perf(tui): batch-drain build events to kill post-completion hang

[raphaelvigee]

Problem

A warm heph run test //... shows 44/44 done, then the inline TUI hangs a few seconds "just processing events" — switching to the all view, the counter keeps climbing right up to exit.

Root cause (consumer-bound)

The engine emits typed events for every target in the transitive closure (ResultStart/ResultEnd + LocalCacheHit per dep) — tens of thousands on a warm run, flooded near-instantly into the unbounded channel. Both backends folded at most one event per tokio::select! iteration, so the burst drained one-at-a-time across thousands of iterations (interleaved with the 80ms ticker) after the headline had already painted. apply() itself is cheap; volume × per-iteration dispatch was the cost.

Fix

fold_buffered (src/tui/backend/mod.rs): on each build-event wake, greedily try_recv the rest of the buffered backlog in one pass. N select iterations collapse to one; the ticker still repaints per tick. Used by both the live arm and the final drain in interactive + ci backends. Engine emission unchanged — events stay typed, collapsed only at render.

Tests

• fold_buffered unit tests: one wake folds the whole K-event backlog and leaves the channel empty; empty channel folds nothing.
• Existing interactive loop tests (ticks / bg-pending / held) stay green.
• Full suite green; clippy -D warnings + fmt clean.

🤖 Generated with Claude Code