perf(ui): offload TUI initial task load to a worker thread

[Kohei-Wada]

Summary

TUI startup blocked first paint. on_mount scheduled the initial load via call_after_refresh(_load_tasks_if_ready) — a synchronous callback that ran the blocking httpx.Client.list_tasks plus the table/gantt presenter transforms directly on the Textual event loop. Nothing could paint until the round-trip and all view-model builds finished, and on a slow/unreachable server the UI froze for up to the client's 30s timeout.

The rest of the app already offloads sync HTTP (ConnectionMonitor, show/stats/audit dialogs) via run_worker + asyncio.to_thread; only the initial/reload path was left on the loop. This change brings it in line.

Change

TaskUIManager.load_tasks now:

1. Samples widget geometry (gantt_widget.calculate_date_range()) and sort state on the main thread (these must not be read off-thread), then
2. dispatches a worker that runs the blocking fetch + transform via asyncio.to_thread, then
3. applies state/UI mutations back on the loop (after the await), so no call_from_thread marshalling is needed.

• Exclusive "load_tasks" worker group → a reload arriving mid-flight (sort toggle, websocket task change) supersedes the in-flight one (last result wins).
• Error handling moved after the await so on_error/notify runs on the main thread.
• TaskUIManager now takes the app to dispatch the worker (mirrors ConnectionMonitor).
• Call sites are unchanged — every reload funnels through load_tasks, so the shared reload path is unblocked too.

Effect

First interactive paint no longer waits for the list_tasks round-trip + deserialization + 2×N view-model builds; a slow/unreachable server keeps the UI responsive instead of freezing.

Verification

• Startup smoke test (tests/tui/test_startup_smoke.py, runs in the default suite): a deliberately blocking API client no longer blocks mount — the app reaches an interactive state with an empty list while the fetch is in flight, then fills in once released. Confirmed it fails against the old synchronous code (mount blocks ~5s and the cache is already populated at first paint), so it genuinely guards the regression.
• make test-ui (full suite): 905 passed, no regressions.
• TaskUIManager unit tests reworked around the worker decomposition: 26 passed when run directly. Note: tests/tui/services/ is excluded from the default suite (pre-existing, see tests/conftest.py — "fix in a separate PR"), which is why the startup proof lives under tests/tui/.
• ruff ✓, mypy --strict (172 files) ✓, vulture (no dead code) ✓, codespell ✓.

🤖 Generated with Claude Code

[amazon-q-developer]

The worker offload implementation successfully addresses the TUI startup blocking issue. The changes are well-architected with proper thread safety, error handling, and comprehensive test coverage. All 905 tests pass, confirming no regressions. The code follows async Python best practices and integrates cleanly with the existing architecture.

---

You can now have the agent implement changes and create commits directly on your pull request's source branch. Simply comment with /q followed by your request in natural language to ask the agent to make changes.