[tiensonqin]

Summary

Keep scrolling ahead of deferred result rendering: prefetch up to 64 rows while retaining the 12-row initial snapshot. Mount only the native viewport plus two rows on either side, independently of the cached data window. Display an in-flight reply when it improves current viewport coverage, even if the exact prefetch key moved; preserve graph/generation/input/revision isolation. Experimental whole-viewport column overscan was withdrawn after real iOS recordings exposed excess remount work.

Reuse scope, filtering/sorting and column projections in a bounded session-local cache. Invalidate by immutable database identity and local journal day: independent graph projections may share a transaction number, and relative-date queries can change overnight without graph edits. Window reads still encode only requested rows.

Validation:

• Core gate: dune build @shared/native/runtest — 1250 tests passed. Focused paging/result-view tests, Android JNI/packaging checks, and Android unit/lint checks passed. The overlapping-response regression fails before the fix and passes afterward.
• CI follow-up changes only the existing iOS capture flow's drawer-closing action to its visible sidebar toggle. All draft/send/edit assertions remain intact; production sources are unchanged after 12307830.
• Real iOS Release baseline: 1000-row Table, identical fast inputs, 1.84–1.89s full-body blanking during inertia. This is not an independently established one-second post-stop timer.
• Current 12307830 Android profile/x86_64: unchanged 302×9 Table, original viewport and 180ms inputs; no full-body scrolling blank, missing-row region, or newly blank whole column observed. Installed APK/core hashes match the build. Separate cold-open residual: the bottom visible row appears after 7 raw frames, approximately 117ms; horizontal updates remain discrete.
• Current 12307830 iOS Release: three vertical/horizontal pairs and a separate historical R2 replay with the pre-input Row965/Field5–7 viewport matched within 1px. Raw-sample review reported no full-body blank or contiguous missing-row band. Final horizontal smoothness is not established: desktop capture has 150–183ms sampling gaps around H1; these must not be counted as confirmed app stalls or a pass. The approximately matched 300-row case is not a pixel-identical historical replay.

Remaining limitations: simulator/emulator evidence, no physical-device claim, no exhaustive row/column boundary coverage, and no independent touch-release/post-stop timing measurement.

CI follow-up

At a448f3bd, Vendor duniverse, ocaml-core, and WIP passed. Both E2E jobs failed before running the smoke flow:

• Android: installing the API 34 emulator system image exhausted the runner disk (No space left on device).
• iOS: the cold native build could not read _build/install/mobile_host/lib/ppxlib/ast/ast.ml.

The iOS selector follow-up has therefore not yet been exercised by CI. These latest failures have not been dismissed as flaky or pre-existing. Re-running the failed jobs was rejected with Resource not accessible by integration; rerun and further investigation remain outstanding.

Link to Devin session: https://app.devin.ai/sessions/f2fb979c38d945a9a836b086a5132f95
Open in Devin Desktop: https://app.devin.ai/desktop/session/f2fb979c38d945a9a836b086a5132f95?variant=devin
Requested by: @tiensonqin

---

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

• Disable automatic comment, CI, and merge conflict monitoring

[devin-ai-integration]

✅ Devin resolved the finding on a448f3b

Fixed by Devin (1)

• Date-relative query windows stay stale overnight

View all findings in Devin Review