refactor(sidebar): redesign agents panel to uniform two-line entries

[dgageot]

The agents panel in the TUI sidebar had two distinct rendering modes: a multi-line "focus card" for the current agent and compact "roster rows" for the others. This asymmetry made the layout shift noticeably when switching agents and required maintaining separate layout computation paths.

This change replaces both modes with a single renderAgentLine that produces a consistent two-line entry for every agent. Line 1 shows the agent name in its accent color, a thinking-effort badge right-aligned in a shared column, and a ^N switch shortcut flush against the right edge. Line 2 shows the provider/model string, indented and left-truncated so the informative tail is always visible. The current agent is marked with ▶ (or the spinner while working); other agents pad the marker column so names stay aligned. Agent descriptions are removed from the sidebar — they remain accessible in the agent-details dialog.

Click zones are updated so both rendered lines of an entry are owned by the agent, while the blank separator lines between entries remain unclickable. The ^N shortcut displayed matches the existing ctrl+1..ctrl+9 keyboard handler.

Removed now-unused helpers — renderAgentCard, renderAgentRow, computeRosterLayout, cardThinkingLine, renderTreeNodes, wrapDescription, ellipsizePlain, rowShortcutCell, rightAlign — and the rosterLayout type. Added small padLeft/padRight/badgeColumnWidth helpers and moved badge-column-width computation into agentInfo so it is derived once rather than per render. Layout constants rowShortcutWidth/rowIndentWidth are replaced by agentMarkerWidth/agentShortcutWidth.

One pre-existing edge case (not a regression): below the MinWidth auto-collapse threshold combined with very long agent names, tab.Render may wrap body lines and misalign click zones.

[docker-agent]

Assessment: 🟢 APPROVE

This refactor replaces the two-mode rendering system (renderAgentCard + renderAgentRow) with a unified renderAgentLine that produces consistent two-line entries for all agents. The new implementation is clean and logically sound.

Width/layout calculations — The nameWidth formula is internally consistent with the right-column formula. The - 1 separator correctly accounts for the literal space between badge and shortcut columns; the gap always resolves to exactly minGap=2. No off-by-one errors.

Left-truncation of model strings — TruncateTextLeft(modelText, max(1, contentWidth-agentMarkerWidth)) correctly limits model text to the available space, producing a total line width equal to contentWidth.

Click zone mapping — The agentLineOwners approach is preserved. buildAgentClickZones correctly applies the tabHeaderLines=2 offset and skips blank owner entries. The updated click tests thoroughly verify the new 2-lines-per-agent layout.

Badge column width — badgeColumnWidth now includes the current agent (unlike the old computeRosterLayout which skipped it), which is correct since the current agent is now rendered as a regular line.

Marker logic — The spinner/▶ precedence in the switch is correct: working agents (any) show a spinner; the current non-working agent shows ▶; all others get blank padding.

No actionable bugs were found in the introduced code.