fix(tui): preserve streaming scroll-follow across Ctrl+T thinking toggle

[alejandro-ao]

Summary

Fix a re-regression of #175: pressing Ctrl+T to toggle thinking-token display during an active stream broke the smooth bottom-following scroll restored by #175 / PR #177. The toggle now reconciles thinking widgets incrementally while streaming instead of triggering a full transcript redraw.

Motivation / problem

While an assistant message is streaming and the viewport is pinned to the bottom, pressing Ctrl+T to show/hide thinking tokens stopped the smooth auto-follow scroll: the viewport either stopped following or jumped/snapped, and only recovered after the turn ended.

action_toggle_thinking always called _refresh() → TranscriptView.update_from_state() → _redraw(), which does a full remove_children() + remount of every transcript block. During an active turn that:

1. Tears down _active_assistant_widget / _active_thinking_widget (reset to None), so the next streamed delta mounts a fresh assistant widget (visible jump / lost incremental markdown stream) — the "thinking stream appears in a different way" symptom.
2. Reflows the whole transcript; the transient content-height change moves scroll_y, and TranscriptView.watch_scroll_y interprets that as a user scrollback and flips _follow_output = False, snapping the viewport back to the bottom.

This is exactly the "scroll state reset during token deltas" failure mode #175 warned about. The earlier "Reduce transcript toggle latency" change (ae42fd0f) optimized the toggle for the idle transcript and didn't account for the toggle firing mid-stream.

What changed

• TranscriptView.apply_thinking_visibility(state, theme) reconciles only the thinking-related widgets against state.items thinking blocks, never remounting the active assistant streaming widget so its incremental markdown stream stays intact.
  • Showing: remove all thinking widgets (finalized TranscriptMessageWidget blocks, live StreamingTranscriptMessageWidget blocks, and the hidden-thinking placeholder), then mount a fresh widget for each thinking item ordered ahead of the assistant widget. StreamingTranscriptMessageWidget(item) re-renders the full item.text on mount, and the next thinking delta continues appending to the freshly assigned active widget.
  • Hiding: remove all thinking widgets and, if a thinking item exists, mount a single hidden-thinking placeholder ahead of the assistant widget.
• Follow/scrollback is preserved across the reflow: _follow_output is snapshotted, _suppress_follow_update makes watch_scroll_y ignore layout-induced scroll movement during the reconcile, and a call_after_refresh callback restores _follow_output and re-requests the bottom-follow scroll after layout settles. A mid-stream toggle therefore does not change follow mode.
• action_toggle_thinking uses the incremental path while state.running, and keeps the full _refresh() redraw for the idle case. The reconcile is synchronous (fire-and-forget mount / remove_children, matching _redraw) so it stays atomic with respect to the streaming worker.

Acceptance criteria / verification

• Ctrl+T during an active stream keeps the transcript pinned to the bottom with smooth following scroll when the user was already following.
• Ctrl+T during an active stream, after the user scrolled up, preserves the user's scrollback position (no jump back to the bottom).
• Toggling thinking on/off repeatedly mid-stream does not cause reflow jumps or scroll resets.
• Multiple thinking blocks (across tool calls) all re-render on show and collapse to one placeholder on hide.
• The active assistant streaming widget survives the toggle and keeps receiving subsequent deltas.
• Behavior matches Fix TUI scrollback regression while assistant messages stream #175/Fix #175: restore streaming transcript scrollback #177: auto-follow only while at the bottom; opt-out on user scroll-up; resume when returning to the bottom.

Regression tests in tests/test_tui_app.py:

• test_tui_thinking_toggle_during_stream_preserves_follow_scroll
• test_tui_thinking_toggle_during_stream_preserves_user_scrollback
• test_tui_thinking_toggle_during_stream_keeps_all_thinking_blocks
• test_tui_thinking_toggle_via_event_stream_preserves_assistant_widget

uv run pytest tests/test_tui_app.py -k "thinking_toggle_during_stream or thinking_toggle_via_event_stream"
uv run ruff check . && uv run mypy

All 610 tests pass, ruff + mypy clean.

Related files or areas

• src/tau_coding/tui/widgets.py — TranscriptView.apply_thinking_visibility, _show_streaming_thinking, _hide_streaming_thinking, _remove_thinking_widgets, watch_scroll_y, _suppress_follow_update
• src/tau_coding/tui/app.py — action_toggle_thinking
• tests/test_tui_app.py — new regression tests
• dev-notes/architecture/fix-thinking-toggle-stream-scroll.md

Notes

• TUI-only fix; does not touch the core agent harness.
• Related: Fix TUI scrollback regression while assistant messages stream #175 (closed, fixed by Fix #175: restore streaming transcript scrollback #177). Key commits: ae42fd0f, cb488805/a5dff600/19e4c498.