[elhigu]

Description

This is part 2 of the performance fixes listed in my fork's README; part 1, #1790, is merged. It is rebased on current main (211ef55) and works with the new compacted decoded-image cache.

The decoded-image cache had two problems:

• It could return stale images. It was keyed by path alone. A file rewritten under the same name kept returning the old pixels, for example tethered captures from a camera that reuses names like capt0000.nef. So did images decoded before a raw pre-processing setting changed: changing linearRawMode never cleared the cache, and a decode that finished after a settings change could put an image with the old settings back.
• Full-resolution previews didn't use it. generate_preview_for_path decoded the file from disk on every call, even when the photo was open in the editor.

Where it makes the app faster

generate_preview_for_path renders the full-resolution previews of:

• The culling view. Coming back to the culling view after editing a photo, or culling photos you have recently opened, renders them without decoding the raw again. The cache holds the photo open in the editor and up to imageCacheSize recent ones (5 by default), within its memory limit.
• The collage modal, for photos that are already decoded.
• The tethering panel, for captures that are already decoded. Opening a new capture in the editor now shows the new file even when the camera reuses its name; before, the editor could show the previous capture.

When the culling view and the editor ask for the same photo at the same time, it is now decoded once instead of twice.

Benchmark

Measured with the pipeline benchmark from my fork (tools/pipeline-bench, rapidraw bench). It runs the app's own functions without a window, times each step and hashes its output. main (211ef55) and main with this PR, median of 5 runs each, on a 12 MP Sony A7S II raw and a 26 MP Sony a6700 raw with their edits:

Step	12 MP, main	12 MP, this PR	26 MP, main	26 MP, this PR
Full-resolution preview of a decoded photo (generate_preview_for_path)	556 ms	263 ms	1598 ms	886 ms

Full-resolution previews of a photo that is already decoded are 1.8-2.1x faster. Decoding, thumbnails, opening a photo, the first editor preview and export are unchanged within run-to-run noise, and every step's output is identical with and without this PR. Both builds ran with MIMALLOC_ALLOW_THP=0 so that allocator settings don't distort the comparison.

Type of Change

• Bug fix
• New feature
• Breaking change
• Performance improvement
• Code refactoring
• Documentation update
• UI/UX improvement
• Build/CI or Dependency update

Changes Made

• fix(cache): reuse decoded images only while file and settings match. Each cache entry carries a fingerprint of the file's size and modification time and of the six settings that change decoded pixels (highlight compression, linear raw mode, colour noise reduction, sharpening, pre-processing of non-raws, Apple RAW). A lookup that no longer matches is a miss and drops the entry. Entries keep their fingerprint while the cache holds them compacted. Decoding moves into one helper, get_or_decode_pristine, which also makes concurrent requests for the same file share one decode. load_image uses it, after compacting the other entries as before, is_image_cached checks the fingerprint without expanding a compacted entry, and the background AI jobs check it before reusing a cached decode.
• perf: reuse decoded images for full-resolution previews. generate_preview_for_path uses get_or_decode_pristine and prepares the cached image with prepare_base_image_cow, so AI denoise, negative conversion and AI patches are applied exactly as before, and the cached image is borrowed instead of cloned when none of them changes it. It doesn't compact anything, so the photo open in the editor stays in full precision. For a photo the cache holds compacted, the preview uses that copy, the same as reopening the photo in the editor does.

Trade-off: full-resolution previews now add their photos to the decoded-image cache, within its existing count and memory limits.

Screenshots/Videos

No UI changes.

Testing

• These changes were tested locally by a human and confirmed to work.
• I haven't added any automated tests to the code because the codebase currently lacks a test suite.

Test Configuration:

• OS: NixOS 26.11, Linux 7.2, GNOME on Wayland
• Hardware: AMD Ryzen AI 9 HX 375 with Radeon 890M (RADV, Mesa 26.2.4), 32 GB RAM

Automated checks:

• The benchmark above: every step's output hashes identically on main and with this PR.
• Every commit builds and passes cargo fmt --check and cargo clippy --all-targets --all-features -- -D warnings.
• Six unit tests of the cache pass: a rewritten file and a changed setting change the fingerprint, a stale entry is dropped instead of returned, the least recently used entry is evicted first, concurrent decodes of one file run once, and an entry compacted to f16 is still found by its fingerprint and dropped when it goes stale. They are not part of this PR; they're on my fork (test/decoded-cache-unit-tests) if you want them.

Checklist

• My code follows the project's code style
• I haven't added unnecessary AI-generated code comments
• My changes generate no new warnings or errors

Additional Notes

Suggested release-notes bullet:

Faster full-resolution previews: the culling view, collage and tethering previews reuse already decoded photos, and the cache no longer shows outdated images after a file is overwritten or a raw setting changes.

AI Disclaimer:

Please state the involvement of AI in this PR:

• This PR is created by an AI agent
• This PR is mostly AI-generated but edited/merged together by a human
• This PR was handwritten with AI assistance (spell check, logic suggestions, error resolving)
• This PR contains only blood, sweat, and coffee (AI-free)

The code, measurements and this description are 100% AI-generated with Claude Code.

🤖 Generated with Claude Code

[elhigu]

Rebasedand fixed conflicts

[elhigu]

rebased, fixed conflicts, updated benchmark numbers

[CyberTimon]

Thanks a lot for the PR, looks good to me. Also sorry for all the merge conflicts I've caused ;)