fix(desktop): keep clarify prompts answerable

[lidises]

Summary

• Keep clarify tool lifecycle events visible to Desktop even when tool progress is set to off, so the interactive prompt can still render and resolve.
• Reconstruct a pending clarify tool row from clarify.request when the earlier tool.start row was missed during reconnect/hydration.

Why

Desktop can otherwise show that a session needs input while providing no answerable clarify prompt, leaving the backend blocked on clarify.respond.

Test plan

• ./.venv/bin/python -m pytest tests/test_tui_gateway_server.py::test_tui_clarify_lifecycle_events_emit_when_tool_progress_off tests/test_tui_gateway_server.py::test_tui_non_interactive_tool_lifecycle_stays_hidden_when_tool_progress_off -o 'addopts=' -q
• cd apps/desktop && npm run typecheck

[teknium1]

Thanks for identifying a real blocked-input path. Current main still suppresses tool.start and tool.complete when tool progress is off (tui_gateway/server.py:3606, 3664), while Desktop needs a tool-call part to render ClarifyTool (apps/desktop/src/components/assistant-ui/thread/message-parts.tsx:35-37).

Problems

• The Desktop file changed by this PR was moved to apps/desktop/src/app/session/hooks/use-message-stream/gateway-event.ts by 51a710e57e; the change needs to be ported there.
• The synthetic row is keyed with requestId, but _block() creates that ID independently (tui_gateway/server.py:2034-2040) from the lifecycle callback’s tool_call_id (tui_gateway/server.py:3594, 3621). It cannot merge with the later real completion as the comment claims, so the synthetic pending row can remain alongside a separate settled row.

Suggested changes

• Preserve the narrow gateway lifecycle exception, but rework missed-start recovery to use an ID shared with the later completion (or carry that tool-call ID in clarify.request), and cover that sequence in Desktop tests.

Automated hermes-sweeper review.

[teknium1]

requestId is generated independently by _block() (tui_gateway/server.py:2034-2040), while the actual tool.start/tool.complete use the agent's tool_call_id. These IDs cannot merge in upsertToolPart; after completion this synthetic pending row can remain beside a separate completed row. Use a shared lifecycle ID or carry the tool-call ID in clarify.request.

[OutThisLife]

Superseded by #69795 — salvaged with your commits (the Python half cherry-picked verbatim; the TS half re-authored onto the post-split use-message-stream/gateway-event.ts since the file you patched was refactored on main). Authorship preserved via cherry-pick + Co-authored-by, and added a contributor mapping for your email. Thanks @lidises! Closing in favor of the rebased version.