Fix flaky diff_pipeline_computes_added_lines_and_ghost_blocks TUI test

[warp-dev-github-integration]

Description

Fixes the flaky TUI test diff_pipeline_computes_added_lines_and_ghost_blocks in crates/warp_tui/src/tui_file_edits_view_tests.rs (e.g. it flaked on a CI run for #13486, failing assert_eq!(ghosts.len(), 1) with left: 0).

Root cause

Two races combined to make the test flaky:

1. Bounded spinning raced the async layout channel (dominant race).
Ghost blocks travel through RenderState's layout channel, and spawn_stream_local polls that channel on the background executor before forwarding each item to a foreground dispatch task (crates/warpui_core/src/core/app.rs, spawn_stream_local). So after expand_diffs, ghost delivery makes a round trip through a background thread and its latency is unbounded under load. The test's wait — for _ in 0..100 { ... yield_now().await } — spins through the foreground executor in microseconds and gives up long before the background thread forwards the LayoutTemporaryBlock action, leaving ghosts empty. Instrumented traces confirmed failing runs where refresh_diff_state produced the ghost block but set_temporary_blocks never ran before the loop exhausted.

2. Stale diff computations could clobber newer results.
reset_content schedules its own diff computation (the buffer reset emits ContentChanged, recomputing against the just-set base), and the recompute triggered by apply_diffs only best-effort aborts it: ctx.spawn starts the future on the background executor immediately, and compute_diff_internal yields per diff op, so the two in-flight computations interleave on the background worker and can complete in either order. A superseded computation that already completed still delivered its result, overwriting the newer diff status and emitting a misleading DiffUpdated (which, with diff nav expanded, re-runs refresh_diff_state with the stale status and can clear the ghost blocks).

Fix

• DiffModel::compute_diff (app/src/code/editor/diff.rs): the completion callback now drops results that are no longer the latest requested computation (compared against the version stored alongside the abort handle), so stale results never clobber a newer diff. This also hardens the production diff pipeline, where last-write-wins by completion order was incorrect.
• The test (crates/warp_tui/src/tui_file_edits_view_tests.rs): both waits are now event-driven instead of bounded spinning. It subscribes to DiffUpdated/LayoutInvalidated on an unbounded channel, waits until the diff model actually contains the applied hunk before expanding, and then re-checks for ghosts after each LayoutInvalidated (emitted right after the render state stores the temporary blocks) rather than spinning a fixed number of iterations.

Linked Issue

N/A — flaky-test fix; no tracked issue.

Testing

• Reproduced the flake at the base commit: 20/300 failures with all CPU cores saturated (always ghosts.len(): left: 0).
• With the fix: 0 failures in 1100 runs (500 + 300 under full CPU load, 300 unloaded).
• cargo nextest run -p warp_tui --lib: 102/102 passed.
• cargo nextest run -p warp_editor: 458/458 passed.
• cargo nextest run -p warp code::editor: 95/95 passed (covers diff_tests and editor model_tests).
• ./script/format clean; cargo clippy -p warp -p warp_tui --all-targets --features tui -- -D warnings clean.

This is a headless-test/infrastructure change with no user-visible behavior, so no manual ./script/run testing or screenshots.

Agent Mode

• Warp Agent Mode - This PR was created via Warp's AI Agent Mode

CHANGELOG-NONE

Conversation: https://staging.warp.dev/conversation/b96d06d2-7f01-4e47-acef-6039567955c8
Run: https://oz.staging.warp.dev/runs/019f5d59-095b-7b38-a79c-8cc11cf4b5c6

This PR was generated with Oz.