[k-kostin]

Summary

A ready disk-cache PNG currently waits behind running video/PDF/image generation in the provider's one pool. Probe cache entries in a separate two-worker pool; send misses to the existing generation pool, where freshness/cache is checked again. Text pages go straight to their existing generation route.

image request → background cache lookup → fresh PNG: finish
                                    → miss: existing generation queue

No disk access moves to the UI thread, no RAM cache or dependency is added, and generation concurrency remains unchanged. A missing PNG is detected before querying source metadata, avoiding an extra network-source lookup on ordinary cold misses. Cache freshness uses the existing policy.

A response is separate from runnable tasks. Cancellation is checked around the cache stage and before queued generation; destruction drains cache work before generation because cache misses may still enqueue jobs. Every response finishes once. This does not import #117's running-helper cancellation.

Evidence

• Before: with real generators held in every slot, an existing disk PNG cannot make a QML Image ready before the gate opens. After: it becomes Ready while generators are still held. The test observes persistent QML state, not a racy finished-signal subscription.
• Eight alternating pairs per direct-provider scenario after warm-up: busy-cache response median 507.9 → 0.57 ms, p95 508.5 → 0.59 ms. This deliberately holds generators for 500 ms; these are response completion times, not displayed-frame latency.
• Eight 12 MP synthetic JPEGs (baseline/progressive mix), cold thumbnail cache: all responses 210.7 → 209.8 ms median. Idle warm-cache batch: 1.33 → 2.33 ms, reflecting the two-reader bound; first response stays about 0.7 ms. Warm OS page cache throughout.
• Real Backend/FolderGrid, three pairs each: all eight pictures 227.0 → 227.6 ms, jump to the last of 512 entries 353.3 → 350.2 ms, navigation during generation 43.20 → 43.26 ms. All 18 measured runs complete and drain without failure markers. Offscreen Qt Quick software rendering; physical frame latency was not measured.
• Release build, diff check and complete local tests: 198 passed, 0 failed, using Qt image plugins matching CI dependencies. Tests cover stale/corrupt/missing cache, removed source, cache recursion and 64 cancelled queued misses.
• ASan/UBSan with rapid real QML creation/cancellation: candidate 2378 and base 2377 requests all finished and were destroyed while the reader remained alive; no sanitizer errors (leak detection disabled). Immediate engine destruction instead leaves a few responses undestroyed on both versions; that boundary remains recorded and is not fixed here.

Merge Danger

Door: two-way.

Blast Radius: thumbnail admission. Two additional bounded workers can themselves wait on slow cache/source I/O; the existing generation limit is preserved. Pending #67/#117 changes will need the usual rebase and validation before integrating on their final base. This PR targets main independently; #67/#117 are not base dependencies.

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; no actionable issues were found.

Summary

Separates disk-cache reads from thumbnail generation so cached pictures do not wait behind busy generators.

• File grids and previews can load cached thumbnails while other files are still being generated.

Acknowledged limits:

• k-kostin notes that slow cache/source I/O can occupy both cache workers.
• k-kostin records responses left undestroyed during immediate engine destruction on both versions; that boundary is intentionally outside this change.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    Request[Thumbnail request] --> Kind{Text page?}
    Kind -->|Yes| Generate[Existing generation pool]
    Kind -->|No| Cache[Two-worker cache pool]
    Cache --> Fresh{Fresh cached image?}
    Fresh -->|Yes| Finish[Finish response]
    Fresh -->|No| Generate
    Generate --> Recheck[Check cache again and generate if needed]
    Recheck --> Finish

Loading

_{Reviews (1) · Last reviewed commit: "Observe fast cache completion through QM..." · Reviewed by Greptile}