Replace XOR thinking-collapse logic with intent-based expand/collapse

crates/tui/src/tui/history.rs:435 computes collapsed = folded ^ !verbose ^ thinking_default_expanded. Triple-XOR inverts intent: verbose ON plus default-expanded ON renders collapsed, and explicitly folding while verbose renders expanded. Replace with intent logic (explicit fold wins; otherwise verbose or default-expanded expands), add a keyboard inline expand/collapse for thinking blocks (currently mouse-click only via ReasoningAction::Expand), and cover the truth table with unit tests. Work in progress on a fix branch.

Start in crates/tui/src/tui/history.rs around line 435 and trace the existing ReasoningAction::Expand mouse path. Define the expand/collapse truth table for explicit folding, verbose mode, and the default-expanded setting, then add keyboard handling and unit tests covering it. Done means intent-based rendering works and thinking blocks can be toggled inline by keyboard as well as mouse.

Written by the indexing model from the issue text.