rerun for this benchmark review. Current affected checks are listed above where production code changed.
